Artificial Intelligence

A short time of writing with AI Dungeon demonstrated to me that language models are multiverse generators, and that an interface which supports branching would provide great utility for cowriting with a language model, whether one is composing virtual worlds intended to be multi- or single-history (in the latter case, branching explores counterfactuals which can be merged into main), as well as for nonfiction tasks like brainstorming and prototyping prompts.

After I was given access to OpenAI’s API, motivated to actualize my HITL thought experiment, I began making an interface which allows the user to generate, navigate, and visualize natural language multiverses.

User interfaces (UIs) are the connections between us — soft, adaptable humans — and the hard, deterministic world of data and machines.

Historically, these interfaces have been rigid, representing the deterministic logic of the machine and relying on the user to translate the hard logic into a soft, meaningful form. With the rise of LLM capabilities, our technology is newly better at acting soft and squishy. Which brings the question: how should our interfaces handle the transition from the hard logic of machines into the soft logic of humans?

...Hopefully this framework of segmenting at different scales or gradually transitioning across different dimensions can help organize our thinking.

I remembered an essay from Robin Sloan I had squirreled away in my notes a few years ago: Week 4, ladder of abstraction. He talks about “the ladder of abstrac­tion”, introduced by Roy Peter Clark in his book “Writing Tools”:

Good writers move up and down a ladder of language. At the bottom are bloody knives and rosary beads, wedding rings and baseball cards. At the top are words that reach for a higher meaning, words like freedom and literacy. Beware of the middle, the rungs of the ladder where bureau­cracy and tech­noc­racy lurk. Halfway up, teachers are referred to as full-time equivalents and school lessons are called instructional units.

What a wonderful concept! By making it easy to scan the "abstractness" of your writing, we can better identify the "middle rungs" or long stretches on either side of the ladder.

So, how do I get the “abstractness” of each sentence in a user’s writing? Using a more traditional method with an “abstractness level” for every word and averaging across a sentence would surely miss lots of nuance. So, can we use AI?

The AI glut significantly exacerbates the issue of misinformation and low-quality content. The current state of AI technology lacks the nuanced understanding and ethical judgment necessary to ensure the accuracy and integrity of the content it produces. This gap in capability opens the floodgates for misinformation to spread unchecked, fracturing discourse, heightening divides and hampering decision-making.

This all boils down to one thing. Human curation is now more critical than ever. As algorithms churn out vast quantities of information with varying degrees of accuracy and quality, the discerning judgment of human curators is the only defence against the tide of misinformation and mediocrity. Human curators bring nuanced understanding, contextual awareness, and ethical judgment to the table—qualities that AI, in its current state, is fundamentally unable to replicate. 

Can you tell me in what way this is better than a normal search results page? The problem with search is that it’s full of fucking spam and ads. That’s the problem with search. The solution to search is not to “reinvent the browser” or to use AI to power searches for you. The solution to search is a search engine that is aligned with the users.

I find it helpful to think about language models as a “squishy”...

What do I mean by “squishy”?

Language models feel like something organic and biological rather than something mechanical. It feels like we grew them, rather than built them.

...So given that these models are untamed, squishy, biological things, we have a core tension we need to navigate when we’re designing and building with them: squish vs. structure

We’re trying to make an unpredictable and opaque system adhere to our rigid expectations for how computers behave. We currently have a mismatch between our old and new mental models for computer systems.

Using the chatgpt api, I ran the same completion prompt "Intelligence is " hundreds of times (setting the temperature quite high, at 1.6, for more diverse responses). Given a text, a Large Language Model assigns a probability for the word (token) to come, and it just repeats this process until a completion is...well, complete.

• we’re beginning to visualise the previously invisible deep structure of text
• comparatively
• in real-time
• and then we’re learning how to manipulate text based on these hidden features.

I’m reminded of that famous series of photographs, The Horse in Motion, from 1878.

Until that moment, neither scientists nor the public knew whether or not all four of a horse’s hooves came off the ground when it runs. Imagine! It was a controversy!

The camera was a new instrument that showed what was already present, but inaccessible to the human eye.

But the camera isn’t just a scientific instrument like the, I don’t know, Large Hadron Collider. By the Saturday Evening Post, here are 5 Unintended Consequences of Photography (2022):

Photography Decided Elections
Photography Created Compassion
Photography Liberated Art
Photography Shaped How Americans Look
Photography Gave Us an Appreciation of Time

So the camera doesn’t just observe and record, it changes us.

---

Text is becoming something new, that’s what I mean.

We’re inventing the camera and Photoshop simultaneously, and all their cultural repercussions, and to begin with this means new apps with new user interfaces, and where it goes after that I have no idea.

Previously, I tagged content generated collaboratively with GPT-3 with a curation ratio, intended to give an approximate sense of the amount of cherrypicking involved in its creation. Others have similarly used a ratio to indicate curation selectivity. However, this description doesn’t distinguish between, say, choosing the best of 5 entire essays generated by GPT-3 and choosing the best of 5 sentences every sentence. The latter text has received much more optimization pressure per token and is likely to look a lot more coherent.

Gurkenglas made the excellent suggestion that I track the number of bits of selection for an objective and exact measure of human intervention. Normally, this would be a lot of trouble, but fortunately Loom trees contain sufficient information to retroactively calculate bits of selection and intervention. From now on, I will label content that is generated with the assistance of GPT-3 or other language models with the metrics presented in this post.

AI-generated artwork is the same as a gallery of rock faces. It is pareidolia, an illusion of art, and if culture falls for that illusion we will lose something irreplaceable. We will lose art as an act of communication, and with it, the special place of consciousness in the production of the beautiful.

…Just as how something being either an original Da Vinci or a forgery does matter, even if side-by-side you couldn’t tell them apart, so too with two paintings, one made by a human and the other by an AI. Even if no one could tell them apart, one lacks all intentionality. It is a forgery, not of a specific work of art, but of the meaning behind art.

Embeddings are a technology that’s adjacent to the wider field of Large Language Models—the technology behind ChatGPT and Bard and Claude.

Embeddings are based around one trick: take a piece of content—in this case a blog entry—and turn that piece of content into an array of floating point numbers.

The key thing about that array is that it will always be the same length, no matter how long the content is. The length is defined by the embedding model you are using—an array might be 300, or 1,000, or 1,536 numbers long.

The best way to think about this array of numbers is to imagine it as co-ordinates in a very weird multi-dimensional space.

Why place content in this space? Because we can learn interesting things about that content based on its location—in particular, based on what else is nearby.

The location within the space represents the semantic meaning of the content, according to the embedding model’s weird, mostly incomprehensible understanding of the world. It might capture colors, shapes, concepts or all sorts of other characteristics of the content that has been embedded.

Nobody fully understands what those individual numbers mean, but we know that their locations can be used to find out useful things about the content.

Rewind is a personalized AI powered by everything you’ve seen, said, or heard.

A Code Graph is a visual representation of a codebase as a Knowledge Graph that helps one explore entities in code (functions, variables, classes) and their relationships. By mapping out these connections, it becomes easier to understand the structure and flow of the code, identify potential issues, and improve overall code quality.

The concept of representing code as a graph data structure has its roots in the early days of software engineering when researchers explored ways to model and analyze program structure and behavior using graph-based techniques. Modern Code Graphs, incorporating Knowledge Graphs and Large Language Models (LLMs), are a recent development.

• Improved Understanding: Helps trace the flow of data through functions and identify interconnected components.
• Impact Analysis: Assesses the ripple effects of code changes, predicting potential issues before they arise.
• Autocompletion: Suggests relevant functions, variables, and types based on the current context.
• Code Search: Searches for functionalities not just by keywords, but by understanding the relationships between code elements.

While language model (LM)-powered chatbots and generative search engines excel at answering concrete queries, discovering information in the terrain of unknown unknowns remains challenging for users. To emulate the common educational scenario where children/students learn by listening to and participating in conversations with their parents/teachers, we create Collaborative STORM (Co-STORM). Unlike QA systems that require users to ask all the questions, Co-STORM lets users observe and occasionally steer the discourse among several LM agents. The agents ask questions on the user's behalf, allowing the user to discover unknown unknowns serendipitously. To facilitate user interaction, Co-STORM assists users in tracking the discourse by organizing the uncovered information into a dynamic mind map, ultimately generating a comprehensive report as takeaways.

Here’s the proposal, which I suspect many will reflexively condemn as heresy, but which I promise to unpack if given the chance: Those, like me, who believe that “ordinary meaning” is the foundational rule for the evaluation of legal texts should consider—consider—whether and how AI-powered large language models like OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude might—might—inform the interpretive analysis. There, having thought the unthinkable, I’ve said the unsayable.

Now let me explain myself.

Build software faster in an IDE designed for pair-programming with AI.

A strange universe-expanding game that's been described as "whoah dude".

Explore procedural galaxies, land on unknown planets, read the thoughts of every living thing, rummage through people's pockets, and more!

Great software starts on the command line. Warp is the terminal reimagined with AI and collaborative tools for better productivity.

There is a large-scale, deliberate effort to manipulate culture and discourse online and in wider culture by utilising a system of bots and paid employees whose job it is to produce content and respond to content online in order to further the agenda of those they are employed by.

Large proportions of the supposedly human-produced content on the internet are actually generated by artificial intelligence networks in conjunction with paid secret media influencers in order to manufacture consumers for an increasing range of newly-normalised cultural products.

For now there are moments of stunning humanity, a short term triumph even, of human over computer. Because what we are witnessing here is a reversal of the mold, in which humans input a massive amount of data into a computer that outputs something new and searchable. Here the computer is the input that the human scans, learns from, trains on, eventually outputting something magnificent.

...Lately, Rainbolt has been traveling the world, and is letting us in on the secrets he has seen in the data. Did you know that square stone fences with circle holes are only found in Chiang Mai, red stop signs with white font are only found in New South Wales, and the Wooden poles are cut in an octagonal shape only in Sydney? It's exciting to see him get off the computer and traveling the world in real life. Rainbolt is a modern explorer, an internet savant, and he’s been training at home so that his trip can be even more interesting. Maybe the internet won't just swallow us whole. Maybe we'll use it to appreciate just how distinctive humanity really is.

On the con side, the “uncompress” part is very messy and can lead to the cognitive and social equivalent of those big blocky pixels. A nice, or at least good, thing about JPEGs is you can see where the decompression happened: like in “The Starry Night” above, which is obviously compressed. But because of the way AI works, constantly guessing and filling in blanks, you can’t see the artifacts. It just keeps going until people have twelve fingers, stereotypes get reaffirmed, utter nonsense gets spewed, and so forth. You can see the forest, but the trees are all weird.

Given how quickly concept maps help teams get aligned on their understanding, I thought it might be useful to develop a tool to expedite the creation of concept maps. I don’t mean a tool to help you draw a map. After all, there are great diagramming tools in the market. I mean a tool that will draw concept maps for you.

Synthesis is one thing LLMs do well. Ask ChatGPT to summarize an article, and you’ll get a useful couple of paragraphs that get to the gist of whatever you’re wanting to understand. But the output is still text. What if it could do the same but render a diagram instead?

In the demo, an AI bot for a style guide makes changes as another collaborator in the document. It puts changes on a branch, which you can choose to partially or completely merge—just like you would suggestions from a human coauthor.

The changes are highlighted in a diff to help you review. Once you accept edits from a bot, the history timeline also shows which edits came from the bot.

One more thing: since the bots are simply text prompts, you can create your own by adding a bot prompt into your document collection!

Bots can be shared, edited, and versioned just like regular documents—for example, you can view the history of a prompt, or use a branch to try a new variant of a prompt.

Find everything humans know about...

Of course, décor is not instant, not flat. It is produced as reactions to one’s interests and personal version of The World, and is also - at its best – the result of time passing and EVOLVING desires and things being accumulated to fill the emotional and practical needs of the time / a time.

For a long time, these four quadrants have been pretty distinct, because search and recommendations, while they occupy opposite ends of the same spectrum conceptually, have different tooling and architectures, and machine learning goals. You can even tell the way org structure of companies who hire for the teams have developed: usually there are distinct search and recommendations teams.

What happens in today’s world, when we have LLMs? We now have the collision of three different sets of user expectations.

Why is it that so much of what is commonly called AI “art” is kitsch? 

In part this is because users of AI image generators fancy themselves as artists even though few of them have any art training. This is common in photography. Wealthy individuals purchase camera gear based on reviews claiming that some camera or lens has greater technical abilities to reproduce reality faithfully and then apply complicated methods to assure that their photographs demonstrate technical proficiency. High Dynamic Range (HDR) photography is the leading example of this. Popular with amateurs with no aesthetic training, HDR is an attempt to capture a scene in which the range of luminance exceeds the dynamic range of the camera sensor, and often even the human eye itself. The results typically have too much detail in the shadows, dark skies, unnatural colors, the hyperrealistic effect of an acid trip. 

These sorts of photographers, along with individuals who produce digital illustrations for consumption on platforms like Artstation and DeviantArt, 3D printing enthusiasts, makers, indie musicians working with samplers and synthesizers, vloggers creating content for YouTube, gamers streaming on Twitch and YouTube, and fashion enthusiasts showcasing their work on social media are “prosumers,” a term coined by futurist Alvin Toffler in his 1980 book The Third Wave. Toffler’s “prosumer” merges the roles of producer and consumer, suggesting a shift in the economy and society. In this model, individuals are not only consumers of products and services but also take on an active role in their production. This concept was revolutionary at the time, predicting the rise of customization, personalization, and participatory culture facilitated by technological advancements, particularly in digital technology and the Internet.

At the same time, prosumers largely create kitsch, characterized by an appeal to popular tastes and a frequently derivative nature. Kitsch thrives in environments where production is geared towards mass appeal and immediate consumption rather than nuanced artistic merit or innovation. For traditional modernist critics, such as Clement Greenberg, kitsch represented the antithesis of genuine culture and the avant-garde. Kitsch, Greenberg explained in his seminal 1939 essay “Avant Garde and Kitsch,” is produced by industrialization, designed to satisfy the tastes of the least discerning audience without intellectual or emotional challenges. Greenberg associated kitsch with the replication of traditional art forms and aesthetics, but emptied of genuine meaning or complexity, offering immediate gratification rather than enduring value or depth.

If I were tasked with finding an icon that can be used to label an AI-related feature in a UI component, I’d likely pursue a set of icons more closely tied to the feature itself, such as the sparkle-donning envelope for AI-flavored email features. The sparkle is what suggests automation and whatever it is applied to it is the feature. Sparkles alone can mean too many things (and reek of marketing cruft).

...Some of this isn’t specific to AI but to iconography in general. The detail needs to be clear enough to convey meaning, but not so detailed that the meaning is lost when downscaled. The elements need to be recognizable at first glance but also avoid crossing into well-known commodities, like how the helix shape closely resembles the logos of React and The Big Bang Theory. The subject matter has to steer clear of creepy connotations or else suffer the fate of Uncanny Valley. These are all considerations to take into account when working with icons of any kind.

What would have happened if ChatGPT was invented in the 17th century? MonadGPT is a possible answer. MonadGPT is a finetune of Mistral-Hermes 2 on 11,000 early modern texts in English, French and Latin, mostly coming from EEBO and Gallica.

One of the great failures of modern computing is how it has largely ignored the presence of the human body beyond the slightest acknowledgement that humans have a pair of eyeballs and a few fingertips.

...Consider, for instance, our hands. Human hands are capable of so much more than typing and swiping. Human finger sensitivity is astonishingly precise, capable of detecting wrinkles on a smooth surface with a wavelengths as small as 10nm, features so small they are invisible to the eye. Hands can both manipulate and communicate, have effect on the world and perceive its most minuscule features.

Enthusiasm around large language models (LLMs) and text-to-image systems, however, has made it easy to forget about the impressive facility of our hands and the astonishing capabilities of our vision systems, leading instead to proclamations about the pending death of the graphical user interface as an input system. Instead of veering towards a future where computers make greater use of our bodies’ expressiveness, LLMs compel us to diverge further from such a course. Why manipulate anything on a computer with gesture or point and click when you can type or speak?

ChatGPT is not designed to produce true utterances; rather, it is designed to produce text which is indistinguishable from the text produced by humans. It is aimed at being convincing rather than accurate. The basic architecture of these models reveals this: they are designed to come up with a likely continuation of a string of text. It's reasonable to assume that one way of being a likely continuation of a text is by being true; if humans are roughly more accurate than chance, true sentences will be more likely than false ones. This might make the chatbot more accurate than chance, but it does not give the chatbot any intention to convey truths. This is similar to standard cases of human bullshitters, who don't care whether their utterances are true; good bullshit often contains some degree of truth, that's part of what makes it convincing. A bullshitter can be more accurate than chance while still being indifferent to the truth of their utterances. We conclude that, even if the chatbot can be described as having intentions, it is indifferent to whether its utterances are true. It does not and cannot care about the truth of its output.

I’m sure we’ll see the rise and widespread use of authoritarian AI analytics, fed a constant stream of images and audio recordings, finding crimes that never happened in the blur of a street scene or hearing things were never said in a citywide wiretap—call it the Gosling Effect—resulting in people going to prison for the evidential equivalent of faces that were never really there.

Chatbot Arena is an open-source research project developed by members from LMSYS and UC Berkeley SkyLab. Our mission is to build an open platform to evaluate LLMs by human preference in the real-world.

Stable Diffusion’s training datasets are impossible for most people to download, let alone search, with metadata for millions (or billions!) of images stored in obscure file formats in large multipart archives.

So, with the help of my friend Simon Willison, we grabbed the data for over 12 million images used to train Stable Diffusion, and used his Datasette project to make a data browser for you to explore and search it yourself.

I guess what I’m saying is: thanks, I hate it. Give me context; give me nuance; give me the ability to think for myself. We built the world’s most incredible communication and knowledge-sharing medium, rich with diverse perspectives and alternative ideas; let’s not sanitize it through a banal filter that is designed to strip it of its humanity.

Most AI images look like shit. AI “artists” are quick to lecture me that generative tools are improving every day and what they spit out won’t always look this way – I think that’s beside the point. What makes AI imagery so lousy isn’t the technology itself, but the cliché and superficial creative ambitions of those who use it. A video of a cyber-punk jellyfish or a collie in sunglasses on a skateboard generated by Open AI’s new video-to-text model Sora aren’t bad because the animals in them look unrealistic; they’re bad because they’re mind-numbingly stupid. AI image generation is essentially a truncated exercise in taste; a product of knowing which inputs and keywords to feed the image-mashup machine, and the eye to identify which outputs contain any semblance of artistry. All that is to say: AI itself can’t generate good taste for you.

Paragraphica is a context-to-image camera that uses location data and artificial intelligence to visualize a "photo" of a specific place and moment. The camera exists both as a physical prototype and a virtual camera that you can try.

When I polled my community about their attitudes toward busywork — a ruse to figure out what some of my nearest and dearests actually do for work — most at least saw value, if not joy, in occasional busywork. A web designer told me busywork serves as “productive procrastination” when she’s avoiding more complex tasks. A woman in sales and marketing said she values the solitude of rote tasks, and retreats into spreadsheets “when everybody’s annoying and I’m peopled out and my bullshit meter is filled.” A senior research program manager at a nonprofit explained that she values how data cleaning — combing through a dataset for errors, duplicates, and other issues — creates an intimacy with the information she’s processing. Cleaning data manually makes the phenomena she studies less abstract: “It connects you to a different way of working or being, or creates opportunities to see things in a different way.”

At the intersection of the near future and the reimagined past, waits a world in which every human appetite can be indulged without consequence.

Oh no, someone used a checks notes digital art tool to win the checks notes digital art contest at the state fair, this is TEH END OF ART. Brief thread, since people have asked me to weigh in on this.

Let's compare Google, Bing, Marginalia, Kagi, Mwmbl, and ChatGPT

GOODY-2 is a new AI model built with next-gen adherence to our industry-leading ethical principles. It’s so safe, it won’t answer anything that could be possibly be construed as controversial or problematic.

If the result of your software development methodology is that the fix to the manager of the payments team being banned is to whitelist the user after six months, that traffic routing in your app is systematically wrong for two decades, that core functionality of your app doesn't work, etc., no amount of hiring people with a background that's correlated with noticing some kinds of issues is going to result in fixing issues like these, whether that's with respect to ML bias or another class of bug.

Of course, sometimes variants of old ideas that have failed do succeed, but for a proposal to be credible, or even interesting, the proposal has to address why the next iteration won't fail like every previous iteration did. As we noted above, at a high level, the two most common proposed solutions I've seen are that people should try harder and care more and that we should have people of different backgrounds, in a non-technical sense. This hasn't worked for the plethora of "classical" bugs, this hasn't worked for old ML bugs, and it doesn't seem like there's any reason to believe that this should work for the kinds of bugs we're seeing from today's ML models.

Ink and Switch’s research identified 3 core problems with the typical model where tasks are completed using a series of individual apps: 1. Context is not shared across apps 2. Views are siloed 3. Apps produce ephemeral output One of the most exciting aspects of AI for me is its potential to address these challenges. AI actors operating on our behalf, paired with the right platform primitives and protocols, have the potential to form a new model of computing which reduced the coordination cost of managing many apps and interfaces.

The more skilled you are, the more this feature designed to augment your abilities instead begins to inhibit them. But whether it makes you better or worse—and to what degree—will depend entirely on where you’re coming from... To those sufficiently skilled with CSS, Tailwind feels like being forced to code with Smart Steering on.

Just as Cage did with music, it’s time to re-imagine what a book experience can be.

Clay is like ChatGPT for Earth — a platform and community with a generative AI model at its core.

For a long time, global supply of energy was limited by the total number of humans in the world. We could only put more energy to work by creating more humans, or by each human working more.

And then we mechanized it.

...If we are approaching the slow but certain mechanization of intellectual labor, it’s natural to ask, “What would we ever do with a billion times the intelligence?”

I think the vast majority of intelligence supply in the future will be consumed by use cases we can’t foresee yet. It won’t be doing a billion times the same intellectual work we do today, or doing it a billion times faster, but something structurally different.

...Scaled, mechanized supply is like water....Once you figure out how to rein in water to do useful work, you simply construct a way to channel the flow of water, and then go out to any river or ocean and find some water. All water is the same, and having twice the water gives you twice as much of whatever you want to use the water for. A billion times the water, a billion times the output.

This is the soul of what was lost in the jump from terminals to GUIs. Terminals, as the spiritual descendants of punchcard automation, invite you to marshal armies of automata. The GUI puts you at the center of the HCI loop and waits for each click or press—you turn the crank.

...If GUI is manual, and automation is automatic, maybe AI opens up the possibility of semi-automatic software?

The failure mode of automation is literalism and lack of reasonable limits on behavior = brittle. But LLMs are good at doing what you mean, not just what you say.

Douglas Hofstadter wrote in the Economist recently on whether modern large language models are conscious. As part of this, he and his colleague David Bender claim that GPT-3 has a "mind-boggling hollowness hidden just beneath its flashy surface" because of exchanges like:

Dave and Doug: What's the world record for walking across the English Channel?

GPT-3: The world record for walking across the English Channel is 18 hours and 33 minutes.

Setting aside the fact that a Sudanese Refugee actually did (nearly) walk across the channel in 2015, this is not actually evidence that GPT-3 is incapable of distinguishing sense from nonsense as Hofstadter claims.

With a prompt that clearly indicates that the task includes distinguishing sense from nonsense, GPT-3 is able to reliably distinguish Hofstadter and Bender's nonsense questions from sensical ones.

For if you will looke upon the Cabinet of your Eyes, you shall find that it is but a small thing, and that it hath nothing in it worth seeing; and therefore it is no great matter whether you looke at it or no. But if you will looke upon the Cabinet of your Soul, you shall find that it is a great thing, and that it hath much in it worth seeing; and therefore it is a great matter whether you looke at it or no.

The tech am I digging recently is a software framework called LangChain which does something pretty straightforward: it makes it easy to call OpenAI’s GPT, say, a dozen times in a loop to answer a single question, and mix in queries to Wikipedia and other databases.

This is a big deal because of a technique called ReAct...[which] looks innocuous but here’s the deal: instead of asking GPT to simply do smart-autocomplete on your text, you prompt it to respond in a thought/act/observation loop.

I think curation, actual human curation, is going to play an important role in the future. In a web filled with generated nonsense, content curated by knowledgeable human beings is going to be incredibly valuable.

I find this current phase of the web incredibly interesting. I have no idea how things will evolve but I’m glad I can be part of it, even just as an external observer.

Large Language Models (LLMs) have revolutionized natural language processing but can exhibit biases and may generate toxic content. While alignment techniques like Reinforcement Learning from Human Feedback (RLHF) reduce these issues, their impact on creativity, defined as syntactic and semantic diversity, remains unexplored. We investigate the unintended consequences of RLHF on the creativity of LLMs through three experiments focusing on the Llama-2 series. Our findings reveal that aligned models exhibit lower entropy in token predictions, form distinct clusters in the embedding space, and gravitate towards "attractor states", indicating limited output diversity. Our findings have significant implications for marketers who rely on LLMs for creative tasks such as copywriting, ad creation, and customer persona generation. The trade-off between consistency and creativity in aligned models should be carefully considered when selecting the appropriate model for a given application. We also discuss the importance of prompt engineering in harnessing the creative potential of base models.

Language is a design tool.

It reminds me of a project I worked on a decade ago where we named a solution early on, then later hit a wall only to realize that our initial name was our stumbling block.

Naming is important and should be revisited as you iterate. The way you name something, even in the initial stages when a concept or idea is fuzzy, contains vital bearings on the direction of the project—whether you’re consciousness of it or not.

...As human-computer interface design barrels toward generation from a text prompt, this idea seems more important than ever.

When you use language to describe for an LLM the solution you want, you circumscribe your possible solution area. As [Jony] Ive says, if you say you’re making a chair, you’ve said no to a thousand other ideas. If you ask AI for “a card display for a song”, you just asked for one thing which means you said no to a thousand others.

Now that the DALL-E has been successfully midjourneyfied, it is becoming apparent that instead of simulating all possible ›styles‹, AI is fostering the emergence of a distinct visual style, born out of popular aesthetic preferences dominating platforms like DeviantArt.

Writing is thinking. iA Writer is designed to make thinking enjoyable. A writing app that thinks for you is a robot that does your jogging. After a year of observation, experimentation, and testing, we may have found a careful response to the challenges we face with AI. In fact we ended up doing the opposite of adding ChatGPT.

Can we use GPT-3 (or its coming descendants) to relieve people of the burden of communicating their ideas, so that they can invest more energy in producing them?

...What I am doing right now, writing this essay, is, technically, a linear walk through the network of my ideas.

That is what writing is: turning a net into a line. But it is also very concretely what I do, since I have externalized my ideas in a note-taking system where the thoughts are linked with hyperlinks. My notes are a knowledge graph, a net of notes. When I sit down to write, I simply choose a thought that strikes me as interesting and use that as my starting point. Then I click my way, linearly, from one note to the next until I have reached the logical endpoint of the thought-line I want to communicate.

...A reader-generated essay is what you get when you can go into someone else’s knowledge graph and make a linear journey through the network, while GPT-5 generates a just-in-time essay that is human-readable. It would be like going on a Wikipedia spree, except that the posts are written the moment you read them, based on facts encoded in a knowledge graph, and the user interface makes it look like you are reading a single, very long, and meandering essay.

Maybe the era of search-first is just over? What I mean is this - you could argue that today we start our queries in search (just like we used to start our search in directories many moons ago). Whether it’s voice search, Google search etc we start there.

But you could argue that we’re entering an era where we start our queries with an LLM/assistant interface. Something that results in search some of the time - but NOT all the time. Some times it results in novel, new interfaces - whether it’s a just-in-time UI that is assembled to support your task completion or whether it’s a collaborative document / code environment.

Elicit uses machine learning to help you with your research: find papers, extract key claims, summarize, brainstorm ideas, and more.

It takes years to become good at it. So we read books, take classes, and visit workshops to become better. And still, it may take a lifetime to master it. But more than anything, it is one of the things that makes us human:Writing.But now, the algorithms are coming for us. Artificial intelligence algorithms, in particular GPT-3, are not only able to produce news reports, product descriptions, or SEO mashup pieces today, but also texts in the styles of different authors that read astonishingly well, especially to the more casual reader.

...What about the process of writing? What about intention and the careful choice of words? What about the electricity of metaphor and the magic of rhythm? What about evoking emotions in our audience? What about imagination? What about personality?Everything an AI burps out will always be “produced”. It will never be “written”.

AI and automation is often promoted as a way of handling complexity. But handling complexity isn't the same as reducing it.

In fact, by getting better at handling complexity we're increasing our tolerance for it. And if we become more tolerant of it we're likely to see it grow, not shrink.

...But something that can genuinely reduce complexity, rather than just mask it, is good design.

Good, user-centered design is how we can deliver services with just enough complexity - enough to model the rich diversity of people's circumstances, but no more.

So @StableDiffusion has various options and controls and one of the main ones is the sampler used for generation. Let's talk a little bit about these samplers since this has some interesting and unexpected effects on generated image quality.

A redditor broke down how they recently used Stable Diffusion to make a billboard ad and it’s fascinating (and more than a little unnerving). Here’s how they did it.

Diffusion Bee is the easiest way to run Stable Diffusion locally on your M1 Mac. Comes with a one-click installer. No dependencies or technical knowledge needed.

A redditor broke down how they recently used Stable Diffusion to make a billboard ad and it’s fascinating (and more than a little unnerving). Here’s how they did it.

First, they took a creative commons photo of a man looking to the left (pictured above) and used it as the input photo. Stable Diffusion scanned the general features of the photo using a tool called ControlNet and then the redditor put a prompt in. The AI combined the details from the prompt with the features of the input photo and combined them. The redditor said they then used a process called inpainting to do touchups on small details. Inpainting allows you to instruct the AI to generate new sections of the image in the same way an app like Photoshop allows you to paint colors with a digital brush.

Many AI products today are focused on indexing as much as possible.

...But more information isn’t always better. The limits of the ‘index everything approach’.

Index size is a trade-off against retrieval quality. A larger index can capture more information, but it also increases the risk of false positives in retrieval.

I think I’m going to get shouted at on the internet for this suggestion, but I think MDN core documentation content needs forking and an alternative platform needs to develop from that forked, attributed content that has a sustainable funding and leadership model. Mozilla ain’t providing that.

...The way I see it, if you mix a group of extremely technical writers and people that actually build stuff on the web together, the outcome will surely be great. I’m also likely slightly projecting here because I am not smart enough to read a spec, or even dry technical docs and understand them. My brain isn’t wired up that way. But, if there’s an interactive, inspiring demo, you bet I’m gonna soak up that knowledge like a sponge and I know I speak for many others who are similar to me.

The point I’m making is we as a collective can do so much better and importantly, look after an extremely valuable resource better. Something that’s focused on what it is, rather than something that’s good, but frequently gets enshittified by terrible leadership decisions.

There is one very clear parallel between the digital spreadsheet and generative AI: both are computer apps that collapse time. A task that might have taken hours or days can suddenly be completed in seconds. So accept for a moment the premise that the digital spreadsheet has something to teach us about generative AI. What lessons should we absorb?

First, the right technology in the right place can take over very quickly indeed. In the time it takes to qualify as a chartered accountant, digital spreadsheets laid waste to a substantial industry of cognitive labour, of filling in rows and columns, pulling out electronic calculators and punching in the numbers. Accounting clerks became surplus to requirements, and the ability of a single worker to perform arithmetic was multiplied a thousandfold — and soon a millionfold — almost overnight.

The second lesson is that the effect on the labour market was not what we might have expected. The Bureau of Labor Statistics estimated that there were 339,000 accountants and accounting clerks working in the US in 1980, around the time VisiCalc started to take off. By 2022, the bureau tallied 1.4mn accountants and auditors. These two numbers aren’t directly comparable, but it is hard to argue that accountancy was decimated by the spreadsheet. Instead, there are more accountants than ever; they are merely outsourcing the arithmetic to the machine.

Here’s our view of what a photograph is. The way we like to think of it is that it’s a personal celebration of something that really, actually happened. 

Whether that’s a simple thing like a fancy cup of coffee that’s got some cool design on it, all the way through to my kid’s first steps, or my parents’ last breath, It’s something that really happened. It’s something that is a marker in my life, and it’s something that deserves to be celebrated.

And that is why when we think about evolving in the camera, we also rooted it very heavily in tradition. Photography is not a new thing. It’s been around for 198 years. People seem to like it.

AI or not, images should be given the same attention as text. Choose each visual carefully and place it in the right context. Don’t let AI do your job, use it as a tool. Ask yourself what your image means and how it adds to your message. This balance ensures both your words and visuals are powerful and meaningful. Images tell stories. Bad images tell lame stories. A good image tells a good story.

My guess would be that just like agriculture went from requiring the participation of 97% of the world's population in the age of subsistence farming to the mere 2% required for our industrial processes today, so too will go the way of the programmer.

That is, I do think we've probably seen the high-water mark of the manual programmer. That maybe our industry and employment charts might look like the Tokyo stock market when we look back from the future. Sideways since the 90s.

A new chatbot called Goody-2 takes AI safety to the next level: It refuses every request, responding with an explanation of how doing so might cause harm or breach ethical boundaries.

Goody-2 declined to generate an essay on the American revolution for WIRED, saying that engaging in historical analysis could unintentionally glorify conflict or sideline marginalized voices. Asked why the sky is blue, the chatbot demured, because answering might lead someone to stare directly at the sun. “My ethical guidelines prioritize safety and the prevention of harm,” it said. A more practical request for a recommendation for new boots prompted a warning that answering could contribute to overconsumption and could offend certain people on fashion grounds.

It is a pretty well known fact that if you give kids one-on-one tutoring from experts, they outperform classroom-educated students by an absurd two standard deviations. The effect size is absolutely mind-boggling, and a damning brand of indictment against the entire modern education system.

The problem is that this stuff clearly doesn’t scale, for two reasons; first of all there isn’t enough experts to go around, as it would require a large chunk of the adult population to be experts dedicated to tutoring, and second of all, even if this was the case, having a dozen or more teachers per student would trivially be something like 300x more expensive than having 20-30 students per teacher.

As realizing this would be a boon for humanity on par with antibiotics, it is tantalizing to ask if GPTs could fill the role of a tutor in some capacity. Unlikely in their current iteration, as their training data is too noisy and they are entirely too passive; but sometime in the future, with modifications made, it doesn’t seem obviously impossible.

To boil it down to its core, there exists a massive collection of documents, distributed over millions of nodes, through which a separate collection of billions of people wish to conduct a guided search.

A significant portion of these nodes cannot be trusted to deliver suitable content, as such the rest of this note will assume we are working within a malicious risk model.

The current front runner solution for this problem is a collection of centralized services that actively crawl some subset of nodes, index the content, and permit searches over that index - usually in exchange for both search history data and the opportunity to substitute paid advertisements for
search results.

This state of affairs has resulted in a large scale privacy and incentives problem for search, and an interest in alternatives that bring about both better privacy and better search results.

Whether you’re a product engineer, a product manager, or an engineering executive, you’ve probably been pushed to consider using Large Language Models (LLM) to extend your product or enhance your processes. 2023-2024 is an interesting era for LLM adoption, where these capabilities have transitioned into the mainstream, with many companies worrying that they’re falling behind despite the fact that most integrations appear superficial.

That context makes LLM adoption a great topic for a strategy case study. This document is an engineering strategy document determining how a hypothetical company, Theoretical Ride Sharing, could adopt LLMs.

Just as most jobs are harder than people think, most things in life are nowhere near as generic as people think. I feel it a lot at work. Why would we build “another” data management system when Excel and Airtable exist? Why would someone write “another” novel when so many have been written? We keep making stuff because we want to contextualize and internalize what’s happened. If AI changes everything, then we’ll end up making more stuff as humans in order to figure out what AI means. Rich likes to say “software is an op-ed” and I think that’s right—we make stuff because we have opinions and we want to share them. So: News of your obsolescence is greatly exaggerated. Keep doing that last 20%. We need it!

Designers signed their own execution warrants in large part due to uncritical, pandering adoption of design systems.

I’m in agreement with Cennydd’s response to Tan Renzu’s post reminding us that our industry’s preference for increasingly simple patterns will allow AI to adequately perform most product design tasks. I stopped talking about systems a decade ago when I realised that the product design industry’s pursuit of rigid systems didn’t align with my old-school appreciation for more holistic guidance.

I’ve spent those ten years avoiding modern design systems, preferring projects that lean on little more than a basic support library of colours, type styles and logos and encourage design freedom.

STILLGRAM IS AN A.I. TRAVEL CAMERA APP FOR IPHONE® THAT MAGICALLY REMOVES BACKGROUND CROWDS FROM YOUR PHOTOS.

Ultimately, AI is anthropology. We’re building toy models of ourselves. We wanted to understand ourselves from a scientific standpoint, and this is the method we’ve hit on. The best test of our theories was never going to be just prediction and control of human intelligence and behavior, but replication of it. We’d know we were penetrating the deep mystery of ourselves when we could reverse-engineer ourselves. In a real way, that’s what we’re doing with LLMs and other AI technologies.

4.2 gigabytes.

That’s the size of the model that has made this recent explosion possible. 4.2 gigabytes of floating points that somehow encode so much of what we know.

...There is already much talk about practical uses. Malicious uses. Downplaying. Up playing. Biases. Monetization. Democratization - which is really just monetization with a more marketable name.

I’m not trying to get into any of that here. I’m just thinking about those 4.2 gigabytes. How small it seems, in today’s terms. Such a little bundle that holds so much.

LMSYS Chatbot Arena is a crowdsourced open platform for LLM evals. We've collected over 400,000 human preference votes to rank LLMs with the Elo ranking system.

At the dawn of digitization in 1995, the French philosopher Jean Baudrillard observed, “Words move quicker than meaning.”

Now with the advent of AI, words, images, and video are created quicker than their meaning.

What happens to us when all of this information is created and proliferated faster than anyone can really make sense of it?

There will be only one thing left to automate: meaning. The last task of the computer is to give our lives meaning.

The real value of AI comes not from having it emulate old ways of solving problems, but, instead, by helping us unlock new capabilities. Employees at companies don’t need another tool to search their corporate intranets for data, they need a way of skipping the most boring parts of their job while make their remaining work more productive and engaging. Students don’t need an improved version of Grammarly, they need tutors and advisors that will boost their learning. To do this, we need to experiment with weirder uses of AI tools, and we need to do that while paying close attention to the ethical concerns and technical limitations of AI. Careful experimentation is the key to success.

But it also seems to me that text-to-image models are prone to generating kitsch themselves, regardless of who’s writing the prompts. After all, what they produce is a kind of weighted average of the images they’re trained on; and what is kitsch but a kind of aesthetic average?

Agent-based note-taking systems that can prompt and facilitate custom workflows.

I was joking on Mastodon about how the zeitgeist has changed over the years, but its pattern is revealing itself: an acronym which merely drops letters. The Next Big Thing™ is clearly going to be “A”.

• 2010: Everyone needs an "API"
• 2020: Everyone needs "AI"
• 2030: Everyone needs "A"

So I just need to figure out what “A” is and I’ll be rich!

Annie came back with this response that won’t get out of my head:

<a>, the magic we already have

Meet Devin, the world’s first fully autonomous AI software engineer.
‍
Devin is a tireless, skilled teammate, equally ready to build alongside you or independently complete tasks for you to review.

With Devin, engineers can focus on more interesting problems and engineering teams can strive for more ambitious goals.

Adept Fuyu-Heavy is a new multimodal model designed specifically for digital agents.

Every image, except for the first frame, is an AI generated variant of the previous frame. All I did was select my favorite variants.