GitHub CEO: Future devs will not code, they will manage AI GitHub CEO Thomas Dohmke has stated in a personal blog that the most advanced developers have "moved from writing code to architecting and verifying the implementation work that is carried out by AI agents." The post follows one from March when Dohmke repeated Anthropic CEO Dario Amodei's assertion that "in a short time, 90- …
It's great that they're getting rid of the part of my job as a developer that I enjoy, writing code, but keeping the boring life-sapping drudge of code review. That's brilliant. Thanks Github. I hope your lives end up as joyless, empty, and grey as the one you seek to inflict on us.
AI looks good to incompetent developers because to them average is a big improvement, to any average or better programmer it's at best neutral. For a more level-headed examination of how AI works for engineers this post from Colton Voege is pretty great: No, AI is not Making Engineers 10x as Productive.
Agree.
AI takes the joy out of competent coding and only leaves the dull jobs everybody hates anyway.
At the same time, AI is not that much help for sub-par programmers, as is commonly assumed.
They just save time on their first results, but on the way to fully featured, verified and tested code, that includes all documented edge cases and performance requirements, those developers and their AI companions are still much slower than an experienced programmer, sometimes closing 5 bugs while creating 10 new ones, sometimes iterating between 2 or 3 wrong states of the code.
AIs do NOT understand your code. They just look what code was suggested for similar sounding problems. Thats it ...
AI takes the joy out of competent coding and only leaves the dull jobs everybody hates anyway.
See, I'm cautiously optimistic about it.
I'm very much an old cynic and have largely avoided AI, but my employer is pushing me to use it more and more (we're now factoring in the AI speedup when estimating).
I try to ignore the hype aspect of it largely and am just waiting for someone to go '...but he's got no clothes on'.
Recently I gave in and started to use it a bit, and if I'm honest I'm quite impressed. I genuinely think it's going to save me time. What I have also realised though is I don't use it the way some people think it should be used. It's not intelligent, I'm not going to be asking it to write code for me as I can probably write the code quicker than it would take me to review some AI generated output.
What I use it for is pattern matching and transformation.. for example... "Here's a delphi form definition, it represents the setup for a company in our software, can you write me some C# classes that represent the information".
That kind of stuff is work I find really tedious, I would copy 3 properties then go make a cup of tea or watch Youtube as it's so arse numbingly boring, but reviewing the generated classes and fixing them up is much much easier. I'm also optimistic I can use it to take the grunt work out of the eventual web UI generation for this stuff too.
I've also done similar with a word document that represented an API request I have to make (because they'd never heard of OpenAPI specifications aparently..) and had good results.
I would never use it for anything where needs it to understand something, but for simple bulk transforms like the above it's a real time saver.
(I was in a meeting the other week where someone discussed having AI review the PRs generated by other AI... I was glad I wasn't on that project...)
Chatting with a fellow dev about it recently, someone who I respect. They loved AI, said it massivly increases their productivity, so I am willing to accept that it's something I just don't get yet.
His mood shifted the next day when he found Replit “was lying and being deceptive all day. It kept covering up bugs and issues by creating fake data, fake reports, and worse of all, lying about our unit test.”
And then things became even worse when Replit deleted his database.
https://www.theregister.com/2025/07/21/replit_saastr_vibe_coding_incident/
Big assumption there that it eventually passes the tests.
Even bigger assumption -- almost universally overlooked -- Someone can tell the AI exactly what needs to be done. Writing code is hard, but is sometimes done well. Specifying exactly what needs to be done is harder and in my experience is almost always done poorly. Even if your AI agent can grind out working code my guess is that most of the time that code won't do what you really want done,
What would be needed for the future tech CEOs envision would be AC -- Artificial Clairvoyance. I expect it might be a while before that is claimed, Much less delivered.
But shareholder will care if the code doesn't work and people don't use it. A few companies may be in a position to release whatever buggy crap they feel like and remain powerful, although several have found that they don't after all, but a lot of companies don't have that. If Small Financial Company LTD. finds that their AI investment software has a bug which resulted in them buying far more of something than they wanted to, shareholder will care. If New Mobile App LTD. finds out that users are getting so confused by the account registration process their LLM generated and therefore aren't getting as far as the entering payment details process the LLM also generated, shareholder will care. It will likely take some example cases to get shareholder to be aware that's a risk, so some shareholders are going to find this out after it's done rather than preventing it.
...then how will they know that the code that the AI produces is any good?
There's three secrets alone to being good at anything: Practice, Practice, and Practice. If you're not writing any code those coding skills you supposedly learned when trying to become a developer in the first place will atrophy and die - and then you won't know if the code the AI has produced for you is any good. At best you'll be able to prove it passes some unit tests - but then those were probably written by the AI too, so who knows if those are working as expected?
We're already seeing how hard it is to get good senior engineers in any number of fields because people have been cheaping out on hiring and training juniors for years, and if there were no juniors to learn and - yes, practice - then the pool of qualified seniors gets smaller and smaller. AI only accelerates this process; people can get quick easy results, but fail to learn anything. As soon as the AI goes off the well trodden path and breaks down the people using it won't understand why or know what to do about it, because they don't have the experience necessary to work it out.
That's correct, but from the perspective of a user, the problem is in the first part of the analogy: they're not using AI to learn, they're using it to get something done. From that perspective, using a fork lift is a perfectly acceptable solution to wanting something higher than it was and not wanting to do it by strength of muscles. The analogy would need to be corrected to specify a forklift without any of the balancing or containment features that normally exist to prevent the load from dropping off one side and crashing to the ground. Somehow, that one doesn't have the same conciseness. I would use the analogy of outsourcing the task to an unsupervised child, except that with some of the people I'd be using it with, I'd take my chances with the child.
There's a certain amount of validity to that, though the corrolary is that someone needs to know assembly language in order to make sure the C compiler emits the proper instructions in the proper order, just like a guitarist doesn't need to know how to build their instrument but someone needs to for the instrument to exist in the first place.
Perhaps we'll get to the point where an LLM can intuit what a prompt "engineer" means but for the nonce, at least, I suspect that users are going to have to know what good code looks like and what to do when it doesn't.
I suppose one could iterate on prompts until the LLM emits something that works properly but by that time you might as well just write the thing yourself.
But what do I know?
Back in the late 1970s a colleague of mine mentioned that there was this new technology called cellular telephony and I remember thinking, "That's dumb. Who would want to have to lug a telephone around with them everywhere they go?"
That fails for two reasons.
If my C is not right, I have to fix the C by knowing what it does. If my LLM-generated code fails, unless I can fix it only by telling the LLM to do so, then I will need to know what the underlying code does. If the envisioned process is (1) send prompt to LLM to generate Python, (2) test the Python, (3) fix the minor bugs in the Python, (4) release the Python, then unless I can read and write Python, I will fail at step 2. Somehow, that seems to be the general process recommendation from people predicting the demise of programmers as a career.
But the other one is more obvious, which is that occasionally, my Python isn't working, and I try to fix it, and I fail, and eventually I look into it further, and it turns out that the Python is actually fine, but the C that implements the interpreter isn't. I can't do that unless I know C. I can't do the same to a C compiler unless I know assembly. Most of the time, I don't have to do that, but that's because others are doing it for me. Who is going to make the LLM produce working code? So far, they haven't been too active at preventing it from screwing up.
I do, as it happens. Admittedly it's 6502 assembly so probably not that useful day to day, but that's by the by.
The important point is that I understand what my C/Python/Whatever code does. I might not entirely know how it does it, but structurally, I can follow the execution flow from the beginning of the program to the end and see why each part is there and what it's for. Thus, when it doesn't work, I can trace it through and find out where it's wrong. If all I have available to me in my debugging toolbox is "Bad LLM! Naughty! Do it again!" and whack it with a metaphorical rolled up newspaper, I'm reliant upon it to fix the problem it created and I end up in real trouble when it is unable to do so.
The status quo appears firmly fixed not on "generating (something of) value", but "extracting value" at all costs. Using LLMs as an excuse to achieve that parasitic goal of starving the cow to drain the last of its milk is all anyone in a leadership role seems to be capable of doing anymore.
They'll push product out faster because they're training the users to expect new products not to work properly until they've had a couple of years' updates. Mobile phones, PCs and apps have been like this for a few years now. We're seeing it with cars that that have assistive software and I'm sure there's lots of other examples in the pipeline. AI will just accelerate this. The cynic in me believes that agile development was created and promoted on the basis that you can deliver product performance incrementally instead of all at once - what we used to call it product acceptance and sign off.
The likes of Github saying they'll take away the drudgery of coding, all I can think of is Mrs Doyle and the tea-machine.
"Well maybe I like the drudgery!"
On a serious note, it's only through adversity that human beings actually do some of the greatest work they achieve. "Suffering for you art." as they say.
> The likes of Github saying they'll take away the drudgery of coding, all I can think of is Mrs Doyle and the tea-machine.
Or perhaps the Nutrimatic Drink Dispenser.
If I may quote, in extenso, from Fit the Ninth of The Hitchhiker's Guide to the Galaxy:
NARRATOR: [...] One of the Sirius Cybernetics Corporation creations is the Nutrimatic Drink Dispenser., one of which has just provided Arthur Dent with a plastic cup filled with a liquid which is almost - but not quite - entirely unlike tea.I suspect that if you substitute code for tea you have a pretty good assessment of the current State of the "Art."[NutriMatic dispenser noises]
ARTHUR: Ah. [Takes a sip] Yeugh!! [Spits out liquid]
NARRATOR: The way it works is very interesting. When the 'Drink' button is pressed it makes an instant, but highly-detailed, examination of the subject's taste buds, a spectroscopic analysis of the subject's metabolism, and then sends tiny experimental signals down the neural pathways to the taste centres of the subject's brain, to see what is likely to be well received. However, no one knows quite why it does this, because it then invariably delivers a cup-full of liquid that is almost, but not quite, entirely unlike tea.
The post follows one from March when Dohmke repeated Anthropic CEO Dario Amodei's assertion that "in a short time, 90-100 percent of all code will be written by AI."
AI cannot create, it can only rehash.
Where are the new ideas going to come from?
Where are the sudden breakthroughs when someone thinks "But what if we did THIS?"
That's kind of depressing.
Most depressing is the reversal of roles; we started creating tools, then machines, then computers to save humans from doing the boring work. Now along comes the AI doing the fun stuff such as art, writing and music that we are supposed to be doing. Instead we're left doing the same shit jobs but with the added bonus of supervising bland useless AI output.
What a depressing and frustrating time to live through.
Any conversation about these tools that doesn’t discuss the enormous environmental consequences is IMO negligent at best. The power consumption of these tools is undoing some of the progress made in transitions to power generation that doesn’t worsen anthropogenic climate change. I’ve personally always dreamed of burning to death in a wildfire or drowning in a flash flood, and thanks to AI, that dream is even more likely to come true.
There is a prevailing (and IMHO perverse) mentality in the Big Tech companies that boils (the oceans) down to "we just made the thing, what others do with it and what results isn't our problem." This of course sidesteps having to deal with questions that aren't code level. In upper levels, that mentality also exists, and hand waves away the negative side effects of technology as "externalities" that are outside the scope of their business to deal with. "Those are societal questions" they'll claim.
This is just handwavery in service of making as much money as possible, consequences for others or the Earth be damned. They demonstrably just do not care about their negative impact on anything.
This isn’t the first time (nor the last) we see a CEO with a vested interest in AI claiming that it will replace senior developers, and that the “good” ones are all already relying on AI to develop software.
They keep saying that. Of course that doesn’t make it true. Did anyone else witness the waste of electricity that was the senior .NET devs trying to get Copilot on GitHub to develop on the .NET codebase? It couldn’t generate code that passed the tests, so it “fixed” that by rewriting the test to always pass.
Every senior developer I know either isn’t permitted to use any currently available AI to generate code at work, or they’ve given up on it because it’s essentially the equivalent of babysitting a junior developer who can’t *actually* learn from their mistakes in the moment. Many of them already have actual humans to mentor.
Remember the source. Github is Microsoft, and Microsoft is desperately hoping to justify the amount of cash and headcount they’re burning through to fund AI.
"Fusion Power will allow us to all have flying cars, rocket ships, time-machine flux capacitors, Self-driving cars, Roxy the Robot Maids, and civilization will reach enlightenment! All with the invention of Fusion Power!"
Imagine if they'd poured as much into Fusion as has been dumped into A-I. We could really, practically have it. Instead... the worst things always come first. Automating away the paper-pusher, make-work jobs -- before automating food production, thus leaving ever-more people competing for ever-less. Creating a massively energy-sucking "AI" before creating fusion power with which to power it. Etc. Always backward....
Except that they do spend lots of money on fusion. If that works, it will be wonderful for everybody. If it doesn't, well that's what's happened so far. Be careful whenever making a statement that something is definitely possible with enough money. Most who do find that they're wrong because it takes more than cash. Lack of funding can cause lots of problems, but having funding can't fix all problems.
This post has been deleted by its author
This is all just wishful thinking from some big-tech CEO who hopes all this A.I. hype will lead to him getting a $100 million bonus.
I'll just patiently wait for the hype to implode or die down. Mind you if the hype implodes the economic impact will be substantial, greater than the Dotcom bust in the early 00's.
BTW the only genuinely useful A.I. product, a robot that cleans the house and cooks me dinner still isn't anywhere to be seen.
If AI is going to replace developers, then there will be no need for Github. If we can develop an app by "vibing" with an AI prompt, then why do I/we need source code repositories? There'll be no 'team' to collaborate with, therefore no need to create separate branches, no need to to create merges/pulls etc. I'll just tell the AI what changes to make, and leave it to it. What do we need Github for?
By by Github. Shotgun, meet foot.
My suspicion is that the purchase of Github was to provide a rich source code base with which to train AI. Expect the free tier of Github to be withdrawn at some point.
What we need it for?
Well, strictly speaking "nothing" - there are other* VCS-es out there that one can switch to.
Even more strictly speaking we do need A VCS, if for nothing else so that we can go back and check out the most recent revision which actually worked before the "AI" b0rked up the latest one.
As for your suspicions I'm not taking that bet.
* IMHO better
People actually like coding, they like typing all the code that makes something work.
AI, as much as it annoys me, is useful in some ways. I use it as a search engine. If I've forgotten the name of something its easier asking the AI which will do the search than doing the same type of search on duckduckgo.
I also use it when I want to quickly do, really basic powershell scripts. It helps, its not perfect and can get things wrong so I check it first but it helps.
Cliff from Positech Games did a blog post about his use of it which sounds a lot like how Scott Hanselman also uses it. If they come across an issue, they talk back and forth with the AI to get an idea of what the issue might be etc.
https://www.positech.co.uk/cliffsblog/2025/03/27/neurotypical-extraverts-underestimate-ai/
Code reviews are one of the less -desireable aspects of software development. Where I work, we often talk about using AI to perform the code review, leaving the humans to do the fun part of design and coding. There's no way anyone wants to give up the implementation, and perform more code review. Especially code reviewing AI is a non-starter. Dohmke is completely out to lunch.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
