AI skeptics are looking increasingly funny by the day. I was one of them as recently as a few months ago. And I ain't no cheerleader. I think this is going to create a world a million times worse than the one we have.
I don't know man. I just don't think the hype around the new models is fake, at all. Too many people with absolutely no incentive to hype it feel the same way. I think we've just reached the point where the skeptics are covering their ears and thinking that the train won't hit them if they try not to hear it. I'm done with the skeptics just dismissing everything as "lol, he has incentive to hype it" when if anything the timelines they gave a year or two ago are being proven *conservative*.
With all due respect, what we have seen is that the timelines haven't held up.
Below are the items that I have seen not materialized within the stated timelines
- In January 2025, Amodei stated that within 3 to 6 months, AI could be writing 90% of all code, and that within one year, it would be writing "all of the code. AI is not currently writing all code, or even 90% of code.
- Sam Altman said in 2019 "Artificial general intelligence could be built within the next five years.” Well it's been 7 years, and no AGI. In fact, no one seems to agree on what AGI even is.
- Elon Musk said that AI would surpass human intelligence by 2025. As of today, there is no evidence of a system that exceeds human general intelligence.
- Fei-Fei Li said that “By 2023, AI assistants will manage daily scheduling, email triage, and travel bookings for most professionals.” Yet agents are still super buggy and we are JUST experimenting with this kind of technology in earnest with Open Claw (I'm actually very excited about OC and how we can use it to perform simple tasks).
If you're going to make claims like "when if anything the timelines they gave a year or two ago are being proven *conservative*." please give actual examples.
Now now, cease being reasonable and trusting your experience immediately!
Serious: Gary, come on dude. I’m grateful you admitted that Ai coding has improved over winter break. It really has. But the straw-men of relative perfection you seem to persistently use to dismiss the positive use cases and successes is feeling like a schtick rather than thoughtful gut-check meant to balance hype.
But he provided zero actual examples. Didn't point to any links, showed us anything, gave us access to look, or proved that anything he's done with it is as good as he is saying.
Speaking of Moltbook, shouldn't it have solved something by now... Maybe calculating tolls? Or solving nuclear fusion? Or how about the human vs agent authentication issue? 1.5+ million agents have to be solving something right? It's the proverbial village!!!
Microsoft analyzed my Form for a closed survey without me asking (don’t need suggestions on how to improve the questions at this point 🧐) but offers no affordance to quickly see who hasn’t responded. There are so many little things you’d think they’d solve. Like, Apple, if another podcast reshares an episode of a podcast I’m already subscribed to, it doesn’t need to be in my feed twice.
Apparently these people just wait until their failures are memory holed and move onto the next sensationalist prediction to accomplish whatever is on their agenda (in his case, attention, which I guess he succeeded at).
I am again reminded of hype during 1995-2000. "Internet will give me perfect answers on any question in a few minutes" (literally said to me during a radio interview). Or "internet will remove the need to have representatives in a democracy, everybody can just take part in decision making" (literally written in a serious newspaper op-ed). Internet has been a massive force, but not quite like that and not only for good.
Most of the criticisms here are valid for how most people use AI coding tools. But I think there's a category of user being missed.
I'm building alone a production SaaS that would normally require a small team. "Let it rip and hope" doesn't work. But that's not what I do.
I treat AI like a junior developer who needs structure: detailed codebase documentation, code review on everything, typed interfaces, pre-commit hooks, tests, git discipline, rigorous testing. I haven't removed myself from the loop - I've changed my role from writing code to managing code production.
Is it perfect? No. But my alternative isn't "write it myself at a comfortable pace with a team." It's don't build it at all, I don't have the resources to build a system like this alone, I would need to fundraise first.
In my opinion, the hype is wrong that it's effortless. This article is wrong that it doesn't work. The truth is in between: it works, but it takes discipline, experience and knowledge. Not magic.
Very true! It is what I call responsible application of AI. But the approach is on the short term much slower than the scenarios that Marcus is so rightfully critiqueing. And that is not a bad thing. Because it will pay off on the long term, Because you don’t have to spend exponential amounts of time fixing compounded technical debt.
The most important question to ask a vibe coder is “can I test it out?” That’s where the rubber meets the road. This guy, Gary Tan, Karpathy or any of the other vibe coding influencers have yet to share a fully functional product. They’re all buggy prototypes.
Gary, you should take a look at FastRender (from Cursor) and claudes-c-compiler. Both are insane failed attempts at automating software engineering work on a serious projects.
Both companies shared the project with naive arrogance only to be humiliated by real engineers and experts reviewing the actual code.
I code using the latest llm technology and it's nowhere near the hype in that blog post. There's a reason LeCun doesn't believe llm's in their current state can reach AGI. And it's easy to see why once you start using them extensively. The marketing hype right now is far beyond their capability.
This whole event reminds me so much of the fast foods industry. Most people are now realizing that cheaply made and highly processed foods are unhealthy and a long way from what your caring parents might have fed you. Real food, prepared with knowledge and care is so much better for you. But, as we know, millions of people eat junk food which is heavily promoted. This also runs over the point that Gary and others have been making for so long: it takes real knowledge to judge the quality and usefulness of AI outputs. The real thinking is in the head of the user who judges the usefulness, errors in and quality of the output. The hype goes on.
What I fail to see is why the risk of compounding technical debt is hardly ever discussed? These systems generate code faster than any human can review. Having it reviewed by AI is fooling yourself. So we let AI generate code that contains errors we don't (take time to) understand, and then let AI fix them. And it works for a while (just like the next heroine hit works for while). And more errors creep in. And we let AI fix it, or at least the AI generated tests pass. In the meantime technical debt creeps (well that is not the right term in this case, rather 'storms') in without us noticing it. This keeps repeating itself until it (AI) cannot fix the errors anymore. Neither can the humans because they never internalized the architecture and code that were generated.
Obviously, and ironically, the better the systems get (as they undoubtedly do) the longer it takes to reach that point of no return.
By the way, I call this approach "Agent Addiction" because the process has a lot of analogies with what junkies go through.
My point, if we delegate these amounts of code generation to AI there is this very dangerous risk of technical debt explosion, much faster than technical debt happened to accrue in traditional software development. Are we going to wait for 6 to 18 month to find this out?
Then there is the other approach, people who are aware of the above risk, they will have to slow down, they will have to take time to review and internalize everything AI generates. On the short term this will (look to be) be orders of magnitude slower than the former approach. But possibly makes them come out as winners after 6 to 18 months. But there is another catch. The people needed to do this review and internalization are experienced software engineers that spent years and years of grinding real software experience. Grinding the skills needed so badly to responsibly make use of AI. But junior engineers come out of school and the majority of them lands in places where they are not offered (or not taking) the opportunity to learn those skills (because AI), which can only be learned by going through that grind.
Basically this means that the skill to responsibly apply AI is slowly waning and not being replenished. This makes our jobs (senior software engineers) in even more demand the coming years. I see a few possible mitigations
- We realize this is happening and will be actively encouraging juniors to keep going through that grind. This is the responsibility of software engineering companies.
- We wait for AGI to actually become as good as a senior developer and that specifically includes long running cognitive tasks (not those stupid smarter than PhD benchmark comparisons) which are absolutely crucial to build reliable systems and even more in order to fix the issues that result from compounding technical debt caused by Agent Addiction. Especially as long as the code is generated by LLM based technology.
Actually I am not really sure about the short term viability of the second option. We currently do not have an agreed upon definition of true AGI let alone a way of measuring whether we have achieved it, and even if we did, there are probably quite a number of systems for which we do not want to trust its internals to AGI only.
Final accountability should always lie with humans, so we need humans to fully understand the systems we build.
In fact it's not just technical debt build up that no-one's talking about - it's any kind of programming principles at all (Murphy's law, MVC, DRY etc.) People seem to have forgotten we have them. Or perhaps never knew about them in the first place. (A large percentage of the very vocal proponents of AI generated code probably aren't coders and never got that particular memo.)
The LLMs distressingly seem to have anti-principles. RYA - "repeat yourself, always!" MICS - "make it complicated, smartypants!" and other gems like, "if it doesn't pass the tests, meddle with the tests!" and "if a requirement in the spec is difficult, skip it and leave an obscure comment in the code, then say you did it anyway!"
You bring it like overstated sarcasm, but the sad truth is that this really happens regularly. I am really curious if we will see a technical debt epidemic months from now. Hopefully the majority of senior engineers have acknowledged these problems and are taking it slowly because of it. Nevertheless numbers show that AI usage by software engineers has grown substantially the last year. But it is hard to guess what the main usage patterns are.
There's much more realistic approach: just let all these things to go till the whole thing would collapse and things that are worth doing rewrite for would be rewritten from scratch.
The majority of code ever written is thrown away, after all… with LLMs we would just learn to accept that.
We already had this situation with freelance-contractors produced code which is often brought to the state where nothing can be fixed and you need to throw it away, AI would just mean that we would have more such throw-away unfixable code, it's not the end of the world.
I've become fluent in ChatGPTese recently - and this essay is definitely largely written by it.
What I think it shows is the way ChatGPT very effectively convinces people that it's turning them into superhumans.
This man - Matt Shumer, that is - plainly believes he's written a really very important essay, quite possibly for all humanity - when in fact it's terrible.
I've noticed this a lot recently - people becoming more dependent on AI to do things we really need to do ourselves as humans - for example relating to other people or articulating our opinions. and it happens very quickly -and these people have a completely impaired view of the quality of work they are producing; it profoundly impacts on their judgement.
This is the angle where things get more interesting. I mean, there's definitely going to be a sort of atrophy that happens for people that rely so much on automation and AI to live their lives for them. What's the wave that happens after the atrophy? Will the people who never got swept up in the hype start to appear like wizened masters simply because they can remember how to do things the hard way -- and because they still have the desire to engage meaningfully and creatively in their lives? Or something else?
Amazing! , I bet it also makes mounds mounds of julienne fries!
As I was reading the BS utterly perfect scenario, all I could think of is how most IT projects fail because of bad requirements. Which made me laugh. Because then the weakest link in this fantasy scenario would be whether he truly gave the AI perfect instructions. So I guess the next step is to build an AI that can come up with a perfect instructions so we don't need the human after all.And all you have to do is think at it. Because the moment you open your mouth, you'll be the source of error.
Put it to the test. Sit down with Claude and recover from a post compaction event while it's writing files to its own VM. While we all watch, live. If it fails - Deer Hunter Roulette - if it works but loses its mind - you lose a finger. If it does something bad but not fatal, the RFK Jr "treatment"
If people lived this every day they would not have these ludicrous greed-stoked ideas informed by nothing. Come down into the trenches, hype-meisters, where the real folks live and join the migrant techno-peasantry for a day.
This is the same viral poison, devoid of any substance, designed only to make people’s hearts jump. I wouldn’t be surprised if this was written by an LLM, iterated many times to regurgitate something meant to farm 'views' and 'engagement.'
It’s sad that a field like AI has become a subject of mockery and infantile content that brings nothing of value to humanity. We are wasting so many resources on spectacle.
It is just marketing nonsense. There is a reason why they provide precisely zero detail about what the LLM actually does
indeed, i should have made that point!
It's HAL on Speed.
This HAL needs to lugs its 100GW nuclear reactor with it.
😂
he described the details specifically: it wrote apps, tested them, and produced a perfect troubleshot app for him, with zero oversight. That's AGI.
Good meme. I appreciate the joke
did you even read the post? Do you use AI? Claude?
(1) "did you even read the post?"
Yes, I did. It is why I said what I did and am making fun of you for your dishonesty.
There is a reason why you provided no details in your attempt to say I was wrong.
(2) "Do you use AI?"
Yes.
(3) "[Do you use] Claude?"
Yes.
...do you?
AI cheerleaders are funny.
AI skeptics are looking increasingly funny by the day. I was one of them as recently as a few months ago. And I ain't no cheerleader. I think this is going to create a world a million times worse than the one we have.
I think one can be skeptical of LLM's technical capabilities while also agreeing with your last point there.
I don't know man. I just don't think the hype around the new models is fake, at all. Too many people with absolutely no incentive to hype it feel the same way. I think we've just reached the point where the skeptics are covering their ears and thinking that the train won't hit them if they try not to hear it. I'm done with the skeptics just dismissing everything as "lol, he has incentive to hype it" when if anything the timelines they gave a year or two ago are being proven *conservative*.
With all due respect, what we have seen is that the timelines haven't held up.
Below are the items that I have seen not materialized within the stated timelines
- In January 2025, Amodei stated that within 3 to 6 months, AI could be writing 90% of all code, and that within one year, it would be writing "all of the code. AI is not currently writing all code, or even 90% of code.
- Sam Altman said in 2019 "Artificial general intelligence could be built within the next five years.” Well it's been 7 years, and no AGI. In fact, no one seems to agree on what AGI even is.
- Elon Musk said that AI would surpass human intelligence by 2025. As of today, there is no evidence of a system that exceeds human general intelligence.
- Fei-Fei Li said that “By 2023, AI assistants will manage daily scheduling, email triage, and travel bookings for most professionals.” Yet agents are still super buggy and we are JUST experimenting with this kind of technology in earnest with Open Claw (I'm actually very excited about OC and how we can use it to perform simple tasks).
If you're going to make claims like "when if anything the timelines they gave a year or two ago are being proven *conservative*." please give actual examples.
Timelines for what, exactly?
Now now, cease being reasonable and trusting your experience immediately!
Serious: Gary, come on dude. I’m grateful you admitted that Ai coding has improved over winter break. It really has. But the straw-men of relative perfection you seem to persistently use to dismiss the positive use cases and successes is feeling like a schtick rather than thoughtful gut-check meant to balance hype.
But he provided zero actual examples. Didn't point to any links, showed us anything, gave us access to look, or proved that anything he's done with it is as good as he is saying.
The Venn diagram of people who were shocked about moltbook and are glazing this article is a 1:1
Sobering
Speaking of Moltbook, shouldn't it have solved something by now... Maybe calculating tolls? Or solving nuclear fusion? Or how about the human vs agent authentication issue? 1.5+ million agents have to be solving something right? It's the proverbial village!!!
But nope.
Microsoft analyzed my Form for a closed survey without me asking (don’t need suggestions on how to improve the questions at this point 🧐) but offers no affordance to quickly see who hasn’t responded. There are so many little things you’d think they’d solve. Like, Apple, if another podcast reshares an episode of a podcast I’m already subscribed to, it doesn’t need to be in my feed twice.
Moltbook is a performance art piece. Is it compelling performance art? To some.
It's very techbro coded, a performance made by a techbro to delight other techbros, catering to the techbro gaze, as it were.
He's a proven fraudster. Remember when he said he built the "best model in the world"? It's right on his Twitter feed.
https://venturebeat.com/ai/reflection-70b-model-maker-breaks-silence-amid-fraud-accusations
Apparently these people just wait until their failures are memory holed and move onto the next sensationalist prediction to accomplish whatever is on their agenda (in his case, attention, which I guess he succeeded at).
that’s linked in my essay of course
That it is! I read too fast. 😅
The link in the essay is broken.
pretty transparent agenda seeing as he is the CEO of an AI company.
An AGI will never forget that!
It's all about managing our context windows effectively.
It reads like he is trying to sell a scamcoin or some sort of men's group retreat where you dance naked in the forest.
I read the first 10000 lines of obsequious blathering and assumed that Matt is an influencer earning his 800K influencer fee.
I am again reminded of hype during 1995-2000. "Internet will give me perfect answers on any question in a few minutes" (literally said to me during a radio interview). Or "internet will remove the need to have representatives in a democracy, everybody can just take part in decision making" (literally written in a serious newspaper op-ed). Internet has been a massive force, but not quite like that and not only for good.
Most of the criticisms here are valid for how most people use AI coding tools. But I think there's a category of user being missed.
I'm building alone a production SaaS that would normally require a small team. "Let it rip and hope" doesn't work. But that's not what I do.
I treat AI like a junior developer who needs structure: detailed codebase documentation, code review on everything, typed interfaces, pre-commit hooks, tests, git discipline, rigorous testing. I haven't removed myself from the loop - I've changed my role from writing code to managing code production.
Is it perfect? No. But my alternative isn't "write it myself at a comfortable pace with a team." It's don't build it at all, I don't have the resources to build a system like this alone, I would need to fundraise first.
In my opinion, the hype is wrong that it's effortless. This article is wrong that it doesn't work. The truth is in between: it works, but it takes discipline, experience and knowledge. Not magic.
Very true! It is what I call responsible application of AI. But the approach is on the short term much slower than the scenarios that Marcus is so rightfully critiqueing. And that is not a bad thing. Because it will pay off on the long term, Because you don’t have to spend exponential amounts of time fixing compounded technical debt.
💯
An AI CEO hyping up gen-AI…shocker!
The only thing that seems to be “happening” lately is how desperate these AI tech bros are sounding.
The most important question to ask a vibe coder is “can I test it out?” That’s where the rubber meets the road. This guy, Gary Tan, Karpathy or any of the other vibe coding influencers have yet to share a fully functional product. They’re all buggy prototypes.
Gary, you should take a look at FastRender (from Cursor) and claudes-c-compiler. Both are insane failed attempts at automating software engineering work on a serious projects.
Both companies shared the project with naive arrogance only to be humiliated by real engineers and experts reviewing the actual code.
Feels like the AI grifters are getting desperate.
I code using the latest llm technology and it's nowhere near the hype in that blog post. There's a reason LeCun doesn't believe llm's in their current state can reach AGI. And it's easy to see why once you start using them extensively. The marketing hype right now is far beyond their capability.
This whole event reminds me so much of the fast foods industry. Most people are now realizing that cheaply made and highly processed foods are unhealthy and a long way from what your caring parents might have fed you. Real food, prepared with knowledge and care is so much better for you. But, as we know, millions of people eat junk food which is heavily promoted. This also runs over the point that Gary and others have been making for so long: it takes real knowledge to judge the quality and usefulness of AI outputs. The real thinking is in the head of the user who judges the usefulness, errors in and quality of the output. The hype goes on.
What I fail to see is why the risk of compounding technical debt is hardly ever discussed? These systems generate code faster than any human can review. Having it reviewed by AI is fooling yourself. So we let AI generate code that contains errors we don't (take time to) understand, and then let AI fix them. And it works for a while (just like the next heroine hit works for while). And more errors creep in. And we let AI fix it, or at least the AI generated tests pass. In the meantime technical debt creeps (well that is not the right term in this case, rather 'storms') in without us noticing it. This keeps repeating itself until it (AI) cannot fix the errors anymore. Neither can the humans because they never internalized the architecture and code that were generated.
Obviously, and ironically, the better the systems get (as they undoubtedly do) the longer it takes to reach that point of no return.
By the way, I call this approach "Agent Addiction" because the process has a lot of analogies with what junkies go through.
My point, if we delegate these amounts of code generation to AI there is this very dangerous risk of technical debt explosion, much faster than technical debt happened to accrue in traditional software development. Are we going to wait for 6 to 18 month to find this out?
Then there is the other approach, people who are aware of the above risk, they will have to slow down, they will have to take time to review and internalize everything AI generates. On the short term this will (look to be) be orders of magnitude slower than the former approach. But possibly makes them come out as winners after 6 to 18 months. But there is another catch. The people needed to do this review and internalization are experienced software engineers that spent years and years of grinding real software experience. Grinding the skills needed so badly to responsibly make use of AI. But junior engineers come out of school and the majority of them lands in places where they are not offered (or not taking) the opportunity to learn those skills (because AI), which can only be learned by going through that grind.
Basically this means that the skill to responsibly apply AI is slowly waning and not being replenished. This makes our jobs (senior software engineers) in even more demand the coming years. I see a few possible mitigations
- We realize this is happening and will be actively encouraging juniors to keep going through that grind. This is the responsibility of software engineering companies.
- We wait for AGI to actually become as good as a senior developer and that specifically includes long running cognitive tasks (not those stupid smarter than PhD benchmark comparisons) which are absolutely crucial to build reliable systems and even more in order to fix the issues that result from compounding technical debt caused by Agent Addiction. Especially as long as the code is generated by LLM based technology.
Actually I am not really sure about the short term viability of the second option. We currently do not have an agreed upon definition of true AGI let alone a way of measuring whether we have achieved it, and even if we did, there are probably quite a number of systems for which we do not want to trust its internals to AGI only.
Final accountability should always lie with humans, so we need humans to fully understand the systems we build.
In fact it's not just technical debt build up that no-one's talking about - it's any kind of programming principles at all (Murphy's law, MVC, DRY etc.) People seem to have forgotten we have them. Or perhaps never knew about them in the first place. (A large percentage of the very vocal proponents of AI generated code probably aren't coders and never got that particular memo.)
The LLMs distressingly seem to have anti-principles. RYA - "repeat yourself, always!" MICS - "make it complicated, smartypants!" and other gems like, "if it doesn't pass the tests, meddle with the tests!" and "if a requirement in the spec is difficult, skip it and leave an obscure comment in the code, then say you did it anyway!"
Haven't you just described the majority of workers on freelance.com or upwork.com ?
AI would easily replace these, the quality wouldn't even suffer all that much.
Yes, sadly. This is the quality of work LLMs are trained on. Maybe they'd be less bad if they didn't need all the data to learn anything.
The question is, do we want to replace bad work with bad work, but cheaper?
Do we really need more bad software faster?
Well, to the extent we do, chatbots now exist.
That made me laugh!
You bring it like overstated sarcasm, but the sad truth is that this really happens regularly. I am really curious if we will see a technical debt epidemic months from now. Hopefully the majority of senior engineers have acknowledged these problems and are taking it slowly because of it. Nevertheless numbers show that AI usage by software engineers has grown substantially the last year. But it is hard to guess what the main usage patterns are.
There's much more realistic approach: just let all these things to go till the whole thing would collapse and things that are worth doing rewrite for would be rewritten from scratch.
The majority of code ever written is thrown away, after all… with LLMs we would just learn to accept that.
We already had this situation with freelance-contractors produced code which is often brought to the state where nothing can be fixed and you need to throw it away, AI would just mean that we would have more such throw-away unfixable code, it's not the end of the world.
I've become fluent in ChatGPTese recently - and this essay is definitely largely written by it.
What I think it shows is the way ChatGPT very effectively convinces people that it's turning them into superhumans.
This man - Matt Shumer, that is - plainly believes he's written a really very important essay, quite possibly for all humanity - when in fact it's terrible.
I've noticed this a lot recently - people becoming more dependent on AI to do things we really need to do ourselves as humans - for example relating to other people or articulating our opinions. and it happens very quickly -and these people have a completely impaired view of the quality of work they are producing; it profoundly impacts on their judgement.
This is the angle where things get more interesting. I mean, there's definitely going to be a sort of atrophy that happens for people that rely so much on automation and AI to live their lives for them. What's the wave that happens after the atrophy? Will the people who never got swept up in the hype start to appear like wizened masters simply because they can remember how to do things the hard way -- and because they still have the desire to engage meaningfully and creatively in their lives? Or something else?
Amazing! , I bet it also makes mounds mounds of julienne fries!
As I was reading the BS utterly perfect scenario, all I could think of is how most IT projects fail because of bad requirements. Which made me laugh. Because then the weakest link in this fantasy scenario would be whether he truly gave the AI perfect instructions. So I guess the next step is to build an AI that can come up with a perfect instructions so we don't need the human after all.And all you have to do is think at it. Because the moment you open your mouth, you'll be the source of error.
I don't know if I can stand so much perfection.
Put it to the test. Sit down with Claude and recover from a post compaction event while it's writing files to its own VM. While we all watch, live. If it fails - Deer Hunter Roulette - if it works but loses its mind - you lose a finger. If it does something bad but not fatal, the RFK Jr "treatment"
If people lived this every day they would not have these ludicrous greed-stoked ideas informed by nothing. Come down into the trenches, hype-meisters, where the real folks live and join the migrant techno-peasantry for a day.
This is the same viral poison, devoid of any substance, designed only to make people’s hearts jump. I wouldn’t be surprised if this was written by an LLM, iterated many times to regurgitate something meant to farm 'views' and 'engagement.'
It’s sad that a field like AI has become a subject of mockery and infantile content that brings nothing of value to humanity. We are wasting so many resources on spectacle.