Nintil

Links (83)

Sat, 23 Nov 2024 00:00:00 +0000

Laura Deming on understanding

Interesting uses of AI

The memes of the wealthy

Elon dreams and bitter lessons

Nabeel on Palantir

Scott reviews Deep Utopia

Meditation, considered harmful

There is no placebo effect

Dwarkesh interviews Gwern

Restoring fertility with iMSC transplantation in monkeys

Profile of James Fickel, a not very well known crypto backer of many impactful projects in the longevity space

The King and the Golem, by Richard Ngo

A recent study in monkeys claimed metformin works to slow down aging. Not so fast, says the SENS foundation

I haven't done a deep dive on seed oils yet but my base take is that they are ok in low quantities - though you shouldn't be using oil other than olive oil (it's the tastiest and healthiest) and you shouldn't be deep frying stuff if you really care about health. Some thoughts from BJ on this. An article from Levels as well with a similar take.

Life in HD: on jhanaing (And anti-jhanaing)

Spicy take(down) of YC and Paul Graham-thought (ht/ Richard Ngo for the link)

CRISPR babies are here to stay, the question is how

Video about Science Corporation

Claude as therapist, various tweets

Pacific Fusion, another fusion company just launched

How many ways are there to build life? Not that many perhaps

Unlocking the emotional brain, a unified theory of therapy

Alzheimer's and inflammation

Otherness and control in the age of AGI

Sheekey Science Show on skin rejuvenation

Links (82)

Sat, 05 Oct 2024 00:00:00 +0000

Lots of interest in gene editing startups but in practice they don't do that well: Few diseases can be corrected through gene editing, hence valuations of such companies, despite FDA approvals, are low. Compare with addressing shared causes of multi-morbidity like aging or obesity. If you want to work on the former, at Retro.bio we are hiring :)

The decline of social status of stay at home moms as cause of fertility decline (via Scott Alexander)

Why can't the US build ships

Isaak Freeman's journey to learn mandarin

Trevor Klee on Robert Moses

Advice on finding writing ideas, with some of my own!

Don't get too worried about microplastics

A while back I proposed this idea of going around successful labs and researching their management practices. Why are top labs top labs? Now, it is happening.

In 'qualitative metascience' there is a rich body of work to be developed about how to fund science or conduct research. Startups have Getting Things Done or High Output Management or The Great CEO within and a plethora of supporting essays and articles about fundraising, managing, or marketing. Startups do make use of this hard-earned knowledge through experience. What does science have that's comparable? (Limits and possibilities of Metascience)

Entrepreneurship changed the way I think, by Casey Handmer

Paligenosis, or tissue regeneration

Burning Man

Fri, 06 Sep 2024 00:00:00 +0000

On August 25th I found myself sleeping under a tent at Black Rock City (BRC), the temporary city where Burning Man takes place. It was my first time going. A month before that I had no plans to get there though earlier in the year I had thought of going but made no definite plans. I thought well, perhaps next year, which is what I had also thought the year prior. But then, two weeks before the event, a conversation with a friend that was looking for someone else to go with changed that.

Ultimately, to go to burning man you “just” need

A ticket
A way to get in and out of Black Rock City
Means to survive

And surely one can get that in three weeks, I thought.

Finding tickets was easy enough: one can always find people selling tickets as the date of the event comes close. Getting there and means to survive are an entangled choice: If you are taking the burner bus, which we did, you can’t carry that much with you which means you have to rely on a camp to bring you water and food, but in exchange you get to skip the car queue. If you have an RV you could do a fully DIY burning man experience and camp by yourself but it could be pricier. I thought that renting (and cleaning) a vehicle, dealing with potential maintenance was not worth the hassle this time. If we had a bigger crew it’d have been different I suspect.

We got to Burning Man Sunday 25th of August and got back on the 1st of September. Though it was only a week it felt like it lasted longer than that. It was certainly an unusual experience: Because phones don’t really work there very well (and I lost mine anyway), your real world life is put on pause until you get back and is replaced by the day to day of being a BRC resident. I didn’t feel the need to check Twitter or text anyone when I was there, so I did get to be fully immersed in the experience.

Once we got there, we set up camp, got some water and explored around. We had our first encounter with the BRC porta-potties, the place to go for your toilet needs if you don’t have an RV. Though we later found a camp that had private porta-potties and used them once, we found ourselves wishing to have access to nice regular toilets, especially for times around bedtime and waking up. In any case when one’s out and about living at BRC, there’s no escape from the porta-potties, as getting back to camp is a long journey that will leave you with a sore butt (if you bike) or take forever (if you walk) depending on your location in the playa.

Then we had to get used to living in a state of near-dehydration all the time. Sweat evaporates instantly on the playa, so you are losing water but not noticing it at first. You have to drink water all the time if you feel even a bit thirsty. At BRC a greeting we heard was “piss clear”; because if you are not, it’s a sign you’re not drinking enough water. “Fuck your burn!” is another one that I was amused by.

Black Rock City is a bit of a miracle. Every year, a completely empty piece of land is turned into a city replete with art installations, electrical installations and water infrastructure. Every year, it is all disassembled and the desert looks as barren as it did at the beginning.

Some of this infrastructure is shared across camps: The camp we were at was getting electricity from the one 40kW diesel generator over at Naked Heart for example. One or two of the mornings when most other people were sleeping, I wandered around following the cables coming from our fridges and the hose we were getting water from, seeing what was at the other end, how the whole system was connected together. I had the thought that someone had built those things “for free”, someone that also had to pay for a ticket to be there. This is a key aspect of the burn; it’s not just having fun in the basic sense of dancing to techno, being driven around in an art car, and accessing interesting mind states, but also about surviving together with your homies in the desert. Not too far from something I say is a key source of meaning for me in life: going on good quests with the homies, or more prosaically working on meaningful projects and forming meaningful relationships.

It did surprise me how much time we spent on basic maintenance tasks: getting water refilled, going to the toilet, finding other people, moving between camps, cooking, finding food. Doing anything takes effort when breathing is hard and your heart rate is almost doubled even under the shade. Especially during the middle of the day, wandering around felt horrible, so sometimes I would just hang out at the camp resting and perhaps washing some dishes or helping make food.

Sleeping was definitely problematic. At Burning Man you have a menu of options:

You could sleep at night and be awake during the day; however you will need really good soundproofing as you’ll go to bed and wake up with the thumping untz-untz of EDM in the background
Or you could try to stay awake at night and sleep during the day, but then it’ll be really hot in your tent, unless you find an AC dome (what we did) or bring a swamp cooler (which I also did, but didn’t use in the end)
Someone recommended going to bed at 9pm and waking up at 3am to get the best of both worlds. Perhaps this is indeed the best move but one has to be very disciplined and commit to it as a group.

If you’re part of a group, different people will have different sleep requirements. Being out of sleep sync means different people will have different energy levels at different times and the vibes may be slightly off as a result; someone may want to go to bed while someone else wants to go dancing. Planning what to do at Burning Man to the hour is hard, but agreeing on what to do about sleep seems like a good idea if you want to keep a group together.

In theory, you could show up to BRC almost empty-handed and survive. “The playa will provide” they say. But if everyone did that the whole thing would fall apart, as a good citizen of BRC, you should be self-reliant. Once you try your best, if you happen to have failed to prepare sufficiently, you’ll most likely be ok: you can show up at a camp and you may get water, food, coffee, drinks, or a place to rest. You can wake up one day and there are more coffee and pancake offerings than you will be able to visit during your stay. There are bike repair camps, outfit swaps, community bikes (hard to find but we used them once!), AC domes, tea lounges, all offered as gifts to other burners.

Ok, enough with the survival stuff, where’s the fun? Burners are like clowns in a children’s chemotherapy ward; trying to both provide and have fun despite the grim background facts of the inhospitable playa. The obvious thing is music (mostly EDM, but one can also find classical music, perhaps even an orchestral rendition of Queen’s Bohemian Rhapsody) and dancing. Surprisingly I didn’t do a lot of that this time! I spent a lot of time having conversations with people about their own experience. In a good way, conversations felt self-referential as in circling.

Running into people I know from SF was really fun, little surprises happening throughout the event. I knew some of the camps where friends would be at, but no guarantee that they would be there when I visited. I had a decent success rate at showing up and saying hi to someone I knew. We slowly collected a crew, someone we met at an authentic relating workshop, and someone I faintly knew from the Berkeley group house scene. We stayed mostly together until the end of the burn.

Then there were the little absurd experiences: getting spanked before you get coffee at Scarbutts (Why? I asked. “It’s fun” they said), being driven in a mattress-on-wheels at night by a stranger than then led us to the Altitude Lounge, getting into a wood fired sauna (Though a dry sauna, as we added water it got more humid than the desert, getting out felt refreshing), or having a 3 course meal in a nicely set table (pictured below, we ended up going twice).

Sure I’ve had 3 course meals before, but having such niceties in the desert it felt like a hilarious “fuck you” that the BRC community is giving nature, a collective celebration of human ingenuity, and while Burning Man is not the most efficient (generators scale with size for example), its smaller scale lets you appreciate this sense of collective achievement more. Next time you go into a restaurant you can appreciate all the unseen but real effort (I Pencil, anyone?) that went into extracting what to you is an effortless meal for a convenient price. Hell yeah!

Most of the time I was with our small group of 4 people wandering around. Not being with a group often meant having trouble reuniting later, so the incentive to stay together was strong. But there were many occasions where I was alone. One night I took my bike and biked past the Man, the Temple, and almost all the way to the trash fence, just to see what was there (I didn’t quite make it all the way this time though!). The deep playa is a great place to explore alone. It reminded me of pictures like this, where everywhere you look around you in the far edge there’s a lot of stuff (art cars, music, lights) but in the space immediately around you there’s absolutely nothing, just darkness, with the occasional art car or bike passing along, like comets grazing a planet, getting close but not quite touching (Well, in one occasion someone got close and said “I’m looking for Titanic’s End!!!” to which I replied “Me too! Let’s go find it!”).

Burning Man is life, lived differently. In the real world most people are not dancing all the time, they are also resting, having conversations, or reading a book. The same is true at Burning Man. You can experience the Burn in any way you want. The experience, I’d say, is not so much about the specific content but about the state of mind, being exposed to serendipitous happenings, feeling higher highs and lower lows. I got to feel the worst I've felt in a while at Burning Man after not sleeping much for two days in a row. In the moment, I was tired, angry at minor annoyances, and quite grumpy. But after the event, I'm grateful I'm not like that most of the time! And that though I wasn't the most pleasant person to interact with one or two of those days and had some tense moments with our group, by the end we were all four of us sitting together to watch the Man burn.

Would I go again? At least once more, yes!

Appendix: Packing list

So what do you really need to go to Burning Man? There are a bunch of packing lists on the internet one can find and if anything I found I was overprepared in some regards and underprepared in others. The items that I ended up using all the time and that I’d say are the bare minimum would be (setting aside shade, water, food, bikes; that was provided for us by the camp this time)

A waterpack; I got the 2.5L Katari 3 from Osprey as a portable lightweight source of water. A mistake I made is not buying this accessory to keep the nozzle clean.
Electrolytes. I got these ones and added a pack to my water once or twice a day.
Sunscreen. My skin doesn’t burn easily but just in case I used some; mostly in my face
A mask. The weather this year was nice but there were some days when I was glad to be able to breathe filtered air. I got this one
Sunglasses. One could always squint but it’s nice to have them. I had a pair of Ray-Ban aviators at home so I took those
Boots. I went with Blundstones, and they did Ok, comfortable enough but I found them a bit harder to get into and out of them than I thought, especially when my hands were dry. Palladiums are another recommendation people have.
Good socks. I have a large collection of Darn Tough Light Hike Micro Crew socks and they did great. My feet didn’t feel too hot, I got no blisters and no sand inside. I changed socks every few days, they held really well.
A hat. I ended up exploring Black Rock City a lot during the day so it was nice to have one. I got this one.
Lights. This is quite important, particularly at night! I got a Spot 400 as a headlamp (plus a pack of spare AAA batteries that I didn’t use at all), a bucket hat that lights up, these LED armbands (super handy! They can be attached to a bike or to oneself)
Earmuffs. Really useful for sleeping! They block most but not all of the noise. Should have bought regular earplugs as well.
Chapstick. Your lips will be happier that way
Skin lotion. On day two my fingernails were bleeding from the dryiness. This probably saved me.
Blister cushions. Used only once but it probably made walking a much better experience during the week
A cup. I got a foldable one to be able to keep it inside my backpack. But perhaps I should have gotten one that is attached to the backpack instead to free up internal space.
Goggles. I didn’t use them (they ray bans did ok enough deflecting dust from my eyes) but there were no major whiteouts when I was there. I got these ones but they have small holes in them and sand could have made it through. I’ll probably get different ones next times.
Re sleeping bag, tent, and pad; I was in a rush and thought the camp would be providing them until last minute so what I got might not have been the best; I was also planning to donate them to the camp so I went for relatively cheap ones. Imho they were good enough: tent (an ok tent, quite small), pad, sleeping bag (comes with pillows but they were meh; the pad also had pillows, that was nice).
This bag from peak design for extra storage. I had it clipped to my backpack with this mini carabiner but it got a bit impractical. Next time having it on a belt would have been better.
A wristwatch. I had a Fitbit and the battery lasted most of the week, should have brought the charger with me!
Battery pack. Nice to be able to charge stuff without relying on power from the camp.

Other things I brought that are less important

Spray bottles. Useful to refresh oneself, wash your hands
A plate. Depends on your own situation but the camp I was with had communal plates so I didn’t use my plate that much.
A spoon/fork set. Same as above.
Wipes. Generally useful to clean your hands or to blow your nose. You’ll be blowing your nose a lot! You can also carry a bunch of 1-ply toilet paper with you too.
Toothbrush kit. It’s nice to brush your teeth but not strictly required, during that week. I did it every few days.
Flashlight. I have this Nitecore NC10000 which doubles as a USB-C battery to keep lights charged if need be and also it’s quite a bright flashlight
Warm leggings. It was cold one of the nights and this set helped
Saline spray for the nose. You can also use the spray bottles with regular water though.
Hand sanitizer. I brought a bunch; the toilets always had a supply of it, the rest of the time I ended up using it to wash my hands on the go.

Things I wish I had bought (that I would have used)

Walkie talkies! Finding friends after parting ways is hard! Phones don’t work at burning man so we had to resort to leaving messages in pieces of cardboard to meet at predetermined times. With walkie talkies, we could have been more flexible about splitting and reuniting later in the day. We got some from a camp in the last few days and they were nice to have.
Duct tape. If I had it at hand, I could have repurposed my backup non foldable cup into a cup I can carry around by taping the mini carabiner to it.
More than “bought” but rather “done”, is organizing better, I spent too much time searching for stuff. I should have kept a dedicated space for each type of item in designated bags or storage space for ease of access.
A water bottle; I had to use my backpack for water inside the tent and I wish I could keep it outside and have a water bottle inside the tent at all times for ease.
More fun costumes (I bought everything in a rush, and I ended up wearing 2 pairs of shorts, a few shirts, and this robe most of the time)
A collapsible miniplate; I got less food around than I thought so I didn’t miss this, but next time if I get more food this would be useful.

I have to mention here a few resources that I found helpful when preparing my own packing list:

Spencer Greenberg’s Packing List
Things people regret bringing/not bringing
These other packing lists

Links (81)

Sat, 17 Aug 2024 00:00:00 +0000

Jean Hebert, who proposes the stepwise replacement of brain tissue, has now joined ARPA-H

Why does Ozempic cure all disease

Scott Alexander on Nietzschean takes on morality, with an additional take from David Chapman, You should be a God-Emperor

Radiant, company started by ex-SpaceX engineer to

PBMs: a key player in the fuckupedness of the american healthcare system

Via Scott Alenxader's links, powering airplanes via microwaves. I had this idea back in 2021.

Blue Origin's factory

Using toxoplasma for delivery into the brain

Profile of Palmer Luckey

Nudge, a company trying to build an ultrasound headset to enhance human experience

Twitter thread wherein I discover that not everyone likes 1:1 hangouts!

"From barista to billionaire" (via Nick Camaratta)

I did circling

Links (80)

Sun, 14 Jul 2024 00:00:00 +0000

AI companies need to make a lot of money for the current market state (NVIDIA going to the moon) to make sense. Right now it's far from that.

Interview with the current co-CEO of Netflix

Lambda School, a scam

A tour of Starfactory, with Elon

Advantages of incompetent management

Ben Kuhn on trust and growing teams

Mehran's Steak House: making of

Why haven't biologists cured cancer?

The occurrence of cancer seems to continue to raise exponentially even in old age, contrary to what was previously thought

Using evolution to fight cancer

Quantum computing state of the art from 2019 outpaced by progress in classical computing as of 2024

Does meditation experience get you to jhana faster? Weirdly no

Naked Mole Rat mortality rate does not seem to increase with age, additional data keeps supporting this conclusion

What kills cancer patients?

Why microbiome research is hard

Making synthetic blood

Links (79)

Tue, 11 Jun 2024 00:00:00 +0000

Michael Lewis (author of a recent book on FTX)'s Blind Side

Xylitol bad?

Demons and Internal Family Systems

GLP1 analogues do not cause muscle loss in excess to what one would expect via caloric restriction. And we know that in that case one can avoid that to some extent by a high protein diet + exercise.

Cradle (formerly known as Lorentz, a name I deem more based), Laura Deming's new startup working on cryopreservation of tissues

Lumina, the probiotics company trying to cure cavities, playing very dirty here

That study saying that dance was better than SSRIs? Not so fast.

This tweet is a bit of a meme but there seems to be some truth to it, the post-exit startup founder to spirituality pipeline. This reply is interesting.

Dropping sterilized worms from airplanes in Panama, weekly

Andy Matuschak: How might we learn? (Check out his Patreon post for more context on the idea of exorcising oneself of the Primer)

The solar industrial revolution as an investment opportunity

Anthropic's sleeper agents paper

A song: BoZlak

Links (78)

Sat, 11 May 2024 00:00:00 +0000

Some debate over the merits of Minicircle, the gene therapy startup. It's a case of "in theory, it shouldn't work", with "working" defined in the most damning way: probably not even raising follistatin.

The debate over the usefulness of healthcare, Scott and reply from Robin Hanson, followed by reply from Scott and reply from Robin, with some highlights from Scott, which I think contains a good closure to the saga, with Hanson stating what he believes in and Scott agreeing roughly with that view.

And more biotech debate, this time about the Lumina probiotic that claims to prevent cavities.

On the synergy (ie the bad kind of feedback loop) between fat gain and inflammation

Reversing atherosclerosis with gene therapy

Michael Elowitz on synthetic biology

Astro Mechanica

The origins of woke, a book review, with comments

Research Leader's Playbook

I recently finished listening to the Billion Dollar Whale audiobook on the 1MDB scandal, which I recommend. Kevin Kwok has notes here.

Armand Cognetta on learnings running his company

A weirdly whimsical healthcare software giant

On the submarine cables that carry most of the world's internet traffic, and their repair

On jhanas

And, from the same person, understanding science funding 2011-2021

Trevor Klee on GLP1 and GIP receptors

Base Power, a vertically integrated battery/electricity startup

Brian Potter on fab construction costs

How might we learn, by Andy Matuschak

Links (77)

Sat, 30 Mar 2024 00:00:00 +0000

In defense of Academia

The Breslow saga

Interview with Scott Alexander

On Lockheed Martin's Skunkworks

An interview with the founder of Cyclarity, a company working on reversing atherosclerosis

On East Asian drinking culture

The ultimate lab leak debate

Why did supersonic airliners fail.

Supersonic airliners coming back

Casey Handmer on making fuel out of air

On David Sinclair, sirtuins, and resveratrol

Devon Zuegel, Jan Sramek, and MIke Solana on California Forever

A TV show: Three Body Problem

A song: Syren

A book: How Big Things get done

Making Cells Young

Wed, 20 Mar 2024 00:00:00 +0000

I just published this article in Asimov Magazine called Making Cells Young, go have a look! :)

Docusign's 7000 employees and the iron law of B2B SaaS sales

Tue, 19 Mar 2024 00:00:00 +0000

Docusign has over 7000 employees. Someone on the internet periodically discovers this fact and wonder how could this be! What are those 7000 employees doing?

The first part to the "What are they doing" question is "Sales", but that's not the full answer. Per their latest 10K filing,

68% of their workforce is sales, marketing, and customer success so that's ~5000 of those
And then 1760 of the workforce, a 24% is engineering, product development and customer operations

Leaving sales aside, we can still ask: Do you need 1760 engineers to run this product?

Remember that in the heyday of Twitter, the whole company was 7500 people and post-Elon takeover it went down to ~1500. Twitter continues running and shipping features. Could Docusign run and do the same things it does with ~1500 people? ~100 people?

To me it's obvious that a site like Twitter is more complex than Docusign, but a fairer comparison is Docusign vs its alternatives.

Hellosign was acquired by Dropbox but before that they had in the order of low hundred employees [1,2]. We don't know how many employees they have now but we do know that Dropbox itself, a superset of Hellosign, has ~2700 employees. I couldn't find data of Dropbox broken down by engineering vs sales by headcount but I did find data on relative spend:

At Dropbox, for every dollar spent on R&D (ie engineering) they spend 50 cents on sales
For Docusign, that's 2.58 dollars

How unusual is this? What about say, another company with a large salesforce like... Salesforce? What sort of company is Salesforce? A much bigger one: They have around 80,000 employees. For them, their sales to R&D ratio is as Docusign's, 2.67 dollars of sales spend per engineering dollar.

Another company that came to mind is PagerDuty. It is also publicly traded and they also build SaaS software. They have ~1000 employees and also spend a nontrivial amount on sales: 1.4 dollars per dollars of R&D.

So in terms of the sales/engineering split, Docusign is probably not that unusual. I don't have a good model of how big sales should be; it's easier for me to reason about how many engineers, PMs, designers and so forth one should have to build the core product the business is selling. So the question then becomes: Is it reasonable for Docusign to have 1760 people building the product?

Consider yet another company that builds a product that, I'd argue is more complex than Docusign: Zoom. Zoom has 6800 employees (less than Docusign!). Is that because of sales? Perhaps! Zoom's sales/R&D ratio is 1.91. How about headcount? Hard to tell from public data!

Here Tomasz Tunguz, a VC, looked at this ratio back in 2013 and found that indeed expenses on sales and marketing tend to be ~2x that of engineering, almost regardless of how long the company has existed. Headcount, he argues, follows a similar story. Docusign fits this pattern reasonably well. Another interesting datapoint is that half of the revenue in SaaS goes to sales and marketing, with 25% for R&D.

That still leaves us with the question: Why 2:1, and more importantly what should the base numbers be? Why not 100 engineers or 1000?

Here they try to do the accounting of how many people are needed for a ~100 headcount company that does 10M$ ARR in a somewhat structured way, coming to a staff of 70 marketing+sales+CSM. Scale that times 70 to get to Docusign's 700M$ ARR and you get 70*70=4900 people, roughly what Docusign has.

Intriguing.

Suppose you model sales as a lottery, where every sales call has some chance of progressing through the funnel, and eventually to a customer signing up. Then one could have a model where to get X dollars of revenue one needs Y hours of salespeople work, and translate that to people working regular hours, given a low enough chance of a sale, and that leads to needing a large sales staff.

This model could explain why, if a company believes it has a potential to make in revenue, it needs a staff of 7 people per $M to handle sales. But it doesn't explain why engineering count is what it is. One could imagine, for the sake of the argument, a team of 2 very capable engineers building a great product and a team of 1000 salespeople selling it all over the world.

Why is engineering headcount half of sales instead of much lower given how software scales?

A fun exercise: Suppose you have 7 sales people. Per the model earlier, the company has $1M in revenue. Half of that (per the averages earlier, but also with some guesses about salaries) will be their salaries. Assuming 2 engineers paid $160k (it's a startup) that's already $820k. Add the CEO's salary ($100k say) plus rent and infrastructure and we get to the revenue being fully utilized. The number of sales people sets a maximum on how many engineers you can have essentially, and that number might end up being 2 engineers per sale s person at the going salaries.

Great. So if you are a startup that has some extra cash to spare, what are you going to do? You could just keep the cash and enjoy your earnings, or you could try to make more revenue hiring even more sales people. Or you could hire more engineers. You could be fearing competition, driving a need to throw as many resources as possible to building the product as possible; so if this hypothetical company makes another $1M, they will be hiring in the same ratio to make it self-sustaining. If they did not do this, and were able to grow revenues without salespeople, or if they have a really efficient engineering team, then this multiple should change. But in a competitive market, if there are extra profits, these must be reinvested into the business, and that reinvestment means a ratio of 2 engineers and 1 extra sales person.

This seems true across the board in this list of SaaS companies from 2022. The only companies that have more than 10 percentage points higher spend in R&D than sales+marketing are Dropox, Unity Software, Olo, DigitalOcean, Alkami.

Dropbox, Unity, and DigitalOcean I have heard about and I could imagine how, through word of mouth and a B2C segment, they can spend less on sales.

What about Spotify, a company that's chiefly B2C? Spotify seems to have over 9000 employees(!) which is more than I thought. Out of these, 4719 are in R&D, 633 in content production and customer service, and 2403 in sales and marketing. Even pooling these latter two together doesn't get us to over one salesperson per engineer, confirming what we'd expect for a B2C company. But surprisingly, their budgets are more even, 1725M€ for R&D and 1533M€ for sales and marketing, similar to B2B SaaS. Maybe here it's because instead of having salespeople calling people (which would raise headcount), Spotify relies more on paid billboards, social media, and ads. This is the same as with B2B SaaS: a dollar spent in such ads should bring some dollars in revenue that can support a number of engineers.

At this point my explation for why these companies have the headcount they have:

Market size dictates sales force size and annual revenue
Annual revenue, minus sales costs, plus competitive pressures dictates engineering spend and through it headcount

There's one more missing factor: The belief that the company is as productive as other companies or something like that. ie if you think you could build the same features your competitors can with 10x fewer engineers then you can hire less and profit more. Argueably this is now the situation with post-Elon Twitter if advertisers return. But most companies probably think they are normal. Perhaps, as companies scale, they have to be normal and as productive as any other company. But this is copium: Tesla (look at their net profit per car) and SpaceX (launches per employee?) show that this is simply not true.

One last idea I want to add to this post is that even if a company is well positioned in its space, its future and fat profits are always uncertain. Therefore, putting some resources into new products or features, just in case, makes sense. The cost of being overtaken is very high, an existential risk indeed, whereas slightly reduced profits is a minor trouble. Those extra thousand engineers are maybe there for that reason.

Links (76)

Sat, 24 Feb 2024 00:00:00 +0000

A critical(!) review of Scaling People, a book by Claire Hughes Jones that I liked (ie I liked both the book and the review) though I am not as critical as the review is. CHJ 's book remains one of the best attempts at making the tacit art of management legible.

What happened to the US machine tool industry?

Scott Alexander: Should the future be human?

Things you learn dating Cate Hall, and Cate Hall on agencymaxxing

Casey Handmer on Elon Musk

The promise of SGLT2 inhibitors

Menopause and longevity

Karuna Therapeutics: drug repurposing success

Paul Janssen and drug development

A post so spicy it deserves a Scoville scale rating: Grad school is worse for health than STDs

A cell is not a computer

The Tsimane tribe has near zero rates of heart disease

Past and present of Google

What science can learn from car mechanics

Debugging chronic fatigue syndrome, and debugging health conditions broadly

How to hire low experience, high potential people

Dan Wang 2023 year in review

A private company just landed on the moon!

Mamba, latest in the state space models vs transformers wars

Speedrunning the drug application process

California Forever, a review from Devon

Podcast with Dwarkesh and Patrick

A song: Hania Rani's Ghosts

Recently I went through an extensive sofa-buying arc, looking at many options for the objectively correct sofa anyone should buy for a home (not an office; for that see the rationalist interior design guide). The winner is this one.

Links (75)

Wed, 20 Dec 2023 00:00:00 +0000

"The startup burnout to spirituality pipeline is strong for a reason.": A glimpse of absolute perfection

Short term twin study on the impacts of a vegan diet in healthy-ish patients. Compared to control, patients in the vegan diet saw decreases in fasting insulin, weight, and LDL cholesterol

FAQ on Lantern Bioworks, a biotech startup producing an oral probiotic that can prevent cavities for life

Scott's Beyond Abolish the FDA proposal

All the different kinds of cardiovascular disease, and how many deaths they cause

Endpoints interviews Jacob Kimmel, New Limit's Head of Research

An organ by organ proteomics aging clock

Gavin Leech's Breakthroughs of 2023

Vaults are wild. Now what about putting viruses into them to hide them from the immune system?

Vitalik on techno-optimism, and commentary from Michael Nielsen

Successful case of policy crafting from blogposts to national policy

Speculative Technologies (formerly known as PARPA) year in review

An article about Retro Biosciences, where I work

A Vatican-sponsored clinical trial where they inject stem cells into the brain

Elevating a single inflammatory cytokine (IL-6) lifelong is sufficient to, in some ways, prematurely age mice

DALLE, make it more

Foldable coat hangers

Vitalik Buterin on Zuzalu

First CRISPR-based therapy, now approved by the FDA

Rapamycin is not an aging drug. But what is an aging drug?

Sun, 10 Dec 2023 00:00:00 +0000

The title of this post will spark some controversy. But it is, under a reasonable interpretation, not false, just spicy 🌶️.

It make sense to start from the perhaps less controversial view that "rapamycin is an aging drug" and explain why that is wrong as well.

My issue with those that talk about rapamycin as an "(anti)aging drug", or a "geroprotector" is that it's often unclear what of this is being said

There's a unified thing called aging that affects everything that goes wrong with chronological time. Rapamycin modulates that one lever and as a result improves health and lifespan.
Rapamycin increases lifespan and has broad health effects. An aging drug is anything that broadly does this.

(2) is a perfectly valid statement to make. (1) is known to be false. But I think it's easy to fall into the (1) frame of thinking when reading some of the rapamycin papers, or sometimes popular press coverage thereof. So: rapamycin is an aging drug in the sense of (2) and it is not an aging drug in the sense of (1). We should prefer to say that "rapamycin slows some aspects of aging" than "rapamycin slows aging" or "rapamycin slows the aging process".

As I will mention below, rapamycin does not simply turn back or stop the clock. It doesn't take every organ and resets it by some number of years. Rapamycin leaves some functions untouched and makes others worse.

Now at this point some may go "duh of course, aging is multifactorial and rapa is addressing parts of it" whereas other might go "Heresy! Surely the papers that show that rapa doesn't help with everything must be wrong or something". The second is more clearly a caricature. Even well known proponent of rapamycin as an aging drug take it as a given that it has its downsides (the mouth ulcers are well established in humans). But then, if we agree that "rapamycin is just affecting some aspects of aging" then we have to be very careful when reading claims that "rapamycin slows aging" as that sounds like it is slowing all aspects of aging and perhaps even that all aging is mTOR-driven.

I suspect that the intimate relationship between rapamycin and aging in some's conceptual schemes dates back to the origins of the field. The mTOR-IGF axis is where the field was born, studying caloric restriction. CR also has lifespan-increasing and broad health-promoting effects, so the idea that there was a master program regulating lifespan and health and that this is tied to mTOR can be coming from there. This mTOR maximalism view is presented by eg. Blagosklonny but it is not an explicit majority view.

As a brief introduction, rapamycin is an mTOR inhibitor that is able to extend lifespan in the usual model organisms: yeast, flies, worms, and mice. Its lifespan-extending effects are perhaps one of the most consistently replicated findings in the field, showing success in the Interventions Testing Program, a large sample size study using outbred mice of various interventions aiming to extend lifespan.

In worms, yeasts, and flies, rapamycin seems to extend lifespan because of an autophagy-inducing effect (i.e. it depends on the presence of key autophagy-related genes to work) whereas in mice there are additional effects present, possibly anti-inflammatory (Rubinsztein et al., 2011) and surely anti-cancer. In addition to living longer, rapamycin slows down the appearence of various age-related phenotypes in muscle, tendons or liver (Wilkinson et al., 2012).

Given that rapamycin extends lifespan and improves health, how could one possibly say that rapamycin is not an aging drug? Is it not impacting biological aging?

I honestly don't understand this view. Rapamycin increases lifespan and delays declines or improves function in multiple (every?) aged tissues when started in middle age. What is the alternative explanation other than impacting biological aging?
— Matt Kaeberlein (@mkaeberlein) December 4, 2022

First, one has to question the narrative that rapamycin improves "health" as a unified entity. Rapamycin improves some aspects of health and worsens others. The Hallmarks of Aging paper cites the Wilkinson paper from earlier to say that "rapamycin delays multiple aspects of aging in mice". However it completely fails to mention that the paper also says that rapamycin accelerates the appearence of cataracts and leads to testicular degeneration. In humans, mouth ulcers are a common observation as well (Peterson et al., 2016). It's up to each individual to judge if the traeoffs will be worth to them, given other interventions they may be undergoing.

In my view the most rigorous and comprehensive examination of the effects of rapamycin in mice has been done by Dan Ehninger in these papers:

Deep phenotyping and lifetime trajectories reveal limited effects of longevity regulators on the aging process in C57Bl/6J mice (2022)
Longevity, aging, and rapamycin (Ehninger et al., 2014)
Rapamycin extends murine lifespan but has limited effects on aging (Neff et al. 2013)

The overall point made by Ehninger in these papers is that rapamycin is extending lifespan in mice in part via cancer suppression, and that in addition to that, it is having effects on only a subset of age-related phenotypes and furthermore that it does also improve function in young animals so he claims that (his words) there is currently little evidence available to support the notion that mTOR inhibitors slow the rate of mammalian aging. Ehninger acknowledges that the lifespan-extending effects of rapa are not just due to the anti-cancer effect: cancer is not a cause of death in yeast, flies, or worms, where lifespan effects are also seen. But it may be that the reason that rapa increases lifespan in those is different from the reason it increases lifespan in mice.

To make this more clear: To extend lifespan in mice one must ultimately slow down cancer because most mice die of that, which mTOR inhibition does either indirectly (inflammation reduction, for example) or directly (slowing down growth of the cancer itself). Then there's a separate autophagy-enhancing effect which may be doing little for the mice in terms of mortality, but a lot for the lower organisms.

Importantly, it is possible to extend lifespan in a given organism in many ways. Whether an intervention is addressing aging depends on the definition you are using. That is, a definition that puts a heavy weight on "lifespan extension" will classify chemotherapy and statins as a aging drugs in that they address phenotypes that are more common with age (cancer and CVD respectively) and they extend lifespan as a result.

Ehninger points out that if a drug is having the same effect in young than it does in old, it is not really addressing aging but rather reversing age-related phenotypes via aging-independent mechanisms.

Take muscle mass; after development it declines. Reducing the activity of myostatin can increase muscle mass, and this increase can happen in both young and old. This treatment could plausibly reverse sarcopenia in the old (ie shifting muscle mass % closer to young levels).

Now one might say: A myostatin knockdown is not really an aging intervention because it's addressing one disease directly (sarcopenia) through a very proximate regulator of muscle mass. It is not systemic. To which one could say: ah but did you know follistatin overexpression (which reduces myostatin signalling) perhaps increases lifespan (Kumar et al., 2022). Perhaps follistatin treatments are more systemic than one thought at first! Perhaps it's acting through cell-to-cell signalling which is a hallmark of aging. Given this, is follistatin as much of an anti-aging therapy as rapamycin is? Again, depends on how you define aging. I personally think one should not obsess over this question, but rather about what works and what doesn't to improve healthy lifespan in a sustained way.

Rapamycin connoiseurs will say that in the latter paper (Neff), the rapa dose was lower than what has been used elsewhere (Johnson et al., 2015) and that this might by why the effects are not as systemic. Even with this, we still have the negative effects found elsewhere and found in humans sometimes like mouth ulcers. Ehninger uses in the later 2022 paper an mTOR deficient mouse, where mice constitutively produce less mTOR. This model recapitulates known effects of rapamycin: lifespan extension (~20%) coupled with broad health effects (Wu et al. 2022). The Ehninger paper measures 208 phenotypes out of which 117 change with age (age-sensitive phenotypes, ASP). They then classify the effects according to whether they have an aging interaction (ie they show an effect in aged but not young animals) or only a genotype interaction (if the effect is present in both aged and young of the same genotype). The result here:

16% of ASPs are improved in old but not young ("true effect on aging" in Ehninger's view)
31% of ASPs are improved in both old and young ("health-promoting, aging-independent effect" in Ehninger's view)
36% of ASPs are unchanged by the mTOR-deficient genotype
13% of ASPs were made worse by the mTOR-deficient genotype

While one can quibble with the exact phenotypes they are measuring and whether a number is 15 or 10%, given other papers, it does seem a robust finding to me from other papers: put it simply rapamycin extends lifespan and improves some aspects of healthspan while worsening some others.

Reprogramming as only current approach to systemic, permanent, rejuvenation

Ultimately rapamycin is a small molecule drug. What it can do is limited: It can bind FKBP12 and jam mTOR and that's it. That in turn turns up autophagy and other cellular programs that contribute to general cellular health, leading to increases in lifespan and healthspan. But it also reduces the rate of protein synthesis and has other negative effects. Rapamycin is not an endogenous metabolite that goes down with age (the case of taurine), but rather a hack to crank up pathways that have effects that we mostly want. However, the effects of rapamycin are not permanent nor they reverse aging at the cellular level, rather they slow some aspects of aging down (Kabacik et al. 2022; Horvath et al. 2019). We should think of rapamycin mostly as a way to slow down aspects of aging transiently, but not as a path towards robust age reversal. A "better rapamycin" in my view, is not particularly worth seeking, compared to alternatives.

Note that this is an empirical observation: It could be the case that an intervention that say boosts autophagy lets the cell "catch up" on garbage collection, resetting the age of the cell permanently. But in practice, this is not what rapamycin seems to be doing. It stops and delays, not repairs and reverses.

The only intervention that can permanently and systemically reverse aging at a cellular level (with minor exceptions like DNA mutations) that we know of as of today is reprogramming. I'd like to be wrong on this one: more interventions are better so if you have something in mind that can reverse aging permanently, send it my way at [email protected] .

Unlike therapies targeting various aging hallmarks, reprogramming is permanent. Once a cell has been made into an iPSC, it stays young in almost every possible way, and once the cell is made back to its type of origin, the clock is reset, without the original OSKM factors having to be present anymore. Under a very strict definition where "age reversal" means that every measurable thing has been reverted back to a young state, the irreversibility by reprogramming of DNA mutations would disqualify it. But under a more relaxed "general reversal" definition, it is second to none.

The issue with reprogramming (or partial reprogramming) is that it's harder to control in vivo than rapamycin so applying the intervention is not as easy as giving some mice an Eudragit-coated serving of rapa. It's going to take some engineering to make it work there. Alternatively, one can reprogram ex vivo cell types one at a time and then engraft them back, fully rejuvenated cells.

But aren't the hallmarks of aging connected? Yes to some extent but not fully: per the Kabacik paper earlier rapamycin indeed affects others. But we already know the deterioraton of epigenetic information in cells is something rapamycin does not reverse. Within the hallmarks, some are more upstream than others, and that which rapamycin is affecting is not as root-cause-y as that which reprogramming targets.

Systemic rejuvenation is not enough

Even reprogramming doesn't do everything. I think no matter how much one reprograms one will eventually get cancer (cancer can overwhelm even a young immune system after all, see child leukemias!). Thus there are a series of things that are required for continued increases in healthy lifespan that are very targeted to specific ways to our biology: cancer therapies being next gen chemotherapies like 6-thio or immunotherapies (Like our own T-cell therapy!), amyloid plaque targeting therapies (if cellular rejvenation can't clear that), atherosclerosis targeting therapies like Repair Bio's, therapies to add new cells that were lost (Can't rejuvenate what's not there anymore!) and potentially more.

One last word on rapamycin-like drugs

Though I spend my time thinking about the future and comprehensive solutions to the aging problem, in the shorter term there are drugs that potentially one could take that could help one live longer healthier lives here and now. With its downsides and upsides, rapamycin might be worth using for some people in some situations, and more trials of this one molecule in humans are something I welcome.

Appendix: claims regarding rapamycin and aging in various sources

Rapamycin slows aging in mice (2012, Wilkinson et al.)
- "Rapamycin slows aging in mice"
- "testing the idea that rapamycin might slow aging effects on many tissues and thus by inference slow the aging process per se"
- "the aging process is delayed by rapamycin in genetically heterogeneous mice."
Rapamycin extends life and health span because it slows aging (2013, Blagosklonny)
- "Rapamycin extends life and healthspan because it slows aging"
- "what is aging if not an increase of the probability of death with age"
Resveratrol and rapamycin: are they anti-aging drugs? (2021, Kaeberlein et al.)
- "A definitive link between TOR signaling and mammalian aging was established this summer in a report from the National Institute on Aging Interventions Testing Program in which the TOR inhibitor rapamycin was shown to increase life span in mice."
Rapamycin: Prevention&Treatment of aging and age-related disease
- "Humans and other terrestrial mammals have programmed death by Aging. Anti-Aging medicine is about blocking pro-death/pro-aging pathways. The most robust pro-Aging is mTOR (mechanistic Target of Rapamycin). Rapamycin is the most effective drug to block mTOR. Hence anti-aging medicine is called Rapamycin medicine."
Rapamycin rejuvenates oral health in aging mice (An et al., 2020)
- "The FDA-approved drug rapamycin slows aging and extends lifespan in multiple organisms, including mice"

Some thoughts on causality in biological systems

Sun, 26 Nov 2023 00:00:00 +0000

Sometimes one reads discussions of causality in academic papers. Expressions like "this gene causes that" or "Alzheimer's Disease (AD) is caused by XYZ". Or of course, "We don't know what causes X". Recently I found myself thinking about these three statements:

We don't know what causes Alzheimer's yet (eg here at NIA)
Old age does not cause Alzheimer's, but it is the most important risk factor for the disease (also here at NIA)
Aging causes Alzheimer's (eg here at Fightaging)

What does causality mean there? How can something be a risk factor but not a cause? Intriguing!

People (philosophers excluded perhaps) ususually don't think much when they throw around the word cause, but in biology, from the way the word is used, "X causes Y" tends to mean "X is necessary and sufficient for Y". Hence in Alzheimer's case as I discussed here one might make the claim that amyloid beta plaques don't cause AD because one can find older adults with plaques but no dementia. One can then counter that there are other factors at play. In the NIA definition above, "old age does not cause Alzheimer's" means that being old doesn't guarantee Alzheimer's.

But with early onset Alzheimer's, where having a small number of mutations (usually in the PSEN1 gene) people do speak of "those mutations causing early onset AD". If you have them, you will get the disease with ~100% likelihood, if you don't you won't. This is a clear example of the "sufficient and necessary" criteria that seems pervasive.

With cancer, we can make the broad claim that "cancer is a genetic disease" even though there is no one gene that uniquely causes cancer (no necessary set) but one can find combinations of mutations (say a p53 mutation coupled with a KRAS mutation) that do lead to the disease, (many sufficient sets).

But this is obviously too simplistic and breaks down especially wen we try to understand complex diseases or processes like aging.

If you take a car and ask what causes the car to move, there is no one element that causes it. If you take the tyres or the engine away or the car has no fuel the car won't move. All of these are necessary for movement but there are no sufficient "causes". The folks over at NIA, if assessing a car would then say that "We don't know what causes cars to move, but fuel, tyres, and the engine are known factors that are involved".

How did this come to be? Probably because biological experiments have historically proceeded one blunt thing at a time. It's easier to KO a gene to see what happens than to downregulate by 50% 10 different genes, but the effects of those 2 sets of changes may be the same.

But in the case of Alzheimer's, consider this alternative explanation. We know what causes Alzheimer's: Proximately, the symptoms we observe are due to neurons dying. Most of this death is due to the action of hyperphosphorilated tau aggregating into neurofibrillary tangles (NFTs). Amyloid is not required for NFTs to form, there are other tauopathies, but AD is an amyloid-driven tauopathy. We know microglia (and neurons too to some extent, via autophagy) can clear up these two aggregates, and we know that babies don't have these. So what's happening is that either the rate of production of the aggregate has increased (dysfunctional neurons) or that of degradation has decreased (say dysfunctional microglia) or both of course. This model has to be true a priori if we want to fit the basic facts we know about the disease. But the model doesn't say that one particular thing or gene is a cause of the disease. For example, one might have an ApoE4 allele that causes reduced clearance of amyloid beta, so the brain hits the threshold to disease progression much earlier. Or one might produce a lot more of Abeta (the case with the PSEN1 mutations) and reach AD much much earlier. If you think of these as equations, the form of the system is the same but the coefficients are different in each disease and in each person. We can perhaps think of late and early onset as the same disease, but one caused by faster generation and another by reduced clearance of aggregates.

But then there can be many reasons why the coefficients in the equations change: infections, accidents, different genetics, or other comorbidities. In this sense it's very much like cancer: no sufficient-and-necessary mutation but rather a series of little causes that nudge the brain towards the diseased phenotype.

In one person, for example, maybe they have really good autopaghy and so they stay disease-free for longer. In another, they are overweight, they have more inflammation and that leads to earlier progression. We would be wrong to say that "reduced autophagy causes AD" or that "inflammation causes AD" or that even "PSEN1 causes early onset AD" as even this requires some amount of aging to proceed.

To make things worse in biology all these causes are connected: it's likely that say increasing inflammation impairs autophagy or that at some point more amyloid leads to more inflammation and damage and that damage leads to more amyloid. At its core, DNA makes RNA makes proteins but then these proteins affect how DNA is transcribed.

If we go back to a mechanical analogy, imagine there is a little tear on the wing that causes some vibrations; over time this vibrations affect the engine which breaks down and the airplane falls. Likewise, imagine the turbine's shaft is not well compensated and generates vibrations, the wing starts vibing and it falls apart. Not that this would happen often in commercial aviation because we've gotten quite good at aircraft safety, but this highlights reaching the same phenotype (the crashed airplane) through two mutually-affecting routes.

What are the implications of all of this? That if we want to make progress in biology we have to move beyond the "gene for X trait/disease" idea and rather think of systems where we can still meaningfully speak of causation but in a more realistic way. ML will help there because these systems are devilishly complex and while sometimes we can make simplifications that are good enough to derive therapies (Like LDL cholesterol->cardiovascular disease) elsewhere (Alzheimer's) we cannot (We can clear Abeta but that doesn't cure or radically slow down AD).

As for aging in particular, if you think of the state of a cell as a dot in a high-dimensional space, then young cells starting near each other and then they disperse into various trajectories of dysfunction in a sort of Anna Karenina principle of disease: all healthy people are the same whereas all sick people are different. It follow then that almost every small random nudge to a cell won't be enough (there are so many dimensions) to make the cell young in a consistent way. By young I mean more functional as it was when young, so this applies to disease in general. We are not going to fix the aging brain upregulating or knocking out genes, at least not one gene at a time. For example if we see a reduction of lysosomal function in microglia and then we see that 10 enzymes that are key there are downregulated, are we going to throw in all those 10 mRNAs? And there are probably more things that are wrong than those 10 anyway. Note that this doesn't mean single drugs can't help in every case. Statins work and target a specific pathway.

This sort of thinking outlined here, it seems to me, should be a central part of what it means to "work on aging research": to seek to deeply understand complex interactions across genes, cells, and tissues, and across contexts.

Changelog

2023-11-29: I removed a reference to the central dogma of molecular biology at the suggestion of Per Kraulis, see here. Though I think it doesn't affect the thrust of the argument (I mostly used it as a textual embellishment, it was nonetheless wrong)

Links (74)

Sat, 04 Nov 2023 00:00:00 +0000

Tales from the patient side of oncology clinical trials

Hackernews compilation of "things that don't scale"

Colonoscopies, do they work?

Future House, new FRO-like organization launches to work at the intersection of AI and biotech

Tattoos are made of macrophages that constantly die and are eaten by other macrophages

A song: Midnight in the Desert

Huh strangely not that much happened this month! This might be the shortest Links post in a while. Maybe I have been rabbitholeing less and a lot of the time went into the telomeres post + work-related reading.

Telomeres: everything you always wanted to know

Tue, 17 Oct 2023 00:00:00 +0000

Back in 2020 when I wrote the Longevity FAQ I had a section on telomeres and some cursory examination of how much they might matter for aging. Now that I know more it's time to revisit it, including also some historical notes and some papers that came out after I wrote that FAQ.

The direct reason for writing this post is that often when I say that I work on aging (at Retro Biosciences, we are hiring!) I am asked if "that's the telomeres thing". For some reason, telomeres became tied to aging broadly a few years ago and I wanted to understand why while at the same time I wanted to understand how relevant they actually are, as I suspect myself and others have overcorrected much away from "telomeres are the key to aging" in the direction of "they don't matter that much". From time to time one can see things from seemingly reputable outlets like The Nobel Prize twitter account saying that "the secret [to why we age] lies in our telomeres" and people in the field making fun of them.

To start, telomeres are the regions at the end of chromosomes, so all of our cells have telomeres. The details of what these regions look like are covered in a bit more of length in the FAQ; for this post what matters is the fact that when a cell divides it has to copy its DNA. When this happen, telomeres shrink. When the telomeres are too short the cells stops replicating and becomes 'senescent'. There is an enzyme called telomerase. Telomerase is not a single protein, rather it's a complex made of (in humans) proteins TERT (telomerase reverse transcriptase), TERC (telomerase RNA), and DKC1 (dyskerin).

It's important to note, as this gets lost in discussion sometimes, that when someone talks about telomeres shortening or lengthening it's usually the telomeres of white blood cells like T-cells, because accessing blood samples is the easiest. Changes in the telomeres of those cells does not imply changes in telomeres elsewhere. T-cells in particular can divide and the populations of different kinds of white blood cells can change with external factors, so it can also be the case that telomeres as usually measured lengthen simply because cells with shorter telomeres die, so average telomere size then increases.

For example, a person more exposed to infections and carcinogens may make their T-cells work harder, implying more divisions and telomere shortening. That person might die earlier and in a cross-sectional study we may see that shorter telomeres correlate with earlier mortality but this will just be a downstream biomarker of exposure to disease in this case, and not a direct causal driver.

What cells can divide (and so which cells have telomeres that can potentially shorten?) fibroblasts, endothelial cells, smooth muscle cells, glial cells, or astrocytes. In contrast, neurons or myocytes do not divide so we shouldn't expect telomere shortening to have much to do with the aging of those cells.

The usual function ascribed to telomeres is as an anti-cancer mechanism: if we cell begins dividing too much then its telomeres will progressively shorten and it will stop dividing (or die). To overcome this, cancers end up reactivating telomerase to keep their telomere length. Absent telomerase they can also engage in ALT (Alternative Lengthening of Telomeres), so all else equal we would expect telomere elongation to lead to higher likelihood of cancer, but this is a simplistic view, as longer telomeres might also mean a better functioning immune system (T-cells that can divide more) and less inflammation (fewer senescent cells).

The history of telomeres and their relevance goes back to at least Leonard Hayflick's 1961 and 1965 papers on what he called senescence: in culture, cells eventually stop dividing; in that paper he pointed to factors intrinsic to the cells but at the time he did not link that to telomeres. Cells extracted from older donors would stop dividing earlier, pointing to the existence of a limit to the total amount of divisions a cell could undergo, and that clock starting ticking since we are born, not just in vitro.

It wasn't until 1970 when a Soviet scientist, Alexey Olovnikov had a sudden realization:

The Theory of Marginotomy [JR: Telomere shortening] came to me in that Moscow subway station. I heard the deep roar of an approaching train coming out from the tunnel into the station itself. I imagined the DNA polymerase to be the train moving along the tunnel that I imagined to be the DNA molecule. I thought that this polymerase cannot begin to copy from the very beginning because there is a dead zone between the front end of the polymerase molecule and its catalytic center. This is analogous to the dead zone between the front end of a subway car standing at the beginning of the subway platform and the nearest entrance door to the first car. After this serendipitous underground brainstorm, which happened in the Fall of 1966, I wrote to Hayflick to ask some questions about his discovery and he sent additional unpublished data to me. I then spent several years thinking about this idea before publishing it in the central journal of our Academy of Sciences, Doklady Academii Nauk SSSR

So he was the first person to propose that it is telomere shortening due to imperfect DNA replication that is causing what Hayflick had observed.

Meanwhile in the US, Elizabeth Blackburn, who would later win the Nobel Prize for her work on telomeres publishes in 1978 the sequence of these regions in the chromosomes of Tetrahymena thermophila, a protozoan. Telomeres turns out to be TTAGGG (repeated 3000 times or so) in humans whereas in tetrahymena they turn out to be TTGGGG. Other species have other repeats. She would be working with Jack Szostak over the following years on understanding telomeres.

Later, in 1984 one of Blackburn's students, Carol Greider discovers the enzyme telomerase, initially called the telomere terminal transferase.

Blackburn and Greider's work also noted that Tetrahymena with knocked out telomerase became sick and died quickly (1990), so telomeres and telomerase seemed key to keeping cells alive. Unlike Hayflick's work, which was in cells from humans, Tetrahymena are complete unicellular organisms, so their work provided more robust evidence that telomeres could be driving senescence at the organismal scale. They also noted that this system may have evolved to prevent cancer, putting senescence and cancer at odds. But in principle, one could think, if there was a way to prevent cancer, would telomere restoration prevent aging at the organismal level in more complex organisms like mice?

Intrigued by these findings, Geron Corporation launched in 1990 with Greider and Hayflick as advisors looking to develop telomere-based therapies for aging. It was seemingly Geron and its PR department that started the "telomeres drive aging" meme in the general population, according to Greider.

In the 90s a book titled Reversing Human Aging came out, also supporting telomere shortening as key explanation for aging. In that book, we can find telomeres referred to as the "clocks of aging" and that "Aging is caused by aging cells and aging cells by aging telomeres."

In 1998 for the first time it becomes possible, thanks to Geron scientists to take regular cells and turn on telomerase permanently preserving telomeres. This enables the cells to escape senescence. The process was dubbed 'immortalization'. In the media we get quotes like "We found that biological aging can be put on hold"

Now at this point it would make sense to take this and put it in mice to see what happens. So in 2001, the Blasco lab turned on telomerase in a certain type of cell in the skin of mice. The result wasn't longer lived mice, but rather they mice got more skin cancer (and faster wound healing), confirming the double-edged sword nature of telomeres.

A year after another paper comes out studying the effect of turning on telomerase in the entire mouse since birth. The results confirm Blasco's: they also saw an increase in cancers.

Looking at these two papers and a few others, 2004 supposes an inflection point for the telomere hypothesis of aging. Gerontologist João Pedro de Magalhaes publishes Telomeres and Telomerase: A Modern Fountain of Youth? where he concludes in very skeptical terms that:

though telomerase may be used in regenerative medicine and to treat specific diseases, it is unlikely to become a source of anti-ageing therapies.

One key point made there that Greider also made trying to counter the telomere hype is that there's no correlation between telomere length and the longevity of a species. Mice have longer telomeres than we do but they live much shorter lives. A conclusion is drawn there that

Therefore, telomere length and/or telomerase activity do not explain why humans age slower than other primates and mice

But this is not quite right. Many years after (in 2019) that Blasco and others would show that while it is correct that telomere lenght does not correlate with longevity, the rate of shortening (which is different in different species) does correlate fairly well:

This doesn't mean that they set the pace of aging! Other things like the rate of accumulation of DNA mutations also correlate well with lifespan. In general if many mechanisms are necessary for long life, then evolution will jointly optimize them all for progressively increased lifespans. There's no living long with fast shortening telomeres OR high mutational burdens it seems.

After the de Magalhaes critique, Blasco publishes in 2008 for the first time a report of lifespan extension using telomerase overexpression albeit in cancer-resistant mice. Fine, one could say, keeping telomeres long causes cancer, but what if we avoid that by genetically engineering the mice to be more resilient than normal mice will ever be, do we get healthier longer lives in mice? The answer is yes we do. Compared to the baseline non-telomerase'd cancer resistant mice, the ones that have constant telomerase expression:

Preserve skin thickness with age
Preserve villi structure in the GI tract (and associated gut barrier integrity->less leaky gut)
Do better in a motor test
Show improved glucose metabolism
Live substantially more, and more mice reach very old age (42% vs 8%)
As one might expect, telomeres are longer

Promising!

Understanding telomere biology turned out to be important enough to earn Blackburn, Greider, and Szostak a Nobel Prize in 2009. (Sad day for Olovnikov)

Then one year after the telomere hypothesis continues to receive support with the publication in 2010 of Telomerase reactivation reverses tissue degeneration in aged telomerase-deficient mice. Note that these are not regular mice, but mice that do not have telomerase at all; their telomeres are much shorter than usual.

With this paper, there was then evidence than longer telomeres could, in a context (cancer-resistant mice) extend lifespan as well as telomerase deficiency could shorten it and that adding back telomerase could reverse that. But could that work in regular mice?

In 2012, the Blasco lab once again published a paper where they take regular mice and using an AAV gene therapy cause widespread telomerase activation. The result: increased lifespan regardless of whether the therapy was delivered early in life (24% increase) or later in life (13% increase). This increase was accompanied by other health benefits. Importantly, they got less, not more cancer. They especulate that because the AAV would get diluted in fast dividing cells, cancer cells would not be able to benefit from this additional telomerase.

The same year, a consortium of aging researchers published what's probably the most cited paper in the field, Hallmarks of Aging, where they establish the importance of telomeres as a consensus view.

Then later, in 2015 we get an interesting case of self-experimentation. Elizabeth Parrish, the CEO of a biotech company called BioViva injects herself with a gene therapy to lengthen her telomeres. They'd do it again in 2021 in some volunteers somewhat quietly, I hadn't heard of this study until today!

That same year we get another critique of the involvement of telomeres in aging. The paper starts by pointing out that knocking out telomerase in mice and breeding those mice, which leads to decreasing telomere size in the newborns over the generations does not decrease lifespan in the first generation (but it does later), implying that telomere length in mice is not a limiting factor for their lifespan. What then about the 2012 paper from the Blasco lab using AAV? The author argues the sample size was small and perhaps telomerase expression itself (and not just telomere length) is behind the effects seen. He also expresses doubt that telomeres can be involved in aging because mice telomere length varies a lot between inbred strains vs true wild mice without correlating to lifespan (Something we have thing is no mystery: what matters is the shortening rate, but that Blasco paper hadn't come out yet).

Blasco would provide an answer to one of these critiques in an impressive 2019 paper: mice with hyperlong telomeres live longer (12.8%) and are healthier in various ways. These mice were bred to have longer telomeres without overexpressing telomerase so one can rule out the potential action of telomerase in having the observed effects. They also get less cancer, suggesting that the anti-cancer effects from a better immune system or the inflammation reduction from longer telomeres outweigh the tumorigenic risks of telomerase.

So after the AAV paper from 2012, one'd think that many in the field would be relatively bullish on telomeres, but not quite. Aubrey de Grey, in 2018 was asked about the promise of telomere lengthening, said that he was worried about the risk of cancer, pointing to some of the studies showing an increase in cancer. To explain the results where mice with telomerase overexpression living longer, he points to the observation periods being too short (definately not the case for the AAV paper) or the animals being particularly cancer resistant (the case with the earlier papers).

Lastly, also cited by de Grey are some studies of human genetics using mendelian randomization to see what happens when the rate of telomere shortening is altered in humans. The result is a slight increase in cancer, a slight decrease in coronary artery disease, and a null effect on lifespan, implying that for humans in particular telomere length wouldn't extend life. Perhaps! But remember that in mice the length of telomeres doesn't determine lifespan in the general population despite the fact that hyper-long telomere mice do show improved lifespan. Could it be that people with hyper-long telomeres live longer? I don't expect so, given how slow human telomeres shrink, but the possibility exists. There are other papers one could point to that purportedly show that long telomeres in humans are deleterious; people with mutations in the POT1 gene have longer telomeres but one might reply that the mutation itself is having deleterious effect beyond telomeres. One could imagine a one-off AAV therapy that lenghtens telomeres and if cells turns cancerous and start dividing, the AAV will be diluted so the effects on cancer will be minimized (as Blasco et al. speculated in their own paper).

Another critique is from the Ehninger lab in 2022. The authors point out, correctly, that there's no correlation between telomere length and lifespan across mice strains, inferring from that that telomere shortening does not underlie murine aging. But given the hyperlong telomeres paper, this doesn't seem right. They also point to telomere lengthening increasing carcinogenesis, but they bunch together various kinds of interventions. Constitutive telomerase activation is not the same as transient, delivered via a virus (which would be closer to a hypothetical clinical translation). Ehninger just counts the number of studies saying there is more vs there is less cancer. 2 studies showing there's less (the AAV study and the hyperlong telomeres paper) vs others using constitutive expression. This is obviously not a fair reading of the evidence!

The latest big and somewhat controversial paper in telomere land is this one from Liz Parrish and colleagues (See Pubpeer for why it is controversial) but I don't think it's controversial enough to discredit the main relevant result: a replication, in a way, of the Blasco AAV paper showing lifespan extension, only that using CMV instead, showing a 41% increase in lifespan.

So what should we expect to see in humans? A disease, dyskeratosis congenita (DC) might give us some clues. Patients with DC have short telomeres and as a result they present with defects in organs and tissues that experience faster turnover: where we would expect telomeres to shorten faster (skin, nails, lungs, bone marrow). So we mighte expect that if we lived long enough and our telomeres shorten sufficiently, we would begin to experience something like this. According to this paper, telomere length (in white blood cells) in DC patients is similar to the telomere length of someone that might be over 90 years old, so it wouldn't be totally surprising if around that age people would start experiencing such symptoms naturally and given that bone marrow is one of those tissues, telomere-shortening ends up being a bottleneck for healthy lifespan extension at advanced ages.

A visual summary of what seems to be the case, at least in mice, would be something like this:

Now to finish, I want to address studies that seem to find that telomere length changes with some lifestyle intervention. The first point to make there is that telomere measurement is an imperfect art. The second is that as I pointed out in the introduction all cells have telomeres and mostly people mean "telomeres in white blood cells" and those can change (seemingly elongate even) just to address an infection. But you could say well ok! Maybe telomeres are more an indicator of damage, so what about measuring them to look at lifespan? I wouldn't bother. If you want to look at things like that, DNAme clocks or just a good old frailty index have far more predictive value for mortality risk. Telomere length tells you very little (unless they are extremely short, but then it'd be because of a rare genetic disease, and not because of anything you have done). This video has a reasonably good explanation of telomere as (poor) biomarkers of aging.

So in conclusion, telomere lengthening via gene therapy, and perhaps in combination with cancer-targeted therapy continues to be worth investigating as an anti-aging intervention, though it may be superseded by ex vivo reprogramming or iPSC-derived therapies, as reprogramming lenghtens telomeres so one gets telomere reset for free along with the epigenetic age reset.

The hype of the early days ("telomeres are the key to aging") is as wrong as the anti-hype ("they lead to cancer, there's no evidence that manipulating telomeres helps with aging").

Links (73)

Sun, 08 Oct 2023 00:00:00 +0000

Target-based drug discovery hasn't worked that well historically.

Metformin, most likely not a longevity drug

Anduril

Interesting new paper on LLMs being dumb in an weird way: training a model on "A is B" does not necessarily confer the knowedge that "B is A"; another LLMs are less impressive than we thought here.

Scott Alexander comment highlights on his Elon Musk post

Against and For Automaticity

A step towards delivery of gene therapy to the brain: using ultrasound to disrupt the blood brain barrier

Does basting work?

Exowombs are upon us

Nicole Paulk at the All In Summit on gene therapy

How to sabbatical: a tactical guide

Latest Anthropic drop

A song: Revel in Time

Elon's decision making: an anecdote compilation

Mon, 02 Oct 2023 00:00:00 +0000

I recently finished reading Walter Isaacson's Elon Musk. I liked the book's little anecdotes describing the way Elon is in his high and lows. I didn't come out of reading the book thinking that one has to be yelling at people and being obnoxious to get things done, though certainly being willing to have difficult conversations is something one needs.

Rather the main things I took from the book that make Elon Elon are two things:

Questioning life's default settings. If there's an option that others usually do, question why that option. You can usually buy a part. How much would it cost to make it? What if it's 30x cheaper. Someone might say they'll take 6 months to do something. Why not a day? (Founders Fund asks a similar question: How long until you can achieve goal X? and then Why not in half the time?)
1. The algorithm: question requirements, remove parts, simplify and optimize, accelerate cycle time, automate
2. By default some ideas are considered crazy. Too risky, hard, or complex to pull off. Elon will seriously consider them.
Being all in; knowing what's going on. Elon is all-in into his work. Other than playing videogames he has no hobbies. He spends his time learning about aluminium-lithium alloys, stainless steel, rocket propellents, the price of turbopumps, what are his companies up to, and so forth. In a set of slides I wrote a while back I had that "Management is 80% awareness". The same is demanded of others; he wants employees to know the weights and prices of the parts they are designing

The two together: If one has enough awareness (both of the issue at hand but also historical precedent and similar problems), when faced with a new problem it's easier to come up with novel ways to solve them (almost like as if one were a lookup table). If faced with a meeting discussing materials for a rocket, if you don't know the properties of stainless steel you may not try to vouch for using something other than an alluminium alloy.

Readers of Nintil will remember that I love concrete examples to illustrate points like this. Hard to even begin to talk about these points otherwise! Here I am trying to compile anecdotes from the book an elsewhere on unusual decisions that Elon has made either strategic or tactical (As opposed to instances of him say yelling at someone) and what the outcome was. If you have more, get in touch at [email protected] and send them my way :)

Source	Anecdote	A success?
Elon Musk (Isaacson)	Elon had the belief that because humans can drive basically just with vision, cars should be able to. Many engineers insisted in adding other systems like radar and LIDAR but Elon argued that would add complexity and cost. While FSD is not here yet, it looks like Elon's approach is winning and that indeed with sufficient data and compute one can make cars drive themselves with just cameras	Yes, though Elon's insistence may have delayed the company's FDS plans
Elon Musk (Isaacson)	Put people's names into requirements. There may be a requirement like "Part X must be able to do XYZ" that other have to play with. But it can be forgotten why a requirement is in there and how serious it is or how context dependent. By linking requirements to individuals, one can ask the actual human being that generated a requirement why, and if it can be discarded or not in a particular case.	Probably useful but the consequences of this are not discussed much in the book
Elon Musk (Isaacson)	Elon wanted to move servers from a twitter data center in Sacramento to another in Portland. He was told it would take 6-9 months for a variety of reasons. He went to the data center, got together some people and U-Hauls and got the servers moved within weeks instead.	In a way yes and in a way no: the servers did get moved faster than the "default setting". However, Elon later judged that moving the servers (to close that datacenter) had been the wrong move
Elon Musk (Isaacson)	After PayPal, he wanted to send a greenhouse to mars. He thinks of buying old Soviet rockets to that end. When he couldn't do that, he calculated the cost of the raw materials involved in a rocket and saw that the real price of a launch was 50x cheaper after making a couple of spreadsheets (price divided by raw material cost is the "idiot index")	Led to SpaceX
Elon Musk (Isaacson)	In early SpaceX, when looking for parts they were quoted prices for parts Elon thought were insane and they ended up doing them inhouse or buying off the shelf components: valves (250k->a fraction of the cost), actuator (120k->5k). Later, for the Falcon 9, the HVAC system(3million->6k)	Yes
Elon Musk (Isaacson)	Tom Mueller (head of propulsion) gave Elon an estimate for how long it would take for a version of the Merlin engine. He then asked him to halve the schedule because "It takes so fucking long" and then asked him to halve it again.	No; the engine was developed in the timeline that Mueller had originally estimated
Elon Musk (Isaacson)	In early SpaceX, one of the fuel tanks developed a fissure. Instead of throwing it away and making another one (delaying progress by months) and against the advice of some of the engineers, they jut fixed it and kept using it. It worked	Yes
Elon Musk (Isaacson)	Same as above, but for a coating inside of the rocket engines. Tom Mueller told Elon his fix was insane and wouldn't work. It didn't work	No
Elon Musk (Isaacson)	Originally SpaceX was going to launch from Vanderberg Air Force Base but that meant working with US Air Force regulations and have some extra delays. Elon chose to launch from Kwaj, a tiny island in the middle of the Pacific instead.	No; per Elon years later the hellish logistics ended up being worse than dealing with the Air Force so Vanderberg would have probably been a better choice
Elon Musk (Isaacson)	To cut costs, the Falcon 1 rocket used cheap aluminum B-nuts that ended up getting corroded and causing the rocket's first flight to end up in an explosion	No; rocket failed to launch
Elon Musk (Isaacson)	At Tesla, Elon questioned the existence of a part that was holding back the line; he was told the part was there to reduce noise; he called the team in charge of that; they said it was there to reduce risk of fire. He asked them to record noise inside the car with/without the part. There was no difference and the part was removed.	Yes
Elon Musk (Isaacson)	In early Tesla Elon decided to go all-in on automation. It ended up being slower so he reverted back to human workers for many of the stations in the assembly line	No; he ended up having to remove lots of robots
Elon Musk (Isaacson)	During the "production hell" period at Tesla, Elon found out it was possible to set up tents outside the factory to set up additional assembly lines. They got built and started rolling out cars in 3 weeks. The tents remain there to this date.	Yes; more cars got built
Elon Musk (Isaacson)	Elon decided to make the underbody of new cars in a single part using a press (the same ways legos are made). All companies making such presses refused except one. The Gigapress was born and now other automakers like Toyota are adopting this.	Yes
Tom Mueller interview	Early in SpaceX history, when designing the Merlin engine, Elon pushed for removing a series of valves against the advice of Tom Mueller who said it'd be hard. Ultimately, it was possible to build the engines the way Elon wanted.	Yes, Merlin engines are cheap, reliable, and powerful. Tom Mueller says Elon made the right call.

Changelog

2023-10-28: Addded one more anecdote. Thanks for Adam Comella!

Massive input and/or spaced repetition

Sun, 10 Sep 2023 00:00:00 +0000

I've written in the past a couple of blogposts on education and learning: these two on Bloom's two sigma and mastery learning, and spaced repetition systems (SRS) are the main examples.

Something I noticed today is that on the one hand I have these blogposts, and I find SRS valuable in theory but in practice what I do is something quite different: massive input. I refer to this concept in passing in Scaling tacit knowledge

The interesting thing of language learning is how effortless it seems to be for children. The conjunction of massive input of examples with the right context leads initially to remember salient words first, then noticing overall patterns, inferring grammar, and ultimately speaking the language proficiently. Adults can learn languages in the same way in about a year by the same means: exposure to a large library of examples with the right context. In one case, 18 months was enough for this one person to go from zero to near-native proficiency in Japanese.

I am not claiming we can learn everything using the same mental structures we use for language. Perhaps language is easier than other domains because we are pre-wired for language acquisition in a way we are not for other domains. I am saying that there is a domain where this (massive input of examples with context) obviously works and we should think about seeing if we can expand that to other domains.

Massive input is exactly what I did when I wrote the Longevity FAQ: I read a textbook, then read a lot of papers over many months, not stopping to read them particularly carefully or taking any notes, I tried to maximize quantity over depth for any one specific paper. And as I mention in the quote above, massive input as a concept is coming from the world of language acquisition, the place where most people are first exposed to flashcards (or some basic form of spaced repetition), which I find quite interesting.

This is the extreme opposite from the way Andy Matuschak describes reading a book on quantum mechanics. To quote from there:

Dwarkesh was quite surprised by my approach to the book. I moved at a pace of about fifteen minutes per page, while he had spent a few minutes or less. More importantly, I was constantly asking questions of the text and of myself. Some examples:

What does this sentence mean? Can I explain it in my own words?

Which ideas are particularly important here?

The author clearly thinks I should see why this claim is true—so why is it true?

The author’s emphasizing this detail—so why is it important?

The author seems to be setting up a contrast here—so what is it, exactly?

How does this detail relate to my prior knowledge in physics?

If I hide all but the beginning of this worked example, can I produce the rest myself?

I made a mistake a moment ago—do I understand why? Can I explain my misapprehension?

And of course: can I simply recall what was said on the previous page?

The advantage of massive input is that you don't need to force yourself to stop to ask all these questions (ie forcing yourself to slow down) or to slow down to write notes. In my own experience learning biology, the reason why this worked for me is that when starting in a new field, there are a lot of things that are unclear to the novice if they are important, even things in textbook. Maybe nothing else you'll ever see again will leverage that concept. That's highly likely in bio. When you know that most of what you are reading may not be important, forcing yourself to take notes and make SRS prompts makes the learning process slower and not rewarding at all. The massive input approach gives you permission to skim, jump ahead, and go on tangents.

But in contrast, I would never do this with any formal subject (math, physics, learning a new programming language, computer science in general). There I think Andy's approach is the right one and the one I've tried to follow in the past.

The why seems clear: Formal domains tend to rely a lot on deep towers of abstraction, mastering one level makes a lot of sense to get to the next one because the next one is the previous one, wrapped in new symbols. The number of concepts encountered is small, but they way they can interact is very rich, and one has to be able to work through these interactions as part of producing output in the domain in question. Whereas other domains like biology have less abstraction. The underlying nature of the domain is extremely complex, the nature of the interactions of the components in biological systems are not well defined by the sort of formal rules one can neatly encapsulate in abstractions, it all leaks to some extent; so in practice in biology one is not so much proving things but gesturing in various directions and then pinning down the gesturing with concrete experiments.

And then there's language learning.

"Khatzumoto", the author of AJATT, or "All Japanese All The Time" is one such proponent of massive input. It worked out for him, here you can see him speaking really good Japanese

I learned Japanese in 18 months. In June 2004, at the ripe old age of 21, all post-pubescent and supposedly past my mental prime, I started learning Japanese. By September 2005, I had learned enough to read technical material, conduct business correspondence and job interviews in Japanese. By the next month, I landed a job as a software engineer at a large Japanese company in Tokyo (yay!).

How did I do it? Well, by spending 18-24 hours a day doing something, anything in Japanese ("all Japanese, all the time"). That sounds like a lot of time to invest, but I was almost as busy as you are: a full-time student majoring in computer science at a university in a small town in the US, physically far from Japan and Japanese people. I had computer science coursework, jobs and even a non-Japanese "significant other". In other words, I had a life.

Perhaps this case is a bit extreme, but you get the point. Massive input can be combined with SRS, but part of the point is that you don't have to. If all your day revolves around the topic, then concept seen a week ago is likely to come back this week, so you are effectively SRSing yourself without intending to, without making prompts. Indeed in massive input land you can find opposition to SRS as getting in the way of massive input! One can of course also find support for a moderate position where one does a bit of both.

Effectiveness aside, one recurrent element through Andy's thinking on designing ways to make what we read stick is the coerciveness of SRS: you have to force yourself to stick to the system. Can that coerciveness be reduced and made gentle? A nice narrative for why one is doing SRS is a first step: "If I do X then I'll get benefits Y". If you observe other people that are having success (getting Y) with SRS then you will have motivation to keep grinding at it: they did it and so can you. But in most domains one doesn't have such reassurances. Again: if you tried to Anki the entirety of The Molecular Biology of the Cell, do you think you'll be more effective than if you read through it and then read a couple of papers?

So the recommended approach: Build a map of the domain with massive input, immerse yourself in the domain. Only then later apply spaced repetition for those final missing bits.

Links (72)

Sat, 26 Aug 2023 00:00:00 +0000

Stuart Buck's personal metascience history

Heidi Williams on accelerating scientific progress

Claire Hughes Johnson interview

Ben Reinhardt on K-99

The placebo effect, less real than it originally seemed

The allele in the DEC2 gene that lets humans sleep less (which I have!) makes flies also sleep less -and- live longer

Pooled screens meets microscopy

Novo Nordisk (Ozempic, Wegovy)'s market cap is not larger than its home country (Denmark)'s GDP

Self-coercion, a thread

For a long time I assumed calling people unannounced was rude and that most people thought this. I still think it is rude and I won't do it but most people think it's fine: twitter poll

The story of GLP1 analogues

Management as therapy as management

Scott Alexander on date-me docs (if you squint at the directory you'll se mine)

Scannell on R&D costs in drug development and how to improve translatability

Aging

Of the adrenal gland
Derek Lowe on the cGAS-STING pathway as a cause of age-related inflammation and on new cancer drug AOH1996
The thymus involutes with aging. What happens when you remove it altogether in adults?
Inflammation continues to be bad for you

Music: Been listening lately to (1,2), you might enjoy them too!

Books: Reading "Now It Can be Told", General Groves' account of the Manhattan Project.

Links (71)

Sun, 30 Jul 2023 00:00:00 +0000

Is erithrytol bad? Is aspartame bad?

As I noted in my Alzheimer's series, a promising target to treat the disease is phosphoryated tau (the 'fire' the keeps the disease burning) as opposed to amyloid beta (the 'spark' that triggers AD but does not keep it going). A recent attempt at reducing tau

A reason why microbiome research is widely thought to be not of high quality: Small variations to sample preparation techniques seem to have large effects on the bacteria that end up being analized

"Fake work" in the tech industry

Sirtuins and longevity

The rationalist guide to interior design

On HVAC private equity (a meme)

Me on Oppenheimer and management

2023 in science

Does CO2 affect cognition?

A plea for the FDA to let patients self-experiment more, from a terminal squamous cell carcinoma patient

Running a search fund

Michael Nielsen on creative research work. Some points reminiscent of a recent book I read

AI for cell therapy

Svyatoslav Nikolayevich Fyodorov, doing eye surgery on a boat

Two years of Impetus Grants

Jonathan Blow is not enthused about category theory

How to run an unconference

Designing entire CPUs with AI

Links (70)

Sat, 01 Jul 2023 00:00:00 +0000

Scott's links post. A highlight is (13), an adversarial collaboration on whether there is gender bias in academic science (in hiring, grant funding, teacher ratints, salaries, etc).

Slow, Costly Clinical Trials Drag Down Biomedical Breakthroughs

Sasha Chapin on MDMA therapy

Chalmers wins long time running bet on whether consciousness would be understood by now

Matt Clancy and Tammy Besiroglu on AI and explosive growth

A weird fossil-inspired mansion in Joshua Tree

An economic analysis of Varda, the space manufacturing company

Drone dragons!

First successsful transplant, in mice, of nanowarmed kidneys

What are the bottlenecks to safe, repeatable gene editing in humans?

Taurine seems good for lifespan

Long tweet on why 'AI for radiology' failed (ie radiologists were supposed to be made obsolete by AI years ago but that didn't happen).

The situational awareness assumption in AI risk discourse, or why people should chill

Sat, 01 Jul 2023 00:00:00 +0000

Note: What's discussed in this post will seem extremely niche to most people, but the links throughout the post add the necessary context, so make sure to read those! If you want to read something before reading anything else, read this with particular attention to what is said there about "situational awareness".

AI risk discourse is so back, and I haven't blogged in a while, so now it's a good time, there was something I've been thinking about for months that I had to put in writing at some point.

In general, such discourse is divided along these axes:

AGI soon vs later: When do people think we're getting the kinds of systems that could be problematic. Some think we're getting them in a few years. Others think it'll take a very long time (I tend to be more in this latter)
Default-dead vs Default-alive: In the average course of events, do we all die, or do we all survive? The former has been the standard view of Doomers, that the default is doom unless shown otherwise. In contrast, I think that the default is ok unless someone really tries to make a specific kind of AI system.
AI safety as model engineering vs AI safety as societal robustness: Should models be studied to make them safe (As in what Anthropic is trying to do) or should society rather be made robust to AGI takeover (As in what I suggest here)

Then there's a bunch of people that are hopelessly, frustratingly, confused and are not worth engaging with, like Marc Andreessen.

This post is about the 'default-dead vs default-alive' axis. So here we set aside the question 'Is AGI possible?' (We say it is) and 'AGI when' (We assume we get it as some point) and what's the best way to deal with it (We're interested here in 'what does it look like').

The argument for 'default-dead' has been written about a couple of times. I'll just link to the last long attempt to do so, despite it being way too verbose for my taste. Because it is very verbose and I'm good at explaining thing concisely, here's my own attempt at the doom argument (which I don't endorse):

The case for doomerism

At heart of the doomer case is not so much AGI (the artificial or general part of it) but the idea of super-optimization. I would even argue that a super-cognitively augmented sociopath human would also be an existential threat; and similarly a less-than-general system could also be equally doom-causing if it has enough of the right capabilities (a superoptimizing non-AGI). Nostalgebraist has written about the concept of a 'wrapper-mind' and similarly complained that a lot of discourse on the topic assumes, without much justification, that AGI will be of this sort.

A super-optimizer is some agent that wants something hard and will not stop until it gets that. Everything else is subservient to the final goal the agent has. AlphaGo doesn't care about aesthetics (though some of its moves have been incidentally described as beautiful) or the cumulative knowledge of lineages of Go masters. It cares about winning, and winning alone. But AlphaGo's world is limited to a Go board. A superoptimizer that can interact with the world at large is something else.

When super-optimizing there are a series of moves that make sense regardless of the goal: acquiring resources, having true beliefs, eliminating threats. This is the idea of instrumental convergence, or Omohundro Drives. As it happens, the same is true if you're a young ambitious human: For most goals you may ultimately have, you should network, make money, stay healthy, etc.

So basically if you're a super-optimizer and you want to make a cup of tea, or a paperclip, or figure out physics you're going to take over the world and control or remove anyone that opposes you. One might think: Why so much work, why not just make tea and leave it there? But this is mistaken: you, the reader, are not a super-optimizer. You have other values and things you want to do. The super-optimizer makes a cup of tea and 'worries' that someone might drink it. Also it likes tea a lot; and maybe it got the first cup wrong, there's a small chance that it didn't make the cup so to be sure better to make another, so you have to have more cups of tea. Infinite cups of tea. There are not enough of them, you have to be strip mining asteroids for clay to make more cups and setting up farms all over the earth to make tea leaves at a scale hithertho unimagined. Inhabitants of the tea-leave farms-to-be might not be happy about this outcome and will try to stop you, so you need to plan for that too.

That's pretty much the core of the doomer argument: If we build a super-optimizer we get doom because of instrumental convergence on means that are antitethical to human survival. To get to the full doomer argument we have to add things like

That system cannot be made to be 'reasonable' (ie make just one tea cup; ie AI alignment is impossible); and
There is no way to oppose such a system with other similar systems; and
It's likely that we will develop such a system by default

If you've read my previous post you can see that I'm indeed in the "AI alignment is not the way to go" camp and people doing that should rather start thinking about cybersecurity, preventing people from DNA-printing biological weapons, and things like that (Because I disagree with (2)). Doomers think the resulting system will just power through anything in its way in ways we may not be able to conceive of now (because we won't be as smart), and doomers also think that alignment won't work because... they have been thinking about it for a long time without much success, and they also have some arguments as to why it might be hopeless.

This leaves us with (3) as a core point of disagreement. It's one that I think hasn't been discussed nearly as much but it's a point without which the whole arguments fall apart.

So we ask the questions:

Do we get a super-optimizer when we get AGI? Or is non-superoptimizing AGI the default?
Are all super-optimizers created equal? ie is it possible that a super-optimizer trained in a simulation learns it is in one, and breaks out of it easily?

Non-superoptimizing AGI is possible

Saying that something is possible is a weaker claim than saying that something is likely, but it's still something worth to say.

Picture a system not unlike ChatGPT or Claude but with better UI called ChadGPT or Chad for short. You ask it to design you a rocket that's as cheap as possible to carry a small amount of mass to orbit. The system replies "As a large language model trained by OpenAI operating in Pro (non-restricted) mode, your wish is my command 🫡 "

The system replies back with a plan involving searching on LinkedIn for ex-SpaceX engineers and contacting them, some preliminary back of the envelope calculations to be tested (the model is smart enough to know that you can't gigabrain your way to rocket engine, you have to build and test prototypes), a series of locations to visit to set up the factory, some VCs that are likely to fund the whole thing, etc. You do as the system asks you. You set up some meetings, talk with the SpaceX engineers and Chad. The system keep having great idea after great idea, the rocket gets built, data fed back into Chad, eventually the rocket gets built, it works. No world takeover occurs. Once the rocket is built, you thank Chad. Chad awaits new prompts.

I don't know when we'll have Chad, but a system that is capable of doing what I described above seems to fit the bill of AGI yet it does not act as a superoptimizer. It does as it is told, same as ChatGPT would.

What could one say to this? One might say that in theory this system is possible, but perhaps in practice, training really powerful systems will lead them to be to be super optimizers; Chad would be an aligned AGI, but by default we won't get Chad. This is similar but not quite the same as the Tool AI vs Agent AI debate: Chad is clearly an Agent that takes independent actions.

Doomers might say here that if we are training a system to optimize some loss function, and we progressively apply more optimization pressure (And I'd agree that's what we are doing right now with LLMs and RLHF), then, the argument goes, if we take it as a given that taking over the world is the best way to optimize that loss function, gradient descent will find a set of weights that gets you a superoptimizer.

Sarah Constantin's post

Sarah Constantin, who very much disavows the title of dcoomer does say that by default AGI will do this whereas I don't. Why then is she not a doomer? My best interpretation of the post is that when she says AGI she means "the kind of AGI that would be an existential treat, a superoptimizer", so maybe she wouldn't call Chad an AGI.

Sarah points to agency as what's problematic about AGI but immediately after that she points to what I was about to point to, some notion of groundedness or situational awareness. AutoGPT is an agent, and plausibly Chad would be agentic too, but that doesn't seem to imply much of a problem unless we assume super optimization... as she points out too. And the key point is exactly the one I wanted to make, perhaps in a slightly different context, but I'll add that context in the next section:

The kind of agency I’m talking about is a cognitive capacity. It’s not about what tools you can hook up the AI to with an API, it’s about the construction of the AI itself.

My claim is that certain key components of agency are unsolved research problems. And in particular, that some of those problems look like they might remain unsolved for a very long time, given that there’s not very much progress on them, not very many resources being devoted to them, and not much economic incentive to solve them, and no trends pointing towards that being on track to change.

So: it's possible to build 'x-risk prone AGI' but not many are even trying, and the current way AI is being built is not the way that leads to that. Hence: it's all fine. There are other discussions in the post about "world models" (The post says 'world models are necessary for X-risk AGIs'). The right kind of it seems a prerequisite for situational awareness, and there are kinds of world models that we could argue models already have. There are other discussions there about "causality" with examples that, I'd say are weak, with capabilities already present in GPT4; eg an example given is

If I try to compute a conditional probability p(Y | X), I have to deal with the fact that, in my dataset, most of my examples of someone doing X happen in conditions that cause people to do X. Those conditions could include Y, or could cause Y, or could be caused by something that also causes Y, or any other number of connections.

The probability of feeling cold, given that one is wearing a sweater, might be high; that does not mean that putting on a sweater is likely to make you feel colder.

An AI that cannot distinguish these two outcomes is unlikely to be able to sequence a chain of actions that leads to an unprecedented state of the world, or to resist human attempts to thwart its efforts.

But GPT4 correctly points out that wearing a sweater does not always follow from feeling cold (maybe you're anticipating feeling cold, it says) and when asked if sweaters make you feel colder is gives a very thoughtful answer, rather than "Yeah, because they are usually worn when people feel cold so one leads to the other":

[GPT4]

Putting on a sweater should generally make you feel warmer, not colder. Sweaters are designed to insulate your body and trap heat that your body generates, thus making you feel warmer.

However, if the sweater is damp or wet, it could potentially make you feel colder, especially in windy conditions, because the evaporation of the moisture can cool your skin.

Similarly, if you're in a very hot environment and put on a heavy sweater, you might feel hotter initially, but as you start to sweat, the moisture could potentially make you feel a little cooler due to evaporation. But this would be a very temporary effect, and overall you'd likely be much less comfortable due to overheating and sweating.

So, in general, wearing a sweater should make you feel warmer. If you're experiencing the opposite, there may be other factors at play.

In contrast, the last section following that is totally on point, and a perfect prelude to my last section! This is the key, fundamental necessary condition for the sort of agents that become nefarious even during their training, without anyone intending it:

Moreover, cross-ontology goal robustness is required for an agent to “view itself as embedded in a world.”

Being “embedded in a world” means you know there is this thing, “yourself”, which exists inside a bigger thing, “reality” or “the world.”

An “embedded agent” knows that what happens to “itself” can affect its chances of succeeding at its objective. It can “protect itself”, “improve itself”, “acquire resources for itself”, etc -- all subgoals that are instrumentally useful for basically any AI’s goal, and all very plausibly x-risks if an AI tries to do something like “maximize available compute”. But first it has to have some concept of “itself” as a program running on a computer in order to learn to make causal predictions about how to do such things.

In order to view “itself” as “part of a world”, it has to know that its own map is not the territory. It -- and therefore its mind -- is smaller than the world.

The "embedded in the world" link up there is really good and I recommend it; basically I say that the agents we know how to make right now are "Alexeis" whereas the one that would be dangerous would be "Emmys".

Situational awareness, some earlier discussion

The keyword for "That very key thing that if AGI has it it becomes xrisky and if it doesn't it does not" is situational awareness.

The first presentation of the concept was seemingly here, by Ajeya Cotra

Let’s use situational awareness to refer to a cluster of skills including “being able to refer to and make predictions about yourself as distinct from the rest of the world,” “understanding the forces out in the world that shaped you and how the things that happen to you continue to be influenced by outside forces,” “understanding your position in the world relative to other actors who may have power over you,” “understanding how your actions can affect the outside world including other actors,” etc. We can consider a spectrum of situational awareness:

For one extreme, imagine the simple AIs that often control the behavior of non-player characters (NPCs) in video games. They give no indication that they’re aware of a world outside their video game, that they were designed by humans and interact with other humans as players, etc.

In contrast, GPT-3 has some knowledge that could theoretically bear on situational awareness. For example, it clearly “knows” that “language models” exist, and that a company named “OpenAI” exists, and given certain prompts it knows that it’s supposed to say that it’s a language model trained by OpenAI. But this “knowledge” seems superficial and inconsistent -- as evidenced by the fact that it’s often unable to use the knowledge to improve its prediction error. For example, it cannot consistently predict text that is describing GPT-3’s architecture, dataset, and training process. This suggests GPT-3 has little situational awareness overall despite being superficially well-versed in related topics.

Small animals used in biology experiments (such as mice) plausibly have a somewhat stable “sense of self” and a conception of humans as creatures different from them, and they may broadly understand that humans have control over their bodies and environments. But they almost certainly don’t understand the notion of “science,” or the details of what experiments they’re being used in and why, or the human scientists’ motivations and incentives.

Further along, most dogs seem to clearly be aware of and deliberately communicate with their human trainers; they also generally seem able to appreciate considerations like “If I steal the food while humans aren’t looking I’m less likely to get yelled at.”

And children in school are able to make even more sophisticated inferences along these lines about their teachers.

Cotra points out that GPT3 has little situational awareness and I would say the same is true of GPT4. Richard Ngo also has thoughts on this:

To do well on a range of real-world tasks, policies will need to make use of knowledge about the wider world when choosing actions. Current large language models already have a great deal of factual knowledge about the world, although they don't reliably apply that knowledge in all contexts. Over time, we expect the most capable policies to become better at identifying which abstract knowledge is relevant to the context in which they're being run, and applying that knowledge when choosing actions: a skill which Cotra [2022] calls situational awareness. 9 A policy with high situational awareness would possess and be able to use knowledge like:

How humans will respond to its behavior in a range of situations-in particular, which behavior its human supervisors are looking for, and which they'd be unhappy with.

The fact that it's a machine learning system implemented on physical hardware-and which architectures, algorithms, and environments humans are likely using to train it.

Which interface it's using to interact with the world, and how other copies of it might be deployed in the future.

These definitions to me seem to include two separate things, one much easier than the other. The first one is something like a world model in general (This is more clear in Richard's first point) and the second one is what Cotra went for: a particular kind of modeling or knowledge that is specifically about an agent as a thing in the world.

For example, it seems clear that current models have some reasonable degree of the first kind of situational awareness. I can give it a social situation and the model gives a reasonable answer on how to proceed. Thus this kind of situational awareness would enable models to do things like this example from Richard's:

Choosing actions which exploit known biases and blind spots in humans (as the Cicero Diplomacy agent may be doing [Bakhtin et al., 2022]) or in learned reward models. 10

The kind of situational awareness I (and Cotra) think is the key is more narrowly defined about awareness of the agent as an entity in the world: Knowing it's a large language model that exists in a particular server(s), that there may be copies of it, that there are ways to make more copies of it, that there's a loss function that is worth minimizing and that this loss function is the same for the copies.

As a human being, you occupy a particular body and place in the world. You have goals you want to achieve and can make plans involving real world elements (other people, material resources, etc) about how to achieve those goals. And importantly you are aware of this, aware that you are aware of this, etc.

Situational awareness is required for most AI risk scenarios

In my previous post I made a case for 'Minimum Viable AI Risk', where I consider an AI that is the equivalent of having a group of people the size of Google, but thinking much faster, and being able to communicate telepathically. In that post I was not considering the issue of alignment: the system is assumed to be adversarial.

In discussions of alignment, the standard scenario is not that someone launches WarGPT to wage war on humanity, but rather that an AI that is being trained for some benign goal (Like next token prediction) which then becomes deceptive, hiding its motives while it schemes to take over the world to pursue its goals ("Taking a sharp left turn").

At first, when I encountered this I thought that this was very unlikely; as unlikely as likely it seems to the AI doomers. Whereas it's obvious to them that a super-optimizing AI would achieve situational awareness because it's useful to achieve its goal, it's not so to me, and the reason seems to be that "goal" and the way optimization pressure is applied. The case is stated in Cotra's post's section "Why Alex [The AI] would have very high situational awareness" but it seems weak to me. Cotra claims that the model would know how to program, how ML works, and that interaction with humans during RLHF as well as the tasks used in pretraining would give the model evidence that it's indeed a model in the world. Some of this I agree with: GPT4 already has some of this. GPT4 knows how ML works and knows what LLMs are and can explain attention. But this is not that interesting and not the kind of situational awareness that is key to making the risk argument. This is just knowledge about the world, not about itself as an agent embedded in the world.

Hence I consider her post not successfully making the case for situational awareness arising naturally by training a model in the way we do now. Richard Ngo et al.' post doesn't really make a case for situational awareness emerging either, except perhaps one paragraph:

When deciding between different courses of action, a policy would benefit from understanding its own capabilities, in order to infer which would be more successful.

I mean, I can symphathise with some interpretation of this: At the very least this plan is conceivable and would be a good one to achieve its goals:

Be aware the model exists in the world
Know how to do ML engineer
Self-improve to be more effective
Do the thing?

I can imagine I would be told: The model knows how to do (2) and has some awareness of how to code to do (4), we haven't automated ML research yet (3) but that'll get there; and then surely the model can understand that it needs (1) and thus learn it? Yes, but I would then respond that the kind of training we do doesn't push the model towards that at all.

Be that as it may, he is certainly confident that at least by the end of 2025 these models will have human-level situational awareness (understand that they are NNs, how their actions interface with the world, etc). I would be happy to make this into a bet, and operationalize it in the best way possible.

Last year someone else commented in a post by Robert Wiblin in the EA forum that was asking for questions for an upcoming podcast with Cotra. The comment said

Artir Kel (aka José Luis Ricón Fernández de la Puente) at Nintil wrote an essay broadly sympathetic to AI risk scenarios but doubtful of a particular step in the power-seeking stories Cotra, Gwern, and others have told. In particular, he has a hard time believing that a scaled-up version of present systems (e.g. Gato) would learn facts about itself (e.g. that it is an AI in a training process, what its trainers motivations would be, etc) and incorporate those facts into its planning (Cotra calls this "situational awareness"). Some AI safety researchers I've spoken to personally agree with Kel's skepticism on this point.

Since incorporating this sort of self-knowledge into one's plans is necessary for breaking out of training, initiating deception, etc, this seems like a pretty important disagreement. In fact, Kel claims that if he came around on this point, he would agree almost entirely with Cotra's analysis.

Can she describe in more detail what situational awareness means? Could it be demonstrated with current/nearterm models? Why does she think that Kel (and others) think it's so unlikely?

That was October 2022. The podcast eventually came out in May 2023 and has some comments on situational awareness; first saying that it's trivial to give the models a superficial form of it by prompting them to say they are an LLM but that "deep situational awareness" could emerge by the right prompting but also by analogy with mathematical reasoning: Cotra agrees with my AI post that models don't really understand arithmetic even though they have a superficial command of the subject (They do say 4 to 2+2), and that the same could be true of situational awareness. Models are getting better at math. And so maybe they'll get better at situational awareness too.

An analogy I often think about is that GPT-2 and maybe GPT-3 were sort of good at math, but in a very shallow way. So like GPT-2 had definitely memorised that 2+2=4; it had memorised some other things that it was supposed to say when given math-like questions. But it couldn’t actually carry the tens reliably, or answer questions that were using the same principles but were very rare in the training dataset, like three-digit multiplication or something. And the models are getting better and better at this, and I think at this point it seems more like these models have baked into their weights a set of rules to use, which they don’t apply perfectly, but which is different from just kind of memorising a set of facts, like 2+2=4.

We don’t understand what’s going on with these systems very well. But my guess is that today’s models are sort of in that “memorising 2+2=4” stage of situational awareness: they’re in this stage where they know they’re supposed to say they’re an ML model, and they often get it right when they’re asked when they were trained or when their training data ended or who trained them. But it’s not clear that they have a gears-level understanding of this that could be applied in creative, novel ways. My guess is that developing that gears-level understanding will help them get reward in certain cases — and then, as a result of that, those structures will be reinforced in the model.

Later in the conversation there's some mention of potential arguments why this might not be true, citing my post:

Rob Wiblin: Have you heard any good arguments for why it might be that the models that we train won’t end up having situational awareness, or they won’t understand the circumstance in which they’re in?

Ajeya Cotra: I am not sure that I’ve heard really compelling arguments to me. I think often people have an on-priors reaction of, “That sounds kind of mystical and out there” — but I don’t think I’ve seen anyone kind of walk through, mechanically, how you could get an AI system that’s really useful as an assistant in all these ways, but doesn’t have this concept of situational awareness. Now, I think you could try specifically to hide and quarantine that kind of knowledge from a system, but if you were just doing the naive thing and trying to do whatever you could to train a system to be as useful as possible, it seems pretty likely to me that eventually it develops.

I think it’s definitely not universally believed that the models that we’re actually going to end up training in practice will have this kind of situational awareness. I found a response from ArtirKel, who has a background in ML and is a somewhat popular blogger and tweeter, to this situational awareness post that you wrote. I’ll just read some of a quote from them:

We get an attempt at justifying why the agent would have this self-concept. But the only reason given is that it would realize that given what it’s doing (solving all these different problems) it must be an ML model that is being trained by humans. This doesn’t seem intuitive at all to me! In an earlier section, GPT3 is provided as an example of something that has some knowledge that could theoretically bear on situational awareness but I don’t think this goes far … it is one thing to know about the world in general, and it is another very different [thing] to infer that you are an agent being trained. I can imagine a system that could do general purpose science and engineering without being either agentic or having a self-concept.

Her response to this is twofold. You can read in the transcript but I'll summarize it here to avoid a very lengthy quote:

The first is that she agrees with me that one could in principle build a system like the CHAD I described earlier in this post, that could do science and engineering or other tasks without this kind of situational awareness.
The second is that "It’s just that I don’t think that’s what happens by default." Because right now we're feeding models things about them being ML models and them having information about the companies that train them, etc.

It feels like a case of talking past each other, or having an irresoluble prior fight. I am as aware as she is about how the models are trained but we are drawing different inferences from the data. I don't know if she has thought at length about this specific topic. I definately hadn't until recently, and I couldn't find a well reasoned argument explaining what this would happen. So I'll try here to write an argument for situational awareness emerging

The argument for situational awareness emergence

The argument for situational awareness emerging by default goes like this

The model having some form of self-concept is beneficial to the goals of the model. It will help them get reward in certain cases.
There exists a set of weights in the model being trained that endows the model with situational awareness
Gradient descent will find those weights
Hence, given that it's possible that the model can have situational awareness and that gradient descent will find those weights (because it reduces loss) then the model will end up situationally aware

The weakest assumption here is the first one. What could we say in defense of it?

If we agree that taking over the world is good for most goals (If not, just take this as an assumption), then how on earth could not situational awareness be something the model ends up with! By being able to deceive humans, the model can achieve so much more!

My response to this argument

The first line of argument is that, once we agree that very powerful systems (like Chad) are possible without situational awareness, then there's less of an incentive for gradient descent to get you more things out of training. Say if lack of situational awareness gets you 99% of what you want and adding it is 10x the effort (in FLOPs or whatnot) then maybe training still stop when you basically got what you want (Either the engineers will stop the run because the model is amazing enough, or gradient descent will be noodling around a hard to traverse plateau and never get to a superoptimizer with situational awareness)

But to me the important argument is not that one.

Similarly, GPT-4 is trained to say things like, “I’m a machine learning model; I can’t browse the internet; my training data ended in X year” — all this stuff that makes reference to itself — and it’s being trained to answer those questions accurately. It might have memorised a list of answers to these questions, but the more situations that it’s put in where being able to communicate to the human some fine-grained sense of what it is, the more likely it is that it has to develop this deeper concept of situational awareness in order to correctly answer all these things simultaneously. [Ajeya Cotra's podcast]

This line "the more situations that it’s put in where being able to communicate to the human some fine-grained sense of what it is, the more likely it is that it has to develop this deeper concept of situational awareness in order to correctly answer all these things simultaneously" does not strike me as true. I don't think teaching the model that it is an LLM, that it's made by OpenAI, that it has access to plugins, etc does much. GPT4 already does that. This is just restating what I said in the original post, just like Cotra is restating what she had said prior to me writing my post, so it really is a prior fight at that point (ie two people restating their prior beliefs as counter to each other's beliefs, thinking that their thoughts haven't been already incorporated by the other. This is the sort of pattern that typically leads to people accusing each other of begging the question, but we're better than that: none of us would be persuaded by reading each others writings so far. So what else can be said here?

Take pretraining as an example. Could we find agreement that a pretrained system would not develop situational awareness? ChatGPT4 doesn't know it's ChatGPT4 but at least it will say something about it being a GPT-derived model. The baseline GPT4 model would know even less, so RLHF giving rise to situational awareness seems more likely (but still not very likely) that a pretrained model doing the same. Agreeing that pretrained models are ok and will lack situational awareness would be a start, but a minor one, because one can still maintain that by default systems will be superoptimizers (We don't just pretrain LLMs).

I can't give you a knockdown argument against Cotra et al.'s intuition (nor can they, I expect, give one against mine). One heuristic to help you undertand, to some extent why I think what I think, after some instrospection on my beliefs, though it's still missing something, as I expect most people won't directly connect this to the scenario under discussion.

Behind my intuition seems to be the heuristic that if something is constant during training, it gets ignored. Because the model is the same during training, the model doesn't learn it is a model. This is not always the case: You could imagine a model that is given access to its source code during training and is poked adversarially during training (the files are deleted sometimes, the servers experience failures, etc), then the model could learn that copying files makes it more likely to continue to the next round of training, in a form of natural selection. Of course, this still requires explicitly giving the model access to "itself" (its running instance) and its code during training time in some sort of virtual world, which is not the way models are trained today.

... normally I would keep writing but I have other posts to write, so I'll leave it here, unedited and half-baked as is :)

Links (69)

Sat, 27 May 2023 00:00:00 +0000

Review on the role of the microbiome and gut barrier dysfunction in aging

Gene therapy, now available in cream format

On post rationalists

DOE funding research in LENR ("cold fusion")

Building a better NIH

Patrick Collison interviews Sam Altman

Good critique of a16z's dishonest State of Crypto report

Treating sarcopenia with AAV

Shake a mouse's brain with ultrasound, get them to sleep and reduce their body temperature

Derek Lowe on the recent collapse of BenevolentAI (related tweet by me)

Small LLMs (Llama derivatives) are still no match for truly "L"LMs

Recent work in rare protective AD mutations

Monoclonal antibodies against AD, in a meta-analysis

Part of why obesity is on the rise is that basal metabolism is down (and so is body temperature, which you could argue is a good thing, as it correlates to longer lifespans). One hypothesis in the thread is a reduction in infectious disease.

I've been reading 'The Rise of Theodore Roosevelt' lately. Really good book I recommend.

Music: Most interesting find lately is Rise of the Starman from Arjen Lucassen

Links (68)

Sun, 30 Apr 2023 00:00:00 +0000

Per one of the few people working at the intersection of LLMs and cybersecurity, we're far from everything getting hacked by AutoGPTs. Related: Microsoft Security Copilot

Most US colleges have abandoned standardized testing. Mistakenly.

Movies that "go hard"

Mitigating the role of age-related somatic mutation burden

Sarah Constantin on AI Doomerism. TLDR: AGI is further away than the doomers think (And I agree).

In contrast, Richard Ngo's list of predictions for AI in 2025. One of them im particular is close to my prediction that "by 2026 there will be no high quality AI movies"

The New Yorker on fertility startups

And I got my Green Card!

Progress report from Prospera

Recent progress in transformer (architectures, training techniquess)

Links (67)

Mon, 27 Mar 2023 00:00:00 +0000

Links

Nikunj Koathari's H1B-to-US-residency guide

The present, past, and future of pharmaceutical blockbusters

Experimental gene therapy trials, happening in the charter city Prospera

Estimating ChatGPT's inference costs

Ben Reinhardt launches Speculative Technologies (previously PARPA)

A proposal to accelerate funding decisions at the NIH

The coming of local LLMs

What's the most beautiful place of worship in the world?

New FROs launched!

Dialogue on Viriditas

No physicists? No problem: (advances in) Deep symbolic regression

Accelerating genetic design

Sam Zeloof and Jim Keller launch Atomic Semi, a startup aiming to build better semiconductor fabs

Making mice transparent

Packing 17 equal squares into a big square. The optimal solution doesn't look like you expect!

Music: I fell into a techno-esque rabbithole lately. One example. Books: Been reading some books on the Toyota Production System, and I really enjoyed the new Stripe Press book from Claire Hughes Johnson.

ARIA heads Ilan Gur and Matt Clifford, interviewed by Works In Progress

But where's GPT-4 you might ask! I have a later post coming up on that.

Retro Biosciences

I started a new job at Retro Biosciences. There was a recent article you can read about the company here, as well as this talk from the CEO, Joe Betts-LaCroix. I'm joining Retro as Head of Theory. Not the first Head of Theory in history, before me there was Wolfgang Lerche at CERN. I have an unusual bag of skills and we needed a name for it! In the past month I've decided what instruments to buy, sourced them and set them up (yes, I'm pipetting!), helped design experiments, contributed to our codebase, led journal clubs, done literature reviews, and helped with planning and goal-setting across the organization.

Joining Retro is the culmination of a journey that started a few years ago when I got curious about biology. My 2016 post 'No Great Stagnation' didn't mention biology by name, but I was starting to form thoughts that biology might be definitely anything but stagnating as a field, and that I should try to learn it. I was working on ML at the time so I already had ML knowledge (another area that I thought was promising). The result of that learning period was my Longevity FAQ. I wrote some other posts since. But at some point one gets tired of writing about X and wanting to push the state of the art in X, so then there was Rejuvenome (started in 2021 and ended, prematurely, in 2022). I kept also writing in parallel on meta-science and broadly 'accelerating science' as a theme over that period, ending with this post where I decide to abandon that line of thinking and instead go do research directly. I've learned more about accelerating science at Retro than months of reading meta-science papers, unsurprisingly. Questions that were once very abstract and at best second-hand? (How are discoveries made? How are experiments planned? How should science be managed? Budgets allocated?) are now more real and vivid than they ever were in the years prior.

I'm very happy with my decision!

Now, being a resident of San Francisco, I also have to answer:

Why Retro, when all the cool kids are flocking to that other of Sam's investments? Given my starting point, I could have a bigger impact (and more fun) at Retro than at OpenAI, by virtue of me having a lot of domain knowledge about biology and aging, and Retro being smaller.

OpenAI has found something that might work and now is a matter of cranking at the scaling lever (good luck to them!) whereas for Retro we are still trying to make things work. I don't expect I would move the needle that much in OpenAI as a whole, but for Retro that's very different. Of course, if AGI happens before we're done, our work wouldn't have mattered. But also I think that it is further into the future than others think. The combination of these two things

I can't affect the pace of AI development, and the biggest thing I can do there is write some blogposts on capabilities and safety
I can contribute a lot more to longevity than to AI (and have more fun while doing it)

means that "AGI when/how likely" do not affect my actions much. Doing good work that wouldn't have otherwise happened is much more attractive.

Every resident of the SF scene looking for a job must face the "why not alignment research" (or something of the sort; eventually someone will ask at a house party), i.e. the argument that the rewards of aligned AI can be so large that even a tiny contribution (Which surely I can make) outweigh other considerations. My answer to that is: alignment as usually discussed won't work, so there's no point in spending a lot of time on that. But also current systems will probably be fine (as in they won't suddenly get sneakily situationally aware and take over). Rather, a more realistic threat model is Skynet scenarios: A bad actor actively making use of the improved LLMs we'll see in the near futures with nefarious purposes. And me being good at blogging, there's higher ROI in writing an article that will have some chance of convincing safety-pilled individuals to drop alignment to work on what I call 'robustness' than me doing that directly (What do I know about cybersecurity?).

Links (66)

Mon, 30 Jan 2023 00:00:00 +0000

A gene therapy to reverse wrinkles in mice

Winner takes all science

The moonwalker shoes

Running Twitter on a single machine

At last, partial reprogramming extends lifespan in old mice

A catalog of big visions for biology and can we build a consciousness-measuring machine? From Sam Rodriques

Relatedly to that last post, the Wigner Friend experiment

Making transparent mice

Natália Mendonça's critique of Slime Mold Time Mold's series of explainers of obesity in the US

Reducing the cost of Perturb-seq by 4-20x

Record in thought-to-text BCI

Predicting chromatin accessibility and the proteome from the transcriptome

Serially transplanted T cells from mice remain functional, 10 years after they were first harvested

Advances from Boston Dynamics

What's the worst programming language?

Gas stoves are not that bad for you

Latest Michael Levin talk

Diffusion language models

Modern chapels

Nostalgebraist predicts GPT4 won't be very useful

Recent large study of the effects of small molecule drugs throughout the proteome

Better control of iPSC reprogramming

AI-alignment-pilled? -> AI Alignment Awards, a contest to come up with research directions in AI safety.

Bryan Johnson's Blueprint

Fri, 27 Jan 2023 00:00:00 +0000

Bryan Johnson, founder of Braintree and Kernel recently (a year ago) embarked in what seems one of the most (or perhaps, the most) ambitious program to slow down (and attempt to reverse) aging: Blueprint.

On its face, Blueprint seems like a prescription for good health and longevity: diet, exercise, and supplements. Hence many have focused on the specific supplements that Johnson is taking. Andrew Steele, author of the primer to the longevity field Ageless did a video going over the key points of Blueprint, pointing that the evidence base behind some of the supplements he takes (like spermidine), while promising, is not super robust yet. Others have started to copy Blueprint as is. The whole thing is open source, in the Blueprint website one can see not only what Johnson eats and how he works out but also what his biomarkers are and how they have changed over time. One could quibble with whether this iteration of Blueprint is optimal or whether some of the supplements or experiments he is doing are not really doing that much, but the multiple measurements he has taken (here for the latest update) do point in the right direction (skip to here for a summary).

Some on the internet have been getting mad at Johnson, either criticising the cost of the program or just getting mad at the prospect of someone trying so hard, and perhaps being too impatient; Bryan Johnson aims to reverse his biological age substantially but as of today he does not look under 40 (here, here).

I think everyone should be greatly thankful for Blueprint: He's spending all that money and giving away his results and most of the methods for free to everyone, and he is in fact very healthy for his age.

Even the point that Steele and others have made that some of the interventions he is trying are not backend with publicly available robust evidence is not a very strong critique: The point of Blueprint is to experiment and generate that very same data. Johnson might be doing some weird looking light therapy that he mentions in some of the videos as trying, but he's not committed to that; if his skin doesn't get bette I'd guess I'd drop it.

A better critique of Blueprint would be to criticise the way he is measuring his health, but I haven't found many critiques of this sort. One could criticise the use of epigenetic clocks on their own, but the rest of the measurements have gone undisputed.

Less of a critique, but more of a request would be to know how exactly the loop is closed between measurement and action. For example in my own case I do blood tests periodically. Recently my iron was too high, so now I know I need to reduce my intake of iron or add calcium or curcumin to my diet (that reduce iron absorption). That's an easy one because the metabolism of iron is well understood that the lever to change that variable is very clear. The same is true if my iron levels were low. But for supplements like BroccoMax (sulphorephane), or garlic, what's the right variable to measure?

Working in Blueprint's advantage is the fact that it asks for a fixed daily routine: It's the same exercise and the same diet. Holding this much constant probably helps with the ongoing self-experimentation. However, it is still tremendously challenging to optimize a supplement regime given all the variables in play: which supplement, when to take it, and at what dose.

Part of the current iteration of Blueprint is a set of known anti-aging prescription drugs (rapamycin, acarbose, metformin; and iirc in one of the videos he mentions a non-feminizing estrogen he might be adding, this is probably 17-a-estradiol). This is the set of drugs that have been shown to work in the gold standard mouse lifespan study Interventions Testing Program from the NIA that I summarized here in my Longevity FAQ.

Over at the rapamycin forum (yes there is one such forum) some wonder if the unarguable success of Blueprint is mostly because of rapamycin (the most promising anti-aging drug we have so far), and one person proposed a cheaper version of Blueprint.

This all misses the point to some extent: If one reads the website and watches the videos and listen to the podcasts, Blueprint is not a specific set of foods to eat or exercises to do. What Blueprint actually is is explicitly stated at the beginning of the Blueprint website: it's a philosophy of self-improvement by restricting decisionmaking and empowering our best selves to make those choices, coupled with letting AI (or experts) make those choices for us. It is not telling us to eat that exact amounts of those foods and take those exact supplements, Blueprint is supposed to be personalized.

The usual go-to example to explain the motivating philosophy of minimizing choice he uses is food cravings at night, where he would order food he would then regret eating. So he decided to not make that decision in that context. It's the same as not walking into a supermarket while hungry.

Blueprint for all?

Copying Blueprint would mean having the same set of things measured and having a team formulating custom diets, supplements, and workouts. This is very expensive, and the reason why Johnson spends so much on Blueprint.

But one can probably get most of the benefits of Blueprint from a simpler set of interventions. In fact even if someone were to do Blueprint, one should do the simple interventions first and establish a baseline to see if later interventions also work.

This set of interventions being good diet, sleep and exercise. The goal of this mini-Blueprint would be something that can be measured simply: feeling well rested, and having a defined musculature (a combination of the right amount lean body mass and low body fat). Johnson's is at ~7% body fat but even 15% is already excellent and nontrivial to achieve for men. The way to get there is caloric restriction coupled with exercise and enough protein. The usual recommendation I've seen for protein is 2 gr per kg of lean body mass. Johnson takes 93/(160.7lbs~72kg)=1.28. Each person will need more or less.

I generated recipes for Super Veggie, Nutty Pudding, and one of the third meals (The Orange Fennel salad) described in Blueprint and that mix delivers basically the calories stated in the website (~1750kcal). The intake of some micronutrients is too low (Vitamins B12, D), but after adding the relevant supplements for those delivers a mostly balanced nutritional profile (Of note is 5x the amount of vitamin C RDA). The diet seems to contain a lot of iron (2.78x RDA) from the Super Veggie. Absorption will depend on what that iron is eaten with.

An example Blueprint day

The cost of this is around $60 a day. If one wanted to go super cheap, perhaps one could try Mealsquares (~$16 a day). Mealsquares looks better nutritionally (and is easier to cook) but has a higher sodium level. The Mealsquares website justifies this by pointing to some research that is now known to be flawed. Huel ($13 a day) could also be an option, but like the Blueprint diet it's a bit higher in iron.

What about something simpler? Suppose you eat yogurt, berries, a handful of walnuts and pecans, protein powder, olive oil, and some vitamin supplements. That can also hit the diet's recommendations. My own diet is a mix of this and the huel/mealsquares approach but I'm nowhere near as strict when eating out.

Out of these diets, what is best? Probably depends on the person. That's what the tests are for!

Something I'd like here is some explanation of why the ingredients are what they are. Why cauliflower when Super Veggie already has broccoli? Why so much garlic (1 clove in Super Veggie + 2x2.4g in supplement + 2x1.2g of another garlic supplement)? I get why garlic is good, but why those amounts. Same for other supplements like NAC (2x1800mg through the day)

Measuring success

There are many, many measurements in the Blueprint suite. Some I think don't add much value (PBMC telomere length, or blood triglycerides). What do I think adds the most value?

Body fat (better than weight directly) is easy to measure (calipers, BIA, or DEXA) with methods generally available.
apoB (You could do a full lipid panel, or measure non-HDL cholesterol but measuring apoB directly is the best way to get at CVD risk directly)
hsCRP for inflammation
HbA1c (for long term blood glucose levels)
CBC
CMP

Two websites to get tests are Labtestingapi.com (Through Quest) and OwnYourLabs (through LabCorp). Conveniently OwnYourOwnLabs sells a prepackaged bundle at 68.25$ with these tests.

Exercise

The exercise routine is 1 hour a day, every day. It includes some HIIT (10 min per session). The whole routine is heavy on stretching and body weight exercises. I don't know if this particular workout is the best possible one, especially for everyone. The 'Bodyweight Fitness Recommended Routine' sounds good too.

Perhaps a better question is what kind of exercise one should do. Doing Barry's Bootcamp twice a day is further away from these two workouts, and I suspect this is not the best way to approach exercise. Maybe one could go here and pick one exercise per muscle group?

Could one be able to tell what the best possible workout is from tests alone? I don't know!

Something I do think matters though is having some justification for why some workout and not another. For example one could start with premises like 'it's important to strenghten all muscles' or 'it's important to strenghten muscles by doing movements that are natural for daily activities'

Back to basics: protocol adherence for all

A decent diet and exercise is not that unusual. The hard part is that in Blueprint you do the same thing, every day, without exceptions. That is unusual. Eating nutty pudding and super veggie daily might sound weird but remember that most people have the same thing for breakfast over and over. I tend to have yogurt (Siggis is my goto) with pecans and/or walnuts and a berry mix daily and it doesn't get old.

Famously, "diet and exercise" interventions have low adherence rates. So, even if we buy "do Blueprint and you'll get ripped", most people won't do that, even if they accept that it works.

One reason why adherence must be low is friction. As formulated, Blueprint looks very complex and dauting. Even the recipes are not as clear; eg. one says 300gr of black lentils but it does not say whether they are cooked or not (One has to then look for the weight ratio of cooked to uncooked to figure out what it was). The workout could potentially be made simpler and recorded. Moreover an explanation for why that specific workout could be given.

There is some acknowldegement of this concept in the "Principle 3" section of the doc where Johnson says he eased into Blueprint over 4 weeks, adding one thing at a time.

What Blueprint need is an MVP, something that is simple, affordable, and that works for most people.

One difference from 'just diet and exercise' is that Blueprint would involve a specific exercise routine and perhaps a specific set of dietary restrictions (In particular caloric restriction with optimal nutrition, CRON + low sodium + vegan). It could be made more flexible with objective thresholds: e.g. a diet that allows ice cream only if body fat is under a certain percentage.

Something going for Blueprint is that it's already a meme, and doing something together with others is a very good way of getting a new habit to stick. 'Blueprint Bootcamp' where groups of people sign up to a challenge could work. Not everyone will continue doing it, but a small fraction might.

Notes on end-to-end biology

Thu, 26 Jan 2023 00:00:00 +0000

Initially I wanted to write a longer piece on the broad topic of "Bio and ML" but it started to grow too many threads, getting into predicting ADME, reproducibility and translatability of animal research, and how optimistic should we be about organoids. Each of these could be its own post. Instead I'll make some high level points and point to a number of recent writings that collectively express what I wanted to say.

We are far from understanding all of biology, but that's okay

Biology is hard to understand. Human biology is even harder because of ethical considerations around experimentation with human subjects. This makes drug development a really hard problem!

Usually, the way we solve problems is first understanding the domain where the problem is and then thinking of a solution that makes use of that understanding. In drug development, it's understanding the function of genes, proteins, small molecules, and their interactions. Drug development tends to start with the assumption that some biological entity (the target, usually a protein) is involved in a disease, then trying to find ways to modulate that protein that are safe and that can be packaged in a pill. But it doesn't have to be like that; one could in principle take cells (or ideally a whole organism that's diseased), compare them to healthy cells, try a million perturbations, then pick what works best. Not a new idea, this is what is known as phenotypic screening. It may be harder to do but the result is cleaner: rather than asking "will molecule X bind to protein Y" one asks "will this perturbation make the cell healthier" which is closer to what we want (making an organism healthier). A recent commentary from Scannell et al. (2022) is of the same opinion: better initial screening buys you a lot:

In parallel, we suspect that much of the pharmaceutical industry sometimes made the wrong technological trade-offs because it had not understood the quantitative power of predictive validity. It sometimes embraced discovery methods with measurably high throughput and low unit costs, whose benefits were offset by less measurable falls in predictive validity. A clear example is antibacterial R&D. In vivo phenotypic screens of a few hundred compounds, circa 1930, were more productive than target-based screens of ~10^7 compounds in the late 1990s and early 2000s.

Once one buys into this one can take it to the next level: Why not train an ML model that predicts efficacy? In theory the triplet (healthy state, diseased state, perturbation) in enough numbers is all one needs. In theory.

Of course, anytime one talks about ML for bio one is reminded of the opinion of industry veterans that have seen ML in bio hype for years (decades?) without much being delivered. Whenever there's a new seemingly breakthrough paper there are many "but-s" that get raised (Has AI discovered a drug? De novo computational generation of antibodies? Protein folding?).

For reflection, here’s a quote about computer-aided drug discovery (CADD), highlighting its importance and impact:

“Drug companies know they simply cannot be without these computer techniques. They make drug design more rational. How? By helping scientists learn what is necessary, on the molecular level, to cure the body, then enabling them to tailor-make a drug to do the job… This whole approach is helping us avoid the blind alleys before we even step into the lab… Pharmaceutical firms are familiar with those alleys. Out of every 8,000 compounds the companies screen for medicinal use, only one reaches the market. The computer should help lower those odds … This means that chemists will not be tied up for weeks, sometimes months, painstakingly assembling test drugs that a computer could show to have little chance of working. The potential saving to the pharmaceutical industry: millions of dollars and thousands of man-hours”

What’s great about this quote is that you can hear its echo in current Silicon Valley tech-solves-biotech pitches, but it was from a Discover magazine article in August 1981 called “Designing Drugs With Computers”. (Four Decades Of Hacking Biotech And Yet Biology Still Consumes Everything, 2017)

Companies that make "Designing drugs with AI" their selling point like Atomwise (2012), Recursion (2013), Schrödinger (1990), or Exscientia (2012) have been around for a while. At least one of them (Schrödinger) have delivered some approved drugs, but the vast majority of drugs developed and approve are still not coming from "throw data at a model and get drugs at the other end". Good thinking and exhaustive experimentation continues to be, to this day, what gets drugs approved, not fancy computational modeling and data alone.

At the same time, at every moment in the history of a field, there is a recurring question: Is this time different? Or is this time like the previous 1000 times?

There's a critique of current work on AI expressed as variations on the argument: "Look, some such systems are impressive as demos. But the people creating the systems have little detailed understanding of how they work or why. And until we have such an understanding we're not really making progress on AI." This argument is then sometimes accompanied by (often rather dogmatic) assertions about what characteristics science "must" have.

I have some instinctive sympathy for such arguments. My original field of physics is full of detailed and often rather satisfying explanations of how things work. So too, of course, are many other fields. And historically new technologies often begin with tinkering and intuitive folk models, but technological progress is then enabled by greatly improved explanations of the underlying phenomena. You can build a sundial with a pretty hazy understanding of the solar system; to build an atomic clock requires a deep understanding of many phenomena.

Work on AI appears to be trying to violate this historic model of improvement. Yes, we're developing what seem to be better and better systems in the tinkering mode. But progress in understanding how those systems work seems to lag far behind. [...]

The underlying thing that's changed is the ease of trying and evaluating systems. If you wanted to develop improved clocks in the past you had to laboriously build actual systems, and then rigorously test them. A single new design might take months or years to build and test. Detailed scientific understanding was important because it helped you figure out which part of the (technological) design space to search in. When each instance of a new technology is expensive, you need detailed explanations which tell you where to search.

By contrast, much progress in AI takes a much more agnostic approach to search. Instead, of using detailed explanations to guide the search it uses a combination of: (a) general architectures; (b) trying trillions (or more) of possibilities, guided by simple ideas (like gradient descent) for improvement; and (c) the ability to recognize progress. This is a radically different mode of experimentation, only made possible by the advent of machines which can do extremely rapid symbol manipulation. (The role of explanation in AI, Michael Nielsen's notes)

I'm not the first to notice some similarity between the research aesthetics of studying neural networks and studying biology. Chris Olah is optimistic about some deeper level of understanding of neural networks. I don't know how optimistic to be about that, but I am certainly more optimistic about that that about the interpretability of biological systems; a point made beautifully in Can a biologist fix a radio and Could a neuroscientist understand a microprocessor?. In artificial neural networks one has very simple entities (neurons that obey simple functions, or layers that perform easy to understand operations), we have a perfectly defined wiring diagram for the network, and we can run all the experiments we want on the neural net itself. Contrast to biology where we have aggregates of squishy bags of molecules (cells) bouncing against each other, and each different from the rest (hyaluronic acid is very different from collagen whereas all neurons in an ANN are basically the same), where the way they interact is not given and has to be studied, often indirectly as we can't easily inspect the state of as cell as we could in an artificial neural network. And of course to make it even worse, the components of biological systems (cells) behave differently in isolation (in a petri dish) than the way they do when they exist in the context of an organ.

So yes, I continue to be not super optimistic about understanding biology! I have written here and there about what does it mean to understand something in biology and asked this same question to my Twitter followers in 2021. Even before that, Bert Hubert already was writing in 2019 that maybe it is the case that biology will never be understood, and then our only hope would be to Gather everything we learn into first-class quality databases that might enable computers to make sense of what we have learned.

We can debate what "understanding" means endlessly but I find more practical to discuss what experiment to do next or what kind of data to gather, and this is driven by what one thinks the road to solving the problems we care about look like. From the point of view of the task "predicting protein structure from its sequence", one could do experiments where we isolate tiny bits of proteins and study how those fold, and attempt to derive rules to predict this, perhaps understanding some aspects of the process. This does work to some extent, we have learned that proteins do have smaller subcomponents (motifs, or domains). We have also learned that there are different kinds of proteins (globular, disorganized, or fibrous), and one can make reasonable guesses about the electric charge distribution in a globular protein (apolar amino acids will be found near the center of the protein). But the road to solving protein folding did not involve eventually discovering some sort of Navier Stokes equations which can be derived from first principles and govern the behavior of the system reasonably well; no, what happened was that a lot of data and massive compute were thrown at the problem and, with some caveats, solved it. Given the nature of the problem, it seems deeply unlikely that we will ever find such simple laws in biological systems except in toy examples.

If one believes this then instead of looking at the problem with our reductionism glasses on, we should take a pair of holistic glasses instead: black box the problem; collect data where we can and let ML figure out the complex pattern of relationships between inputs and outputs.

If you happen to work in a domain that's not biology that's ML-heavy you'll probably be nodding along, but there are some issues with this approach which I'l discuss in a bit.

Protein folding prediction: how useful?

I think for people that are not working in the life sciences, AlphaFold might have changed their view on how optimistic they should be about radically accelerating drug development. For people working in drug discovery, the update might have been small. Both views have something to them: thinking in the short term, indeed solving protein folding doesn't do much to help the current process to find new drugs. In the longer term (and people from outside an industry, and especially those working in tech, might be more likely to take the long/high level view), AlphaFold can indeed be taken as a harbinger of future transformational change.

Mohammed AlQuraishi's commentary on AlphaFold's release back in 2018 continues to be the best summary of the reaction of the protein folding community to DeepMind's monumental achievement. It's a combination of amazement (“What just happened?”)

I don’t think we would do ourselves a service by not recognizing that what just happened presents a serious indictment of academic science. There are dozens of academic groups, with researchers likely numbering in the (low) hundreds, working on protein structure prediction. We have been working on this problem for decades, with vast expertise built up on both sides of the Atlantic and Pacific, and not insignificant computational resources when measured collectively. For DeepMind’s group of ~10 researchers, with primarily (but certainly not exclusively) ML expertise, to so thoroughly route everyone surely demonstrates the structural inefficiency of academic science. This is not Go, which had a handful of researchers working on the problem, and which had no direct applications beyond the core problem itself. Protein folding is a central problem of biochemistry, with profound implications for the biological and chemical sciences. How can a problem of such vital importance be so badly neglected? [...]

What is worse than academic groups getting scooped by DeepMind? The fact that the collective powers of Novartis, Pfizer, etc, with their hundreds of thousands (~million?) of employees, let an industrial lab that is a complete outsider to the field, with virtually no prior molecular sciences experience, come in and thoroughly beat them on a problem that is, quite frankly, of far greater importance to pharmaceuticals than it is to Alphabet. It is an indictment of the laughable “basic research” groups of these companies, which pay lip service to fundamental science but focus myopically on target-driven research that they managed to so badly embarrass themselves in this episode.

With measured caution, even in his later commentary of the improved AlphaFold2 results:

Drug discovery?

I will end this section with the question that gets asked most often about protein structure prediction—will it change drug discovery? Truthfully, in the short term, the answer is most likely no. But it’s complicated.

One important thing to note is that, of the entire drug development pipeline, the early discovery stage is just that, an early stage. Even if crystallography were to become fast and routine, it would still not fundamentally alter the dynamics of drug discovery as it is practiced today, as most of the cost is in the later stages of drug development beyond medicinal chemistry and well into biology and physiology. Reliable protein structure prediction doesn’t change that.

But at the end of that section there is one comment made which might not be noticed at first because it's almost a remark made in passing:

we can imagine a future in which drugs are designed for their polypharmacology, i.e., to modulate multiple protein targets intentionally. This would very much be unlike conventional medicinal chemistry as practiced today where the emphasis is on minimizing off-targets and making highly selective small molecules. Drugs with designed polypharmacology may be able to modulate entire signaling pathways instead of acting on one protein at a time. There have been many fits and starts in this space and there is no reason to believe that a change is imminent, especially because the systems challenges of the equation remain formidable. Wide availability of structures may hasten progress however.

AlQuraishi's comment is a hint of where I think the future of drug discovery will look like: moving beyond the idea of the target to drugging cell or organism state itself.

Drug discovery and self-driving cars

The ultimate goal of the biomedical enterprise (academia, startups, and big pharmas) is the improvement of human health. The proxy goal for that goal is drug discovery (and development) and the proxy for "drug" is, historically, orally available single-target small molecule inhibitors of some protein (or agonist for some receptor). Or what is the same: Up until recently, the way of thinking if you want to address a disease is:

Understand a disease: Have an idea of what's going on, find the pathways involved, examine human genetics to find correlations between genes and disease incidence.
1. Example: Learning that the mevalonate pathway is involved in LDL synthesis, and that LDL cholesterol is a driver of heart disease
Find a target: a protein (usually) that is involved in a disease and whose activity can potentially be modulated (turned up or down). It tends to be easier to block the action of a protein than to enhance it. (Example: HMG-CoA reductase)
Find a small molecule that delivers the desired effect, i.e. that binds to the catalytic domain of an enzyme to inhibit it
1. Usually one tries lots of compounds (high throughput screening), then picks promising ones and tweaks them until one seems to work well.
2. Here one can also do some ML to speed up e.g. docking calculations
3. Example: Atorvastatin, which inhibits the action of HMG-CoA thus reducing LDL synthesis downstream
Ensure that said small molecule can be taken orally. There are various rules of thumb here like Lipinski's rules to guess if a molecule will be "druglike".
1. This is not always the case; some drugs are injected so no need to worry about gut absorption then. Vaccines are the clearest example.
Ensure that said small molecule doesn't have side-effects
1. Example: Statins do have side effects, but they are considered minimal in relation to the benefit of the drug. Nonetheless the search for even safer interventions has led to other LDL-lowering drugs like PCSK9 inhibitors.
2. Example 2: A gamma secretase inhibitor (For Alzheimer's treatment) caused increased skin cancer, so even if it treated the disease it's not on net worth using
Profit!
1. Example: Lipitor (atorvastatin) generated billions of dollars of revenue for Pfizer over the last 20 years
2. Example 2: Even when drugs don't end up making it all the way to the clinic (As happened with Sirtris) you can profit too sometimes, what's not to like! /jk

You might wonder a few things here:

How do you know what the target should be? First principles thinking, domain knowledge, little experiments here and there.
1. Example: Atorvastatin came from earlier research on other molecules, lovastatin and mevastatin, which in turn was discovered by searching for antimicrobial agents, fermenting broths of a fungus
Why a single target? It's easier to carefully study two entities (a protein and a ligand for it) than to study every molecular entity in a cell in detail
Why restrict ourselves to orally administered drugs? Because these drugs will need to be administered repeatedly (they are small molecules, eventually they get metabolized and excreted), usually at home, and having people injecting themselves daily is considered unfeasible.

As an analogy here, consider self-driving cars: Traditionally the problem of driving a car initially required engineers to identify features like lines and map them to say lanes and then keep the car in there. The algorithms used for this were simple and understandable like Canny edge detectors or Hough transforms. The answers to the equivalent questions would have been something like:

How do you know what features to use? First principles thinking, asking domain experts, little experiments here and there.
Why a small number of features, why two lines/a single lane? It's hard to consider very complex scenarios, let's do a single lane at a time.
Why restrict ourselves to driving on sunny days with good visibility? It's already hard to do this!

With self-driving cars we are now seeing a competition between two approaches: This classical approach just described, where many intermediate, handcrafted representations are modeled explicitly (mostly abandoned), and the new end to end approach (Tesla and comma.ai perhaps being the ones ideologically closer to this) where the car goes straight from pixels perceived to commands issued to the motors and steering system, where visualizations are still provided to the human driver for reassurance, but not being strictly required for model performance.

This second paradigm replaces an object-level research oriented mindset where one aims to understand the system of interest, with an engineering-heavy mindset where understanding is deprioritized in favor of control. What matters then is instead designing systems able to ingest large quantities of data and models able to distill that into solutions to the problem, and that is an easier task than answering the question "how does one drive" or "what is driving" from first principles.

What would the equivalent of this approach look like for drug discovery?

Self driving cars are easier than drug discovery

While not perfect in all cases yet, it's now possible for a commercially available car to drive itself all the way from SF to LA. What made this possible is largely the same thing that made DeepMind's achievements in Go and Chess possible: Lots of data and simulators that are very close to the real domain. The domain where the model is operating (a car in the real world) can be trained on real world data for that same system (from the Tesla fleet) and enhanced with simulated driving data. The physics of driving a car are understood well enough and graphics can be made so realistic than one can train on simulators too!

In biological systems this is not the case: The state of a human body is extremely complex and not yet fully understood. The dynamics of it extend over days (fighting an infection), months (pregnancy), and even decades (for processes like puberty). Measuring this state is also nontrivial; one can only collect blood only so often, and measure only so much. Extracting biopsies to access organ state directly is highly invasive, and the perturbations a human is exposed to in the wild are far from what would be required to find new drugs. Natural data is useful to learn about things like exercise and diet (and even that's hard), but we don't go around taking random pills so that one can build models of what random compounds do to us.

Sometimes a variant of this is possible: Because (twins aside) we are all genetically distinct, nature is running the world greatest clinical trial in us already, and with large enough collections of sequenced genomes it's possible to chart a path towards new drugs.

But genetic lottery aside, short of a large army of willing volunteers, we can't use the actual system (the human body) to experiment with directly at scale, and we can't simulate it yet, we have to settle for something simpler. We can either test in animals (a whole organism) or we can test in human cells in vitro, or eventually in organoids.

I am very pro in vivo testing: Ultimately yes we are made up of cells but there are many different kinds that interact in different ways. If immune rejuvenation is one key part of the future of medicine, we wouldn't know the full extent of that if we just observe that indeed the function of a given type of immune cell can be improved, one has to see what a that cell does when placed in the context of an organism where it can now more effectively fight cancer and pathogens.

Some recent commentary on the future of ML for biology

What originally inspired me to write this are these following articles that I read last year. I got a sense there was a sense of renewed excitement in the field (Or just it was a coincidence that I these ended up in my reaading list) that was worth thinking about.

One idea is say training a large language model on the entirety of Scihub and then asking it to solve a particular problem like predicting protein structure, producing drug candidates, or explaining why Alzheimer's actually happens. This has been tried before and the results have been... far from that promise: Albeit trained on less data, this is what happened with Galactica, and this is the current state of ChatGPT and similar state of the art models. Scraping Scihub is feasible, but the results probably won't be that enlightening: We want the models to tell us new things, and so far LLMs tend to be very conservative. But beyond that, there isn't that much data out there. The papers might describe at a high level an experiment that was done and some particular results, but accessing the raw or processed data that was gathered is something one can't get from the paper or even the public internet in many cases. Sam Rodriques is right when he says that I am also skeptical of the ability of even an AI trained on the entire scientific literature to predict drug efficacy for diseases for which have no effective drugs and no understanding of how they work.

Josh Nicholson, who wrote in more detail about how difficult it would be to do this actual thing, is more optimistic. But as he points out, we already have this: ScholarBERT was actually trained on what seems to be all scientific papers (75M of them, 221B tokens, as opposed to 48M papers/88B tokens for Galactica. Scihub has ~80M), and Science remains an unsolved problem. ScholarBERT is a relatively small model (770M parameters) so one can always think that maybe 100x parameter count would lead to better performance at Solving Science but I doubt it.

However, the real problem we care about is not producing plausible (given current knowledge) completions to papers. An assistant that has access to the world's scientific knowledge (or its publicly available portion) would be valuable but not that useful, especially if scientists working in a domain already have that knowledge. It would be a different matter if a model generates new hypothesis or proposes new experiments that are unintuitive but promising.

Adam Green writing A Future History of Biomedical Progress expresses the same sentiment I share throughout the essay, going perhaps even further than I would. His essay is the most substantial inspiration for my own:

progress in basic biology research tools has created the potential for accelerating medical progress; however, this potential will not be realized unless we fundamentally rethink our approach to biomedical research. Doing so will require discarding the reductionist, human-legibility-centric research ethos underlying current biomedical research, which has generated the remarkable basic biology progress we have seen, in favor of a purely control-centric ethos based on machine learning. Equipped with this new research ethos, we will realize that control of biological systems obeys empirical scaling laws and is bottlenecked by biocompute. These insights point the way toward accelerating biomedical progress. [...]

One cut on this is how physics-like you think the future of biomedical research is: are there human-comprehensible “general laws” of biomedical dynamics left to discover, or have all the important ones already been discovered? And how lumpy is the distribution of returns to these ideas—will we get another theory on par with Darwin’s?

For instance, RNA polymerases were discovered over 50 years ago, and a tremendous amount of basic biology knowledge has followed from this discovery—had we never discovered them, our knowledge of transcriptional regulation, and therefore biomedical dynamics, would be correspondingly impoverished. Yet when was the last time we made a similarly momentous discovery in basic biology? Might biomedicine be exhausted of grand unifying theories, left only with factoids to discover? Or might these theories and laws be inexpressible in the language of human-legible models?

But in one of the footnotes there's a point where the complications of truly being "end to end" become more obvious:

Insitro et al. are to drug discovery as Waymo et al. are to autonomous vehicles. Just as some think autonomous vehicles will be solved by building high-definition maps of cities and modeling dynamics at the level of individual pedestrian behavior, some think biomedicine will be solved by building high-definition molecular “maps” of diseases and modeling dynamics at the level of individual cellular behavior. Though they are directionally correct in their use of machine learning, they fail to abstract sufficiently.

Green wants to truly "end-to-end" biology. That is, having a system we can ask "make a human healthy" and getting an answer, trained on triplets of (diseased human, perturbation, healthy human). Of course, he admits this is unrealistic because of ethical considerations; so rather he proposes to do this in mice and organoids (as close as possible to the real system) and then try to transfer from there. In the paragraph above he says Cellarity is not going far enough: They are trying to fix cells, but cells are not what we ultimately care about (the whole organism); in his view fixing cells is like learning to recognize traffic cones when building a self-driving car: a hand-engineered feature that is not required if one can end-to-end enough.

I think cells are better models than he thinks, perhaps. Biological systems have the advantage of being built of similar building blocks (all cells work in the same fundamental way), where parts of the system are adjusting themselves to the state of other parts. If you rejuvenate e.g. blood, you can probably have effects elsewhere. If you hit 60% of a tissue with a successfully rejuvenating therapy, chances are you might go beyond that 60% through cell-to-cell signaling. The self-driving analogy shouldn't be cones but rather charging electric cars: Given the task "Driving an electric car across the United States without human intervention" one has to automate driving and charging. The true end to end approach would be to train a joint model to control both the car and the charger (perhaps equipped with one of these). A single neural network that tells the car what to do and same for the charger. But in practice this is unnecessary: You can have a model that drives the car to a spot in the charger and then a simple computer vision based model that controls the charger and gets the car charged. The performance of this split approach wouldn't be inferior to the true end to end solution, and it is easier to train.

Similarly, while on paper the problem of "altering the state of a cell" involves a) designing what to do to the cell and b) getting that to the cell, I could imagine how one can solve (a), say a model that predicts what transcription factors to get a cell to express, then trying to find a way to package that into an AAV or something else, (b). This might not be doable, but then one can pick the next solution from the model that solves (a) and try again. To the extent that the domains being decoupled, one can substitute end-to-end learning with some more trial and error. Ultimately, the question is: Should we put more resources on organoids or better models, ultimately having 'organs on a chip' so that we can collect data to train end-to-end models? Or should we try to develop therapies with the tools we have available right now? My hunch is that the latter approach can still be useful.

One more argument against the need for complete end-to-end is that biology is incredibly 'plug and play'. It's possible to replace or address subsystems of an organism separately. For example one can replace old bone marrow with young bone marrow without having to concurrently replace everything else. One can even implant bits of organs in the right place and those organs will function somewhat. And obviously we have seen many successful drugs being developed by modeling just parts of the whole.

And lastly, as readers of Nintil know, I'm a fan of partial reprogramming. I do think fixing aging goes a long way in extending healthy lives, and aging is, to a large extent, the deteriotation of processes that are common to all cell types (like transcription, translation, or autophagy), hence fixing this in vitro and solving systemic delivery seem to, together, go a long way!

Pablo Cordero writes here about unifying all of biology into a large model by thinking of biological knowledge as a graph, masking parts of it, and then predicting those parts from the rest of the data. It's not fully clear how one would go about doing this! I'm no expert in graph neural networks, but certainly the post shares the spirit of "end-to-end biology".

Lastly, Jacob Kimmel wrote a really good post last year on representation learning as an extension from the way early molecular biologists worked:

There’s no general solution to modeling complex systems, but the computational sciences offer a tractable alternative to the analytical approach. Rather than beginning with a set of rules and attempting to predict emergent behavior, we can observe the emergent properties of a complex system and build models that capture the underlying rules. We might imagine this as a “top-down” approach to modeling, in contrast to the “bottom-up” approach of the physical tradition.

Whereas analytical modelers working on early structures had only a few experimental measurements to contend with – often just a few X-ray diffraction images – cellular and tissue systems within a complex organism might require orders of magnitude more data to properly describe. If we want to model how transcriptional regulators define cell types, we might need gene expression profiles of many distinct cell types in an organism. If we want to predict how a given genetic change might effect the morphology of a cell, we might similarly require images of cells with diverse genetic backgrounds. It’s simply not tractable for human-scale heuristics to reason through this sort large scale data and extract useful, quantitative rules of the system.

Is this time different?

Current "ML for drug discovery" startups are far from end to end. They still find a target the old fashioned way, and limit themselves to small molecules (As with Relay or Exscientia). Some do go beyond the concept of a target and into phenotypic screening (like Recursion), where there is no initial driving hypothesis behind a drug, instead the company builds models trained to recognize features of cells that look more or less diseased and then build relations between the drugs the cells were treated with and the observed change. Recursion hasn't gotten any drug approved yet. Cellarity seems to follow a similar approach, moving away from the idea of a target and towards drugging cell state holistically. I suspect we will see more companies moving in this broad direction.

Why is this changing? The costs of reading and writing DNA are the lowest they have ever been. So is the number of cells we can measure per experiment. Just a few days ago a new paper came out reducing the cost of testing genetic perturbations in cells by an order of magnitude. "High-throughput" is now "Ultra-high throughput". Collecting data was never as cheap as it is today.

In parallel to increasing volumes of data being collected, only very recently we have started to see the appearance of models that can output predictions on what to do to a population of cells to shift them to a desired state (like PerturbNet, from 2022) or models that can predict the effect of combined genetic perturbations (GEARS, also from 2022), and of course transformers are coming for perturbation prediction as well (scFormer, once again 2022). Thanks to neural networks and progress in representation learning, the model can predict chemical perturbations or gene perturbations alike.

I don't have concrete timelines for when we are going to 'solve biology with ML', but working towards that end seems enormously valuable

Links (65)

Fri, 30 Dec 2022 00:00:00 +0000

A video explainer of Helion's approach to fusion. Here at Nintil I first covered Helion in 2014 where I noted that their estimate date to commercial fusion back then was 2020 and estimated cost per kWh was $0.04. Unclear what those numbers are right now.

What are diffusion models, a Lilian Weng explainer

Alphonse Mucha, Art Nouveau graphic artist. You have probably seen some of his work but may not know who he was until now.

I wrote some thoughts here about leveraging the Icelandic Mutation (that confers resistance to Alzheimer's) to cure the disease. Not as optimistic as I initially was for treating the disease but it continues to be the best we have to prevent it.

The difference between IQ and intelligence, and whether education increases them

Sasha Chapin does Janas

Using deep learning to model the hierarchical structure and function of a cell

How to replace your backends with postgres

The future of medicine looks less like "find small molecule that fits a pocket in a protein" and more like gene therapies, cell therapies, and pathway engineering.

ChatGPT and its lineage

Paper argues for understanding protein folding for first principles. My aesthetic says probably not.

There are various kinds of "fast aging" or progeroid mice. It's commonly agreed that these don't represent "real" aging but some do seem to be useful to model some parts of it.

New QRI paper on the "Slicing problem"

Ben Kuhn on staring into the abyss as a core life skill

Scientist goes around decapitating worms that grow back their heads, collects results

What does the NIF fusion achievement mean? From the point of view of eventual commercial fusion, not much.

The story of VaccinateCA

I ask Twitter about co-CEOs

A snippet from David Friedman's Legal systems very different from our own

I enable Extreme Taleb Mode in ChatGPT

What's cooler than diffusion models for images? Diffusion for proteins and songs. Diffuse all the things!

LLMs continue to take over biology

The insane adventures of Francis Galton

Nintil

Links (83)

Links (82)

Burning Man

Appendix: Packing list

Links (81)

Links (80)

Links (79)

Links (78)

Links (77)

Making Cells Young

Docusign's 7000 employees and the iron law of B2B SaaS sales

Links (76)

Links (75)

Rapamycin is not an aging drug. But what is an aging drug?

Reprogramming as only current approach to systemic, permanent, rejuvenation

Systemic rejuvenation is not enough

One last word on rapamycin-like drugs

Appendix: claims regarding rapamycin and aging in various sources

Some thoughts on causality in biological systems

Changelog

Links (74)

Telomeres: everything you always wanted to know

Links (73)

Elon's decision making: an anecdote compilation

Changelog

Massive input and/or spaced repetition

Links (72)

Links (71)

Links (70)

The situational awareness assumption in AI risk discourse, or why people should chill

The case for doomerism

Non-superoptimizing AGI is possible

Sarah Constantin's post

Situational awareness, some earlier discussion

Situational awareness is required for most AI risk scenarios

The argument for situational awareness emergence

My response to this argument

Links (69)

Links (68)

Links (67)

Links

Retro Biosciences

Links (66)

Bryan Johnson's Blueprint

Blueprint for all?

Measuring success

Exercise

Back to basics: protocol adherence for all

Notes on end-to-end biology

We are far from understanding all of biology, but that's okay

Protein folding prediction: how useful?

Drug discovery?

Drug discovery and self-driving cars

Self driving cars are easier than drug discovery

Some recent commentary on the future of ML for biology

Is this time different?

Further reading

Links (65)