I believe Gary Marcus is quite well known for terrible AI predictions. He's not in any way an expert in the field. Some of his predictions from 2022 [1]
> In 2029, AI will not be able to watch a movie and tell you accurately what is going on (what I called the comprehension challenge in The New Yorker, in 2014). Who are the characters? What are their conflicts and motivations? etc.
> In 2029, AI will not be able to read a novel and reliably answer questions about plot, character, conflicts, motivations, etc. Key will be going beyond the literal text, as Davis and I explain in Rebooting AI.
> In 2029, AI will not be able to work as a competent cook in an arbitrary kitchen (extending Steve Wozniak’s cup of coffee benchmark).
> In 2029, AI will not be able to reliably construct bug-free code of more than 10,000 lines from natural language specification or by interactions with a non-expert user. [Gluing together code from existing libraries doesn’t count.]
> In 2029, AI will not be able to take arbitrary proofs from the mathematical literature written in natural language and convert them into a symbolic form suitable for symbolic verification.
Many of these have already been achieved, and it's only early 2026.
Which ones are you claiming have already been achieved?
My understanding of the current scorecard is that he's still technically correct, though I agree with you there is velocity heading towards some of these things being proven wrong by 2029.
For example, in the recent thread about LLMs and solving an Erdos problem I remember reading in the comments that it was confirmed there were multiple LLMs involved as well as an expert mathematician who was deciding what context to shuttle between them and helping formulate things.
Similarly, I've not yet heard of any non-expert Software Engineers creating 10,000+ lines of non-glue code that is bug-free. Even expert Engineers at Cloud Flare failed to create a bug-free OAuth library with Claude at the helm because some things are just extremely difficult to create without bugs even with experts in the loop.
The bug-free code one feels unfalsifiable to me. How do you prove that 10,000 lines of code is bug-free, and then there's a million caveats about what a bug actually is and how we define one.
The second claim about novels seems obviously achieved to me. I just pasted a random obscure novel from project gutenberg into a file and asked claude questions about the characters, and then asked about the motivations of a random side-character. It gave a good answer, I'd recommend trying it yourself.
I agree with you but I'd point out that unless you've read the book it's difficult to know if the answer you got was accurate or it just kinda made it up. In my experience it makes stuff up.
Like, it behaves as if any answer is better than no answer.
I've done similar loose experiments using a 19th century French novel I have read (L'Agonie by Jean Lombard), and which is quite obscure. There is no English translation of the novel published.
Using Gemini Notebooks I've found it passable at summarizing chapters, listing characters, major theme and so on, and it can do this in English.
"What is the symbolism of the Black Stone?"
> In the sources, the Black Stone (referred to as the Pierre-Noire) serves as the central religious icon of the Syrian Sun cult brought to Rome by the Emperor Elagabalus. Its symbolism is multifaceted, ranging from fertility and virility to a mystical pursuit of universal unity.
> It represents the perennity of the Sun, which fertilizes the world, causing "germs to rise" and spreading them through the atmosphere. It is viewed as the "definitive form" of divinity, intended to absorb and replace the "transitory forms" of all other Roman, Greek, Egyptian, and Persian gods, including the Christian "Kreistos".
> Virility and Phallic Symbolism
>
> The sources explicitly characterize the Black Stone as a gigantic sacred phallus. It is described as:
• An icon of virility and the "organ of generation" • A "unisexual icon" that materializes the generative force of nature • A representation of "virility in activity," which is why it is often paraded and elevated in an "orgasmic" or "colossal adoration"
> The Androgyne and Universal Unity - Beyond simple fertility, the philosopher Atillius explains a more complex, "mad" metaphysical project associated with the stone. It symbolizes "Life One" (Vie Une) and the return to a unisexual state of perfection. • The Androgyne: Atillius believes that by pursuing "the male sex by the male sex," the cult "inutilizes" the female sex to eventually create the Androgyne—a self-sufficient being containing both sexes • Unity: The stone signifies the fusion of all generative forces into a single Unity, reversing the "separation of the sexes" which is viewed as a state of unhappiness and impotence. • Marriage of Moon and Sun: The ritual marriage of the goddess Astaroth (representing the Moon and the female principle) to the Black Stone (representing the Sun and the male principle) symbolizes the merging of the Orient and Occident into this unified life principle
>
> Destruction of the Symbol - The Black Stone ultimately becomes a symbol of Oriental pollution and decadence to the Roman populace. During the final rebellion against Elagabalus, the stone is torn from its temple on the Palatine, defiled with filth, and broken into pieces to ensure that its "signification of Life" would never again dominate Rome.
This is all accurate to the book, even teasing out a couple themes that were only subconsciously present to me.
The NotebookLM version gives citations with links to the original text to support all these assertions, which largely are coherent with that purpose.
The input is raw images of a book scan! Imperfect as it is it still blows my mind. Not that long ago any kind of semantic search or analysis was a very hard AI problem.
Not quite the same analysis. The human is better, no surprise. But the NotebookLM output links back to the original book in a very useful way. If you think about it as fuzzy semantic search it's amazing. If you want an essay or even just creativity, yes it's lacking.
It doesn't have to be the same analysis to put it in a partially overlapping vector space. Not saying it wasn't a useful perspective shuffling in the vector space, but it definitely wasn't original.
LLMs haven't solved any of the 2029 predictions as they were posited. But I expect some will be reached by 2029. The AI hype acts like all this is easy. Not by 2029 doesn't mean impossible or even most of the way there.
LLMs will never achieve anything as long as any victory can be hand waved away with "in the training set". Somehow these models have condensed the entire internet down to a few TB's, yet people aren't backing up their terabytes of personal data down to a couple MB using this same tech...wonder why
I strongly disagree. I’ve yet to find an AI that can reliably summarise emails, let alone understand nuance or sarcasm. And I just asked ChatGPT 5.2 to describe an Instagram image. It didn’t even get the easily OCR-able text correct. Plus it completely failed to mention anything sports or stadium related. But it was looking at a cliche baseball photo taken by an fan inside the stadium.
I have had ChatGPT read text in an image, give me a 100% accurate result, and then claim not to have the ability and to have guessed the previous result when I ask it to do it again.
1) Is it actually watching a movie frame by frame or just searching about it and then giving you the answer?
2) Again can it handle very long novels, context windows are limited and it can easily miss something. Where is the proof for this?
4 is probably solved
4) This is more on predictor because this is easy to game. you can create some gibberish code with LLM today that is 10k lines long without issues. Even a non-technical user can do
I think all of those are terrible indicators, 1 and 2 for example only measure how well LLMs can handle long context sizes.
If a movie or novel is famous the training data is already full of commentary and interpretations of them.
If its something not in the training data, well I don't know many movies or books that use only motives that no other piece of content before them used, so interpreting based on what is similar in the training data still produces good results.
EDIT:
With 1 I meant using a transcript of the Audio Description of the movie. If he really meant watch a movie I'd say thats even sillier because well of course we could get another Agent to first generate the Audio Description, which definitely is possible currently.
Just yesterday I saw an article about a police station's AI body cam summarizer mistakenly claim that a police officer turned into a frog during a call. What actually happened was that the cartoon "princess and the frog" was playing in the background.
Sure, another model might have gotten it right, but I think the prediction was made less in the sense of "this will happen at least once" and more of "this will not be an uncommon capability".
When the quality is this low (or variable depending on model) I'm not too sure I'd qualify it as a larger issue than mere context size.
My point was not that those video to text models are good like they are used for example in that case, but more generally I was referring to that list of indicators. Like surely when analysing a movie it is alright if some things are misunderstood by it, especially as the amount of misunderstanding can be decreased a lot. That AI body camera surely is optimized on speed and inference cost. but if you give an agent 10 1s images along with the transcript of that period and the full prior transcript, and give it reasoning capabilities, it would take almost endlessy for that movie to process but the result surely will be much better than the body cameras. After all the indicator talks about "AI" in general so judge a model not optimized for capability but something else to measure on that indicator
> In 2029, AI will not be able to read a novel and reliably answer questions about plot, character, conflicts, motivations, etc. Key will be going beyond the literal text, as Davis and I explain in Rebooting AI.
Can AI actually do this? This looks like a nice benchmark for complex language processing, since a complete novel takes up a whole lot of context (consider War and Peace or The Count of Monte Cristo). Of course the movie variety is even more challenging since it involves especially complex multi-modal input. You could easily extend it to making sense of a whole TV series.
Yes. I am a novelist and I noticed a step change in what was possible here around Claude Sonnet 3.7 in terms of being able to analyze my own unpublished work for theme, implicit motivations, subtext, etc -- without having any pre-digested analysis of the work in its training data.
Novel is different from a codebase. In code you can have a relationship between files and most files can be ignored depending on what you're doing. But for a novel, its a sequential thing, in most cases A leads to B and B leads to C and so on.
> Re: movies. Get YouTube premium and ask YouTube to summarize a 2hr video for you.
This is different from watching a movie. Can it tell what suit actor was wearing? Can it tell what the actor's face looked like? Summarising and watching are too different things.
Yes, it is possible to do those things and there are benchmarks for testing multimodal models on their ability to do so. Context length is the major limitation but longer videos can be processed in small chunks whose descriptions can be composed into larger scenes.
Gary Marcus' proposal was about a novel, and not a codebase. I think GP's point is that motivations require analysis outside the context window, which LLMs are essentially incapable of doing.
No human reads a novel and evaluates it as a whole. It's a story and the readers perception changes over the course of reading the book. Current AI can certainly do that.
>Can AI actually do this? This looks like a nice benchmark for complex language processing, since a complete novel takes up a whole lot of context (consider War and Peace or The Count of Monte Cristo)
Yes, you just break the book down by chapters or whatever conveniently fits in the context window to produce summaries such that all of the chapter summaries can fit in one context window.
You could also do something with a multi-pass strategy where you come up with a collection of ideas on the first pass and then look back with search to refine and prove/disprove them.
Of course for novels which existed before the time of training an LLM will already contain trained information about so having it "read" classic works like The Count of Monte Cristo and answer questions about it would be a bit of an unfair pass of the test because models will be expected to have been trained on large volumes of existing text analysis on that book.
>reliably answer questions about plot, character, conflicts, motivations
LLMs can already do this automatically with my code in a sizable project (you know what I mean), it seems pretty simple to get them to do it with a book.
> Many of these have already been achieved, and it's only early 2026.
I'm quite sure people who made those (now laughable) predictions will tell you none of these has been achieved, because AI isn't doing this "reliably" or "bug-free."
Defending your predictions is like running an insurance company. You always win.
In my opinion, contrary to other comments here I think AI can do all of the above already except being a kitchen cook.
Just earlier today I asked it to give me a summary of a show I was watching until a particular episode in a particular season without spoiling the rest of it and it did a great job.
Which ones of those have been achieved in your opinion?
I think the arbitrary proofs from mathematical literature is probably the most solved one. Research into IMO problems, and Lean formalization work have been pretty successful.
Then, probably reading a novel and answering questions is the next most successful.
Reliably constructing 10k bug free lines is probably the least successful. AI tends to produce more bugs than human programmers and I have yet to meet a programmer who can reliably produce less than 1 bug per 10k lines.
Formalizing an arbitrary proof is incredibly hard. For one thing, you need to make sure that you've got at least a correct formal statement for all the prereqs you're relying on, or the whole thing becomes pointless. Many areas of math ouside of the very "cleanest" fields (meaning e.g. algebra, logic, combinatorics etc.) have not seen much success in formalizing existing theory developments.
I have seen many people try to use Claude Code and get LOTS of bugs. Show me any > 10k project you have made with it and I will put the effort in to find one bug free of charge.
Besides being a cook which is more of a robotics problem all of the rest are accomplished to the point of being arguable about how reliably LLMs can perform these tasks, the arguments being between the enthusiast and naysayer camps.
The keyword being "reliably" and what your threshold is for that. And what "bug free" means. Groups of expert humans struggle to write 10k lines of "bug free" code in the absolutist sense of perfection, even code with formal proofs can have "bugs" if you consider the specification not matching the actual needs of reality.
All but the robotics one are demonstrable in 2026 at least.
I'm pretty sure it can do all of those except for the one which requires a physical body (in the kitchen) and the one that humans can't do reliably either (construct 10000 loc bug-free).
And why not? Is there any reason for this comment to not appear?
If Bill Gates made a predication about computing, no matter what the predication says, you can bet that 640K memory quote would be mentioned in the comment section (even he didn't actually say that).
I think it’s for good reason. I’m a bit at a loss as to why every time this guy rages into the ether of his blog it’s considered newsworthy. Celebrity driven tech news is just so tiresome. Marcus was surpassed by others in the field and now he’s basically a professional heckler on a university payroll. I wish people could just be happy for the success of others instead of fuming about how so and so is a billionaire and they are not.
> Meanwhile, my cofounder is rewriting code we spent millions of salary on in the past by himself in a few weeks.
Code is not an asset it's a liability, and code that no one has reviewed is even more of a liability.
However, in the end, execution is all that matters so if you and your cofounder are able to execute successfully with mountains of generated code then it doesn't matter what assets and liabilities you hold in the short term.
The long term is a lot harder to predict in any case.
> Code is not an asset it's a liability, and code that no one has reviewed is even more of a liability.
Code that solves problems and makes you money is by definition an asset. Whether or not the code in question does those things remains to be seen, but code is not strictly a liability or else no one would write it.
This discussion and distinction used to be well known, but I'm happy to help some people become "one of today's lucky 10,000" as quoted from https://xkcd.com/1053/ because it is indeed much more interesting than the alternative approach.
Just this week, sun-tue. I added a fully functional subscription model to an existing platform, build out a bulk async elasticjs indexing for a huge database and migrated a very large Wordpress website to NextJS. 2.5 days, would have cost me at least a month 2 years ago.
AI is helping me solve all the issues that using AI has caused.
Wordpress has a pretty good export and Markdown is widely supported. If you estimate 1 month of work to get that into NextJS, then maybe the latter is not a suitable choice.
It's not directly comparable. The first time writing the code is always the hardest because you might have to figure out the requirements along the way. When you have the initial system running for a while, doing a second one is easier because all the requirements kinks are figured out.
By the way, why does your co-founder have to do the rewrite at all?
You can compare it - just factor that in. And compare writing it with AI vs. writing it without AI.
We have no clue the scope of the rewrite but for anything non-trivial, 2 weeks just isn't going to be possible without AI. To the point of you probably not doing it at all.
I have no idea why they are rewriting the code. That's another matter.
I find the opposite to be true. Once you know the problem you’re trying to solve (which admittedly can be the biggest lift), writing the fist cut of the code is fun, and you can design the system and set precedent however you want. Once it’s in the wild, you have to work within the consequences of your initial decisions, including bad ones.
lol same. I just wrote a bunch of diagrams with mermaid that would legit take me a week, also did a mock of an UI for a frontend engineer that would take me another week to do .. or some designers. All of that in between meetings...
Waiting for it to actually go well to see what else I can do !
Oh I've thought this for years. As an L7, basically my primary role is to serve as someone to bounce ideas off of, and to make recommendations based on experience. A chatbot, with its virtually infinite supply of experience, could ostensibly replace my role way sooner than it could a solid junior/mid-level coder. The main thing it needs is a consistent vision and direction that aligns with the needs of nearby teams, which frankly sounds not all that hard to write in code (I've been considering doing this).
Probably the biggest gap would be the ability to ignite, drive, and launch new initiatives. How does an AI agent "lead" an engineering team? That's not something you can code up in an agent runtime. It'd require a whole culture change that I have a hard time seeing in reality. But of course if there comes a point where AI takes all the junior and mid-level coding jobs, then at that point there's no culture to change, so staff/principal jobs would be just as at risk.
I suspect he means as a trillion dollar corporation led endeavor.
I trained a small neural net on pics of a cat I had in the 00s (RIP George, you were a good cat).
Mounted a webcam I had gotten for free from somewhere, above the cat door, in the exterior of the house.
If the neural net recognized my cat it switched off an electromagnetic holding the pet door locked. Worked perfectly until I moved out of the rental.
Neural nets are, end of the day, pretty cool. It's the data center business that's the problem. Just more landlords, wannabe oligarchs, claiming ownership over anything they can get the politicians to give them.
The problem is... you're going to deprive yourself of the talent chain in the long run, and so is everyone else who is switching over to AI, both generative like ChatGPT and transformative like the various translation, speech recognition/transcription or data wrangling models.
For now, it works out for companies - but forward to, say, ten years in the future. There won't be new intermediates or seniors any more to replace the ones that age out or quit the industry entirely in frustration of them not being there for actual creativity but to clean up AI slop, simply because there won't have been a pipeline of trainees and juniors for a decade.
But by the time that plus the demographic collapse shows its effects, the people who currently call the shots will be in pension, having long since made their money. And my generation will be left with collapse everywhere and find ways to somehow keep stuff running.
Hell, it's already bad to get qualified human support these days. Large corporations effectively rule with impunity, with the only recourse consumers have being to either shell out immense sums of money for lawyers and court fees or turning to consumer protection/regulatory authorities that are being gutted as we speak both in money and legal protections, or being swamped with AI slop like "legal assistance" AI hallucinating case law.
> Meanwhile, my cofounder is rewriting code we spent millions of salary on in the past by himself in a few weeks.
This is one of those statements that would horrify any halfway competent engineer. A cowboy coder going in, seeing a bunch of code and going 'I should rewrite this' is one of the biggest liabilities to any stable system.
As someone who is more involved in shaping the product direction rather than engineering what composes the product - I will readily admit many product people are utterly, utterly clueless.
Most people have no clue the craftsmanship, work etc it takes to create a great product. LLMs are not going to change this, in fact they serve as a distraction.
I’m not a SWE so I gain nothing by being bearish on the contributions of LLMs to the real economy ;)
I find it a bit odd that people are acting like this stuff is an abject failure because it's not perfect yet.
Generative AI, as we know it, has only existed ~5-6 years, and it has improved substantially, and is likely to keep improving.
Yes, people have probably been deploying it in spots where it's not quite ready but it's myopic to act like it's "not going all that well" when it's pretty clear that it actually is going pretty well, just that we need to work out the kinks. New technology is always buggy for awhile, and eventually it becomes boring.
> Generative AI, as we know it, has only existed ~5-6 years, and it has improved substantially, and is likely to keep improving.
Every 2/3 months we're hearing there's a new model that just blows the last one out of the water for coding. Meanwhile, here I am with Opus and Sonnet for $20/mo and it's regularly failing at basic tasks, antigravity getting stuck in loops and burning credits. We're talking "copy basic examples and don't hallucinate APIs" here, not deep complicated system design topics.
It can one shot a web frontend, just like v0 could in 2023. But that's still about all I've seen it work on.
You’re doing exactly the thing that the parent commenter pointed out: Complaining that they’re not perfect yet as if that’s damning evidence of failure.
We all know LLMs get stuck. We know they hallucinate. We know they get things wrong. We know they get stuck in loops.
There are two types of people: The first group learns to work within these limits and adapt to using them where they’re helpful while writing the code when they’re not.
The second group gets frustrated every time it doesn’t one-shot their prompt and declares it all a big farce. Meanwhile the rest of us are out here having fun with these tools, however limited they are.
> The whole discourse around LLMs is so utterly exhausting. If I say I don't like them for almost any reason, I'm a luddite. If I complain about their shortcomings, I'm just using it wrong. If I try and use it the "right" way and it still gets extremely basic things wrong, then my expectations are too high.
As I’ve said, I use LLMs, and I use tools that are assisted by LLMs. They help. But they don’t work anywhere near as reliably as people talk about them working. And that hasn’t changed in the 18 months since I first promoted v0 to make me a website.
If you hired a human, it will cost you thousands a week. Humans will also fail at basic tasks, get stuck in useless loops, and you still have to pay them for all that time.
For that matter, even if I'm not hiring anyone, I will still get stuck on projects and burn through the finite number of hours I have on this planet trying to figure stuff out and being wrong for a lot of it.
It's not perfect yet, but these coding models, in my mind, have gotten pretty good if you're specific about the requirements, and even if it misfires fairly often, they can still be useful, even if they're not perfect.
I've made this analogy before, but to me they're like really eager-to-please interns; not necessarily perfect, and there's even a fairly high risk you'll have to redo a lot of their work, but they can still be useful.
I am an AI-skeptic but I would agree this looks impressive from certain angles, especially if you're an early startup (maybe) or you are very high up the chain and just want to focus on cutting costs. On the other hand, if you are about to be unemployed, this is less impressive. Can it replace a human? I would say no its still long way to go, but a good salesman can convince executives that it does and thats all that matters.
> On the other hand, if you are about to be unemployed, this is less impressive
> salesman can convince executives that it does
I tend to think that reality will temper this trend as the results develop. Replacing 10 engineers with one engineer using Cursor will result in a vast velocity hit. Replacing 5 engineers with 5 "agents" assigned to autonomously implement features will result in a mess eventually. (With current technology -- I have no idea what even 2027 AI will do). At that point those unemployed engineers will find their phones ringing off the hook to come and clean up the mess.
Not that unlike what happens in many situations where they fire teams and offshore the whole thing to a team of average developers 180 degrees of longitude away who don't have any domain knowledge of the business or connections to the stakeholders. The pendulum swings back in the other direction.
I just think Jevins paradox [1]/Gustafson's Law [2] kind of applies here.
Maybe I shouldn't have used the word "replaced", as I don't really think it's actually going to "replace" people long term. I think it's likely to just lead to higher output as these get better and better .
Not you, but the word replaced is the being used all the time. Even senior engineers are saying they are using it as a junior engineers while we can easily hire junior engineers (but Execs don't want to). Jevon's paradox wont work in Software because user's wallets and time is limited, and if software becomes too easy to build, it becomes harder to sell. Normal people can have 5 subscriptions, may be 10, but they wont be going to 50 or 100. I would say we would have already exhausted users already, with all the bad practices.
You’ve missed my point here - I agree that gen AI has changed everything and is useful, _but_ I disagree that it’s improved substantially - which is what the comment I replied to claimed.
Anecdotally I’ve seen no difference in model changes in the last year, but going from LLM to Claude code (where we told the LLMs they can use tools on our machines) was a game changer. The improvement there was the agent loop and the support for tools.
In 2023 I asked v0.dev to one shot me a website for a business I was working on and it did it in about 3 minutes. I feel like we’re still stuck there with the models.
In my experience it has gotten considerably better. When I get it to generate C, it often gets the pointer logic correct, which wasn't the case three years ago. Three years ago, ChatGPT would struggle with even fairly straightforward LaTeX, but now I can pretty easily get it to generate pretty elaborate LaTeX and I have even had good success generating LuaTeX. I've been able to fairly successfully have it generate TLA+ spec from existing code now, which didn't work even a year ago when I tried it.
Of course, sample size of one, so if you haven't gotten those results then fair enough, but I've at least observed it getting a lot better.
>Every 2/3 months we're hearing there's a new model that just blows the last one out of the water for coding
I haven't heard that at all. I hear about models that come out and are a bit better. And other people saying they suck.
>Meanwhile, here I am with Opus and Sonnet for $20/mo and it's regularly failing at basic tasks, antigravity getting stuck in loops and burning credits.
Is it bringing you any value? I find it speeds things up a LOT.
> If your metric is an LLM that can copy/paste without alterations, and never hallucinate APIs, then yeah, you'll always be disappointed with them.
I struggle to take comments like this seriously - yes, it is very reasonable to expect these magical tools to copy and paste something without alterations. How on earth is that an unreasonable ask?
The whole discourse around LLMs is so utterly exhausting. If I say I don't like them for almost any reason, I'm a luddite. If I complain about their shortcomings, I'm just using it wrong. If I try and use it the "right" way and it still gets extremely basic things wrong, then my expectations are too high.
It seems like just such a weird and rigid way to evaluate it? I am a somewhat reasonable human coder, but I can't copy and paste a bunch of code without alterations from memory either. Can someone still find a use for me?
I think what they're best at right now is the initial scaffolding work of projects. A lot of the annoying bootstrap shit that I hate doing is actually generally handled really well by Codex.
I agree that there's definitely some overhype to them right now. At least for the stuff I've done they have gotten considerably better though, to a point where the code it generates is often usable, if sub-optimal.
For example, about three years ago, I was trying to get ChatGPT to write me a C program to do a fairly basic ZeroMQ program. It generated something that looked correct, but it would crash pretty much immediately, because it kept trying to use a pointer after free.
I tried the same thing again with Codex about a week ago, and it worked out of the box, and I was even able to get it to do more stuff.
I think it USED to be true that you couldn't really use an LLM on a large, existing codebase. Our codebase is about 2 million LOC, and a year ago you couldn't use an LLM on it for anything but occasional small tasks. Now, probably 90% of the code I commit each week was written by Claude (and reviewed by me and other humans - and also by Copilot and ZeroPath).
For a long time, I've wanted to write a blog post on why programmers don't understand the utility of LLMs[1], whereas non-programmers easily see it. But I struggle to articulate it well.
The gist is this: Programmers view computers as deterministic. They can't tolerate a tool that behaves differently from run to run. They have a very binary view of the world: If it can't satisfy this "basic" requirement, it's crap.
Programmers have made their career (and possibly life) being experts at solving problems that greatly benefit from determinism. A problem that doesn't - well either that needs to be solved by sophisticated machine learning, or by a human. They're trained on essentially ignoring those problems - it's not their expertise.
And so they get really thrown off when people use computers in a nondeterministic way to solve a deterministic problem.
For everyone else, the world, and its solutions, are mostly non-deterministic. When they solve a problem, or when they pay people to solve a problem, the guarantees are much lower. They don't expect perfection every time.
When a normal human asks a programmer to make a change, they understand that communication is lossy, and even if it isn't, programmers make mistakes.
Using a tool like an LLM is like any other tool. Or like asking any other human to do something.
For programmers, it's a cardinal sin if the tool is unpredictable. So they dismiss it. For everyone else, it's just another tool. They embrace it.
[1] This, of course, is changing as they become better at coding.
I’m perfectly happy for my tooling to not be deterministic. I’m not happy for it to make up solutions that don’t exist, and get stuck in loops because of that.
I use LLMs, I code with a mix of antigravity and Claude code depending on the task, but I feel like I’m living in a different reality when the code I get out of these tools _regularly just doesn’t work, at all_. And to the parents point, I’m doing something wrong for noticing that?
If it were terrible, you wouldn't use them, right? Isn't the fact that you continue to use AI coding tools a sign that you find them a net positive? Or is it being imposed on you?
> And to the parents point, I’m doing something wrong for noticing that?
There's nothing wrong pointing out your experience. What the OP was implying was he expects them to be able to copy/paste reliably almost 100% of the time, and not hallucinate. I was merely pointing out that he'll never get that with LLMs, and that their inability to do so isn't a barrier to getting productive use out of them.
Its strong enough to replace humans at their jobs and weak enough that it cant do basic things. Its a paradox. Just learn to be productive with them. Pay $200/month and work around with its little quirks. /s
There’s a subtle point a moment when you HAVE to take the driver wheel from the AI.
All issues I see are from people insisting to use far beyond the point it stops being useful.
It is a helper, a partner, it is still not ready go the last mile
It's funny how many people don't get that. It's like adding a pretty great senior or staff level engineer to sit on-call next to every developer and assist them, for basically free (I've never used any of the expensive stuff yet. Just things like Copilot, Grok Code in JetBrains, just asking Gemini to write bits of code for me).
If you hired a staff engineer to sit next to me, and I just had him/her write 100% of the code and never tried to understand it, that would be an unwise decision on my part and I'd have little room to complain about the times he made mistakes.
> The whole discourse around LLMs is so utterly exhausting. If I say I don't like them for almost any reason, I'm a luddite. If I complain about their shortcomings, I'm just using it wrong. If I try and use it the "right" way and it still gets extremely basic things wrong, then my expectations are too high.
I’m perfectly happy to write code, to use these tools. I do use them, and sometimes they work (well). Other times they have catastrophic failures. But apparently it’s my failure for not understanding the tool or expecting too much of the tool, while others are screaming from the rooftops about how this new model changes everything (which happens every 3 months at this point)
There's no silver bullet. I’m not a researcher, but I’ve done my best to understand how these systems work—through books, video courses, and even taking underpaid hourly work at a company that creates datasets for RLHF. I spent my days fixing bugs step-by-step, writing notes like, “Hmm… this version of the library doesn’t support protocol Y version 4423123423. We need to update it, then refactor the code so we instantiate ‘blah’ and pass it to ‘foo’ before we can connect.”
That experience gave me a deep appreciation for how incredible LLMs are and the amazing software they can power—but it also completely demystified them. So by all means, let’s use them. But let’s also understand there are no miracles here. Go back to Shannon’s papers from the ’60s, and you'll understand that what seems to you like "emerging behaviors" are quite explainable from an information theory background. Learn how these models are built. Keep up with the latests research papers. If you do, you’ll recognize their limitations before those limitations catch you by surprise.
There is no silver bullet. And if you think you’ve found one, you’re in for a world of pain. Worse still, you’ll never realize the full potential of these tools, because you won’t understand their constraints, their limits, or their pitfalls.
>Generative AI, as we know it, has only existed ~5-6 years
Probably less than that, practically speaking. ChatGPT's initial release date was November 2022. It's closer to 3 years, in terms of any significant amount of people using them.
We implement pretty cool workflows at work using "GenAI" and the users of our software are really appreciative. It's like saying a hammer sucks because it breaks most things you hit with it.
I think that even if it never improves, its current state is already pretty useful. I do think it's going to improve though I don't think AGI is going to happen any time soon.
I have no idea what this is called, but it feels like a lot of people assume that progress will continue at a linear pace for forever for things, when I think that generally progress is closer to a "staircase" shape. A new invention or discovery will lead to a lot of really cool new inventions and discoveries in a very short period of time, eventually people will exhaust the low-to-middle-hanging fruit, and progress kind of levels out.
I suspect it will be the same way with AI; I don't now if we've reached the top of our current plateau, but if not I think we're getting fairly close.
I'm not trying to be pedantic, but how did you arrive at 'keep improving' as a conclusion? Nobody is really sure how this stuff actually works. That's why AI safety was such a big deal a few years ago.
Totally reasonable question, and I only am making an assumption based on observed progress. AI generated code, at least in my personal experience, has gotten a lot better, and while I don't think that will go to infinity, I do think that there's still more room for improvement that could happen.
I will acknowledge that I don't have any evidence of this claim, so maybe the word "likely" was unwise, as that suggests probability. Feel free to replace "is "likely to" with "it feels like it will".
I made a joke once after the first time I watched one of those Apple announcement shows in 2018, where I said "it's kind of sad, because there won't be any problems for us to solve because the iPhone XS Max is going to solve all of them".
The US economy is pretty much a big vibes-based Ponzi scheme now, so I don't think we can single-out AI, I think we have to blame the fact that the CEOs running these things face no negative consequences for lying or embellishing and they do get rewarded for it because it will often bump the stock price.
Is Tesla really worth more than every other car company combined in any kind of objective sense? I don't think so, I think people really like it when Elon lies to them about stuff that will come out "next year", and they feel no need to punish him economically.
Exactly: the technology is useful but because the executive class is hyping it as close to AGI because their buddies are slavering for layoffs. If that “when do you get fired?” tone wasn’t behind the conversation, I think a lot of people would be interested in applying LLMs to the smaller subset of things they actually perform well at.
Maybe CEOs should face consequences for going on the stage and outwardly lying. Instead they're rewarded by a bump in stock price because people appear to have amnesia.
This feels like a pretty low effort post that plays heavily to superficial reader's cognitive biases.
I work commercializing AI in some very specific use cases where it extremely valuable. Where people are being lead astray is layering generalizations: general use cases (copilots) deployed across general populations and generally not doing very well. But that's PMF stuff, not a failure of the underlying tech.
I think both sides of this debate are conflating the tech and the market. First of all, there were forms of "AI" before modern Gen AI (machine learning, NLP, computer vision, predictive algorithms, etc) that were and are very valuable for specific use cases. Not much has changed there AFAICT, so it's fair that the broader conversation about Gen AI is focused on general use cases deployed across general populations. After all, Microsoft thinks it's a copilot company, so it's fair to talk about how copilots are doing.
On the pro-AI side, people are conflating technology success with product success. Look at crypto -- the technology supports decentralization, anonymity, and use as a currency; but in the marketplace it is centralized, subject to KYC, and used for speculation instead of transactions. The potential of the tech does not always align with the way the world decides to use it.
On the other side of the aisle, people are conflating the problematic socio-economics of AI with the state of the technology. I think you're correct to call it a failure of PMF, and that's a problem worth writing articles about. It just shouldn't be so hard to talk about the success of the technology and its failure in the marketplace in the same breath.
A year ago I would have agreed wholeheartedly and I was a self confessed skeptic.
Then Gemini got good (around 2.5?), like I-turned-my-head good. I started to use it every week-ish, not to write code. But more like a tool (as you would a calculator).
More recently Opus 4.5 was released and now I'm using it every day to assist in code. It is regularly helping me take tasks that would have taken 6-12 hours down to 15-30 minutes with some minor prompting and hand holding.
I've not yet reached the point where I feel letting is loose and do the entire PR for me. But it's getting there.
I think that's the key. Healthy skepticism is always appropriate. It's the outright cynicism that gets me. "AI will never be able to [...]", when I've been sitting here at work doing 2/3rds of those supposedly impossible things. Flawlessly? No, of course not! But I don't do those things flawlessly on the first pass, either.
Skepticism is good. I have no time or patience for cynics who dismiss the whole technology as impossible.
I would strongly recommend this podcast episode with Andrej Karpathy. I will poorly summarize it by saying his main point is that AI will spread like any other technology. It’s not going to be a sudden flash and everything is done by AI. It will be a slow rollout where each year it automates more and more manual work, until one day we realize it’s everywhere and has become indispensable.
It sounds like what you are seeing lines up with his predictions. Each model generation is able to take on a little more of the responsibilities of a software engineer, but it’s not as if we suddenly don’t need the engineer anymore.
Though I think it's a very steep sigmoid that we're still far on the bottom half of.
For math it just did its first "almost independent" Erdos problem. In a couple months it'll probably do another, then maybe one each month for a while, then one morning we'll wake up and find whoom it solved 20 overnight and is spitting them out by the hour.
For software it's been "curiosity ... curiosity ... curiosity ... occasionally useful assistant ... slightly more capable assistant" up to now, and it'll probably continue like that for a while. The inflection point will be when OpenAI/Anthropic/Google releases an e2e platform meant to be driven primarily by the product team, with engineering just being co-drivers. It probably starts out buggy and needing a lot of hand-holding (and grumbling) from engineering, but slowly but surely becomes more independently capable. Then at some point, product will become more confident in that platform than their own engineering team, and begin pushing out features based on that alone. Once that process starts (probably first at OpenAI/Anthropic/Google themselves, but spreading like wildfire across the industry), then it's just a matter of time until leadership declares that all feature development goes through that platform, and retains only as many engineers as is required to support the platform itself.
Gary Marcus (probably): "Hey this LLM isn't smarter than Einstein yet, it's not going all that well"
The goalposts keep getting pushed further and further every month. How many math and coding Olympiads and other benchmarks will LLMs need to dominate before people will actually admit that in some domains it's really quite good.
Sure, if you're a Nobel prize winner or PhD then LLMs aren't as good as you yet, but for 99% of the people in the world, LLMs are better than you at Math, Science, Coding, and every language probably except your native language, and it's probably better at you at that too...
Ignoring the actual poor quality of this write-up, I think we don't know how well GenAI is going to be honest. I feel we've not been able to properly measure or assess it's actual impact yet.
Even as I use it, and I use it everyday, I can't really assess its true impact. Am I more productive or less overall? I'm not too sure. Do I do higher quality work or lower quality work overall? I'm not too sure.
All I know, it's pretty cool, and using it is super easy. I probably use it too much, in a way, that it actually slows things down sometimes, when I use it for trivial things for example.
At least when it comes to productivity/quality I feel we don't really know yet.
But there are definite cool use-cases for it, I mean, I can edit photos/videos in ways I simply could not before, or generate a logo for a birthday party, I couldn't do that before. I can make a tune that I like, even if it's not the best song in the world, but it can have the lyrics I want. I can have it extract whatever from a PDF. I can have it tell me what to watch out for in a gigantic lease agreement I would not have bothered reading otherwise.
I can have it fix my tests, or write my tests, not sure if it saves me time, but I hate doing that, so it definitely makes it more fun and I can kind of just watch videos at the same time, what I couldn't before. Coding quality of life improvements are there too, I want to generate a sample JSON out of a JSONSchema, and so on. If I want, I can write the a method using English prompts instead of the code itself, might not truly be faster or not, not sure, but sometimes it's less mentally taxing, depending on my mood, it can be more fun or less fun, etc.
All those are pretty awesome wins and a sign that for sure those things will remain and I will happily pay for them. So maybe it depends on what you expected.
Meanwhile $employer is continuing to migrate individual tasks to in-house AI tooling, and has licensed an off-the-shelf coding agent for all of us developers to put in our IDEs.
The irony of a five sentence article making giant claims isn't lost on me. Don't get me wrong: I'm amenable to the idea; but, y'know, my kids wrote longer essays in 4th grade.
What a joke this guy is. I can sit down and crank out a real, complex feature in a couple hours that would have previously taken days and ship it to the users of our AI platform who can then respond to RFQs in minutes where they would have previously spent hours matching descriptions to part numbers manually.
...and yet we still see these articles claiming LLMs are dying/overhyped/major issues/whatever.
Cool man, I'll just be over here building my AI based business with AI and solving real problems in the very real manufacturing sector.
Guessing this isn’t going to be popular here, but he’s right. AI has some use cases, but isn’t the world-changing paradigm shift it’s marketed as. It’s becoming clear the tech is ultimately just a tool, not a precursor to AGI.
The comment wasn't referencing LLMs, but generative AI.
Even then, given the deep impact of LLMs and how many people are using them already, it's a stretch to say LLMs will have no effect on the development of AGI.
I think it's pretty obvious that AGI requires something more than LLMs, but I think it's equally obvious LLMs will have been involved in its development somewhere, even if just a stepping stone. So, a "precursor".
I keep reading comments that claim GenAI's positive traits, but this usually amounts to some toy PoC that very eerily mirrors work found in code bootcamps. You want an app that has logins and comments and upvotes? GenAI is going to look amazing setting up a non-relational db to your node backend.
I’m starting to think this take is legitimately insane.
As said in the article, a conservative estimate is that Gen AI can currently do 2.5% of all jobs in the entire economy. A technology that is really only a couple of years old. This is supposed to be _disappointing_? That’s millions of jobs _today_, in a totally nascent form.
I mean I understand skepticism, I’m not exactly in love with AI myself, but the world has literally been transformed.
It's literally just four screenshots paired with this sentence.
> Trying to orient our economy and geopolitical policy around such shoddy technology — particularly on the unproven hopes that it will dramatically improve– is a mistake.
The screenshots are screenshots of real articles. The sentence is shorter than a typical prompt.
He's such a joke that even LLMs make fun of him. The Gemini-generated Hacker News frontpage for December 9 2035 contains an article by Gary Marcus: "AI progress is stalling": https://dosaygo-studio.github.io/hn-front-page-2035/news
It's going well for coding. I just knocked out a mapping project that would have been a week+ of work (with docs and stackoverflow opened in the background) in a few hours.
And yes, I do understand the code and what is happening and did have to make a couple of adjustments manually.
I don't know that reducing coding work justifies the current valuations, but I wouldn't say it's "not going all that well".
Download models you can find now and forever. The guardrails will only get worse, or models banned entirely. Whether it's because of "hurts people's health" or some other moral panic, it will kill this tech off.
gpt-oss isn't bad, but even models you cannot run are worth getting since you may be able to run them in the future.
I'm hedging against models being so nerfed they are useless. (This is unlikely, but drives are cheap and data is expensive.)
I just used ChatGPT to diagnose a very serious but ultimately not-dangerous health situation last week and it was perfect. It literally guided me perfectly without making me panic and helped me understand what was going on.
We use ChatGPT at work to do things that we have literally laid people off for, because we don't need them anymore. This included fixing bugs at a level that is at least E5/senior software engineer. Sometimes it does something really bad but it definitely saves times and helps avoid adding headcount.
Generative AI is years beyond what I would have expected even 1 year ago. This guy doesn't know what he's talking about, he's just picking and choosing one-off articles that make it seem like it's supporting his points.
I wholeheartedly agree. Shitty companies steal art and then put out shitty products that shitty people use to spam us with slop.
The same goes for code as well.
I’ve explored Claude code/antigravity/etc, found them mostly useless, tried a more interactive approach with copilot/local models/ tried less interactive “agents”/etc. it’s largely all slop.
My coworkers who claim they’re shipping at warp speed using generative AI are almost categorically our worst developers by a mile.
I believe Gary Marcus is quite well known for terrible AI predictions. He's not in any way an expert in the field. Some of his predictions from 2022 [1]
> In 2029, AI will not be able to watch a movie and tell you accurately what is going on (what I called the comprehension challenge in The New Yorker, in 2014). Who are the characters? What are their conflicts and motivations? etc.
> In 2029, AI will not be able to read a novel and reliably answer questions about plot, character, conflicts, motivations, etc. Key will be going beyond the literal text, as Davis and I explain in Rebooting AI.
> In 2029, AI will not be able to work as a competent cook in an arbitrary kitchen (extending Steve Wozniak’s cup of coffee benchmark).
> In 2029, AI will not be able to reliably construct bug-free code of more than 10,000 lines from natural language specification or by interactions with a non-expert user. [Gluing together code from existing libraries doesn’t count.]
> In 2029, AI will not be able to take arbitrary proofs from the mathematical literature written in natural language and convert them into a symbolic form suitable for symbolic verification.
Many of these have already been achieved, and it's only early 2026.
[1]https://garymarcus.substack.com/p/dear-elon-musk-here-are-fi...
Which ones are you claiming have already been achieved?
My understanding of the current scorecard is that he's still technically correct, though I agree with you there is velocity heading towards some of these things being proven wrong by 2029.
For example, in the recent thread about LLMs and solving an Erdos problem I remember reading in the comments that it was confirmed there were multiple LLMs involved as well as an expert mathematician who was deciding what context to shuttle between them and helping formulate things.
Similarly, I've not yet heard of any non-expert Software Engineers creating 10,000+ lines of non-glue code that is bug-free. Even expert Engineers at Cloud Flare failed to create a bug-free OAuth library with Claude at the helm because some things are just extremely difficult to create without bugs even with experts in the loop.
The bug-free code one feels unfalsifiable to me. How do you prove that 10,000 lines of code is bug-free, and then there's a million caveats about what a bug actually is and how we define one.
The second claim about novels seems obviously achieved to me. I just pasted a random obscure novel from project gutenberg into a file and asked claude questions about the characters, and then asked about the motivations of a random side-character. It gave a good answer, I'd recommend trying it yourself.
I agree with you but I'd point out that unless you've read the book it's difficult to know if the answer you got was accurate or it just kinda made it up. In my experience it makes stuff up.
Like, it behaves as if any answer is better than no answer.
So do humans asked to answer tests. The appropriate thing is to compare to human performance at the same task.
At most of these comprehension tasks, AI is already superhuman (in part because Gary picked scaled tasks that humans are surprisingly bad at).
I've done similar loose experiments using a 19th century French novel I have read (L'Agonie by Jean Lombard), and which is quite obscure. There is no English translation of the novel published.
Using Gemini Notebooks I've found it passable at summarizing chapters, listing characters, major theme and so on, and it can do this in English.
"What is the symbolism of the Black Stone?"
> In the sources, the Black Stone (referred to as the Pierre-Noire) serves as the central religious icon of the Syrian Sun cult brought to Rome by the Emperor Elagabalus. Its symbolism is multifaceted, ranging from fertility and virility to a mystical pursuit of universal unity.
> It represents the perennity of the Sun, which fertilizes the world, causing "germs to rise" and spreading them through the atmosphere. It is viewed as the "definitive form" of divinity, intended to absorb and replace the "transitory forms" of all other Roman, Greek, Egyptian, and Persian gods, including the Christian "Kreistos". > Virility and Phallic Symbolism > > The sources explicitly characterize the Black Stone as a gigantic sacred phallus. It is described as: • An icon of virility and the "organ of generation" • A "unisexual icon" that materializes the generative force of nature • A representation of "virility in activity," which is why it is often paraded and elevated in an "orgasmic" or "colossal adoration"
> The Androgyne and Universal Unity - Beyond simple fertility, the philosopher Atillius explains a more complex, "mad" metaphysical project associated with the stone. It symbolizes "Life One" (Vie Une) and the return to a unisexual state of perfection. • The Androgyne: Atillius believes that by pursuing "the male sex by the male sex," the cult "inutilizes" the female sex to eventually create the Androgyne—a self-sufficient being containing both sexes • Unity: The stone signifies the fusion of all generative forces into a single Unity, reversing the "separation of the sexes" which is viewed as a state of unhappiness and impotence. • Marriage of Moon and Sun: The ritual marriage of the goddess Astaroth (representing the Moon and the female principle) to the Black Stone (representing the Sun and the male principle) symbolizes the merging of the Orient and Occident into this unified life principle > > Destruction of the Symbol - The Black Stone ultimately becomes a symbol of Oriental pollution and decadence to the Roman populace. During the final rebellion against Elagabalus, the stone is torn from its temple on the Palatine, defiled with filth, and broken into pieces to ensure that its "signification of Life" would never again dominate Rome.
This is all accurate to the book, even teasing out a couple themes that were only subconsciously present to me.
The NotebookLM version gives citations with links to the original text to support all these assertions, which largely are coherent with that purpose.
The input is raw images of a book scan! Imperfect as it is it still blows my mind. Not that long ago any kind of semantic search or analysis was a very hard AI problem.
"quite obscure" doesn't mean there is nothing in the internet that directly addresses the question.
Here is an english analysis of the text that easily showed up in an internet search:
https://www.cantab.net/users/leonardo/Downloads/Varian%20Sym...
This source includes analysis of "the Black Stone."
Not quite the same analysis. The human is better, no surprise. But the NotebookLM output links back to the original book in a very useful way. If you think about it as fuzzy semantic search it's amazing. If you want an essay or even just creativity, yes it's lacking.
It doesn't have to be the same analysis to put it in a partially overlapping vector space. Not saying it wasn't a useful perspective shuffling in the vector space, but it definitely wasn't original.
LLMs haven't solved any of the 2029 predictions as they were posited. But I expect some will be reached by 2029. The AI hype acts like all this is easy. Not by 2029 doesn't mean impossible or even most of the way there.
LLMs will never achieve anything as long as any victory can be hand waved away with "in the training set". Somehow these models have condensed the entire internet down to a few TB's, yet people aren't backing up their terabytes of personal data down to a couple MB using this same tech...wonder why
1 and 2 have been achieved.
4 is close, the interface needs some work to allow nontechnical people use it. (claude code)
I strongly disagree. I’ve yet to find an AI that can reliably summarise emails, let alone understand nuance or sarcasm. And I just asked ChatGPT 5.2 to describe an Instagram image. It didn’t even get the easily OCR-able text correct. Plus it completely failed to mention anything sports or stadium related. But it was looking at a cliche baseball photo taken by an fan inside the stadium.
I have had ChatGPT read text in an image, give me a 100% accurate result, and then claim not to have the ability and to have guessed the previous result when I ask it to do it again.
I dispute 1 & 2 more than 4.
1) Is it actually watching a movie frame by frame or just searching about it and then giving you the answer?
2) Again can it handle very long novels, context windows are limited and it can easily miss something. Where is the proof for this?
4 is probably solved
4) This is more on predictor because this is easy to game. you can create some gibberish code with LLM today that is 10k lines long without issues. Even a non-technical user can do
I think all of those are terrible indicators, 1 and 2 for example only measure how well LLMs can handle long context sizes.
If a movie or novel is famous the training data is already full of commentary and interpretations of them.
If its something not in the training data, well I don't know many movies or books that use only motives that no other piece of content before them used, so interpreting based on what is similar in the training data still produces good results.
EDIT: With 1 I meant using a transcript of the Audio Description of the movie. If he really meant watch a movie I'd say thats even sillier because well of course we could get another Agent to first generate the Audio Description, which definitely is possible currently.
Just yesterday I saw an article about a police station's AI body cam summarizer mistakenly claim that a police officer turned into a frog during a call. What actually happened was that the cartoon "princess and the frog" was playing in the background.
Sure, another model might have gotten it right, but I think the prediction was made less in the sense of "this will happen at least once" and more of "this will not be an uncommon capability".
When the quality is this low (or variable depending on model) I'm not too sure I'd qualify it as a larger issue than mere context size.
My point was not that those video to text models are good like they are used for example in that case, but more generally I was referring to that list of indicators. Like surely when analysing a movie it is alright if some things are misunderstood by it, especially as the amount of misunderstanding can be decreased a lot. That AI body camera surely is optimized on speed and inference cost. but if you give an agent 10 1s images along with the transcript of that period and the full prior transcript, and give it reasoning capabilities, it would take almost endlessy for that movie to process but the result surely will be much better than the body cameras. After all the indicator talks about "AI" in general so judge a model not optimized for capability but something else to measure on that indicator
> In 2029, AI will not be able to read a novel and reliably answer questions about plot, character, conflicts, motivations, etc. Key will be going beyond the literal text, as Davis and I explain in Rebooting AI.
Can AI actually do this? This looks like a nice benchmark for complex language processing, since a complete novel takes up a whole lot of context (consider War and Peace or The Count of Monte Cristo). Of course the movie variety is even more challenging since it involves especially complex multi-modal input. You could easily extend it to making sense of a whole TV series.
Yes. I am a novelist and I noticed a step change in what was possible here around Claude Sonnet 3.7 in terms of being able to analyze my own unpublished work for theme, implicit motivations, subtext, etc -- without having any pre-digested analysis of the work in its training data.
Yes they can. The size of many codebases is much larger and LLMs can handle those.
Consider also that they can generate summaries and tackle the novel piecemeal, just like a human would.
Re: movies. Get YouTube premium and ask YouTube to summarize a 2hr video for you.
Novel is different from a codebase. In code you can have a relationship between files and most files can be ignored depending on what you're doing. But for a novel, its a sequential thing, in most cases A leads to B and B leads to C and so on.
> Re: movies. Get YouTube premium and ask YouTube to summarize a 2hr video for you.
This is different from watching a movie. Can it tell what suit actor was wearing? Can it tell what the actor's face looked like? Summarising and watching are too different things.
Yes, it is possible to do those things and there are benchmarks for testing multimodal models on their ability to do so. Context length is the major limitation but longer videos can be processed in small chunks whose descriptions can be composed into larger scenes.
https://github.com/JUNJIE99/MLVU
https://huggingface.co/datasets/OpenGVLab/MVBench
Ovis and Qwen3-VL are examples of models that can work with multiple frames from a video at once to produce both visual and temporal understanding
https://huggingface.co/AIDC-AI/Ovis2.5-9B
https://github.com/QwenLM/Qwen3-VL
You’re moving the goalposts. Gary Marcus’ proposal was being able to ask: Who are the characters? What are their conflicts and motivations? etc.
Which is a relatively trivial task for a current LLM.
Gary Marcus' proposal was about a novel, and not a codebase. I think GP's point is that motivations require analysis outside the context window, which LLMs are essentially incapable of doing.
No human reads a novel and evaluates it as a whole. It's a story and the readers perception changes over the course of reading the book. Current AI can certainly do that.
>Can AI actually do this? This looks like a nice benchmark for complex language processing, since a complete novel takes up a whole lot of context (consider War and Peace or The Count of Monte Cristo)
Yes, you just break the book down by chapters or whatever conveniently fits in the context window to produce summaries such that all of the chapter summaries can fit in one context window.
You could also do something with a multi-pass strategy where you come up with a collection of ideas on the first pass and then look back with search to refine and prove/disprove them.
Of course for novels which existed before the time of training an LLM will already contain trained information about so having it "read" classic works like The Count of Monte Cristo and answer questions about it would be a bit of an unfair pass of the test because models will be expected to have been trained on large volumes of existing text analysis on that book.
>reliably answer questions about plot, character, conflicts, motivations
LLMs can already do this automatically with my code in a sizable project (you know what I mean), it seems pretty simple to get them to do it with a book.
> Many of these have already been achieved, and it's only early 2026.
I'm quite sure people who made those (now laughable) predictions will tell you none of these has been achieved, because AI isn't doing this "reliably" or "bug-free."
Defending your predictions is like running an insurance company. You always win.
In my opinion, contrary to other comments here I think AI can do all of the above already except being a kitchen cook.
Just earlier today I asked it to give me a summary of a show I was watching until a particular episode in a particular season without spoiling the rest of it and it did a great job.
Which ones of those have been achieved in your opinion?
I think the arbitrary proofs from mathematical literature is probably the most solved one. Research into IMO problems, and Lean formalization work have been pretty successful.
Then, probably reading a novel and answering questions is the next most successful.
Reliably constructing 10k bug free lines is probably the least successful. AI tends to produce more bugs than human programmers and I have yet to meet a programmer who can reliably produce less than 1 bug per 10k lines.
Formalizing an arbitrary proof is incredibly hard. For one thing, you need to make sure that you've got at least a correct formal statement for all the prereqs you're relying on, or the whole thing becomes pointless. Many areas of math ouside of the very "cleanest" fields (meaning e.g. algebra, logic, combinatorics etc.) have not seen much success in formalizing existing theory developments.
> Reliably constructing 10k bug free lines is probably the least successful.
You imperatively need to try Claude Code, because it absolutely does that.
I have seen many people try to use Claude Code and get LOTS of bugs. Show me any > 10k project you have made with it and I will put the effort in to find one bug free of charge.
Besides being a cook which is more of a robotics problem all of the rest are accomplished to the point of being arguable about how reliably LLMs can perform these tasks, the arguments being between the enthusiast and naysayer camps.
The keyword being "reliably" and what your threshold is for that. And what "bug free" means. Groups of expert humans struggle to write 10k lines of "bug free" code in the absolutist sense of perfection, even code with formal proofs can have "bugs" if you consider the specification not matching the actual needs of reality.
All but the robotics one are demonstrable in 2026 at least.
I'm pretty sure it can do all of those except for the one which requires a physical body (in the kitchen) and the one that humans can't do reliably either (construct 10000 loc bug-free).
This comment or something very close always appears alongside a Gary Marcus post.
And why not? Is there any reason for this comment to not appear?
If Bill Gates made a predication about computing, no matter what the predication says, you can bet that 640K memory quote would be mentioned in the comment section (even he didn't actually say that).
Which is fortunate, considering how asinine it is in 2026 to expect that none of the items listed will be accomplished in the next 3.9 years.
I think it’s for good reason. I’m a bit at a loss as to why every time this guy rages into the ether of his blog it’s considered newsworthy. Celebrity driven tech news is just so tiresome. Marcus was surpassed by others in the field and now he’s basically a professional heckler on a university payroll. I wish people could just be happy for the success of others instead of fuming about how so and so is a billionaire and they are not.
Meanwhile, my cofounder is rewriting code we spent millions of salary on in the past by himself in a few weeks.
I myself am saving a small fortune on design and photography and getting better results while doing it.
If this is not all that well I can’t wait until we get to mediocre!
> Meanwhile, my cofounder is rewriting code we spent millions of salary on in the past by himself in a few weeks.
Why?
Im not even casting shade - I think AI is quite amazing for coding and can increase productivity and quality a lot.
But I'm curious why he's doing this.
> Meanwhile, my cofounder is rewriting code we spent millions of salary on in the past by himself in a few weeks.
Code is not an asset it's a liability, and code that no one has reviewed is even more of a liability.
However, in the end, execution is all that matters so if you and your cofounder are able to execute successfully with mountains of generated code then it doesn't matter what assets and liabilities you hold in the short term.
The long term is a lot harder to predict in any case.
> Code is not an asset it's a liability, and code that no one has reviewed is even more of a liability.
Code that solves problems and makes you money is by definition an asset. Whether or not the code in question does those things remains to be seen, but code is not strictly a liability or else no one would write it.
"Code is a liability. What the code does for you is an asset." as quoted from https://wiki.c2.com/?SoftwareAsLiability with Last edit December 17, 2013.
This discussion and distinction used to be well known, but I'm happy to help some people become "one of today's lucky 10,000" as quoted from https://xkcd.com/1053/ because it is indeed much more interesting than the alternative approach.
Developers that can’t see the change are blind.
Just this week, sun-tue. I added a fully functional subscription model to an existing platform, build out a bulk async elasticjs indexing for a huge database and migrated a very large Wordpress website to NextJS. 2.5 days, would have cost me at least a month 2 years ago.
To me, this sounds like:
AI is helping me solve all the issues that using AI has caused.
Wordpress has a pretty good export and Markdown is widely supported. If you estimate 1 month of work to get that into NextJS, then maybe the latter is not a suitable choice.
It's not directly comparable. The first time writing the code is always the hardest because you might have to figure out the requirements along the way. When you have the initial system running for a while, doing a second one is easier because all the requirements kinks are figured out.
By the way, why does your co-founder have to do the rewrite at all?
You can compare it - just factor that in. And compare writing it with AI vs. writing it without AI.
We have no clue the scope of the rewrite but for anything non-trivial, 2 weeks just isn't going to be possible without AI. To the point of you probably not doing it at all.
I have no idea why they are rewriting the code. That's another matter.
I find the opposite to be true. Once you know the problem you’re trying to solve (which admittedly can be the biggest lift), writing the fist cut of the code is fun, and you can design the system and set precedent however you want. Once it’s in the wild, you have to work within the consequences of your initial decisions, including bad ones.
lol same. I just wrote a bunch of diagrams with mermaid that would legit take me a week, also did a mock of an UI for a frontend engineer that would take me another week to do .. or some designers. All of that in between meetings...
Waiting for it to actually go well to see what else I can do !
The more I have this experience and read people maligning AI for coding, the more I think the junior developers are actually not the ones in danger.
Oh I've thought this for years. As an L7, basically my primary role is to serve as someone to bounce ideas off of, and to make recommendations based on experience. A chatbot, with its virtually infinite supply of experience, could ostensibly replace my role way sooner than it could a solid junior/mid-level coder. The main thing it needs is a consistent vision and direction that aligns with the needs of nearby teams, which frankly sounds not all that hard to write in code (I've been considering doing this).
Probably the biggest gap would be the ability to ignite, drive, and launch new initiatives. How does an AI agent "lead" an engineering team? That's not something you can code up in an agent runtime. It'd require a whole culture change that I have a hard time seeing in reality. But of course if there comes a point where AI takes all the junior and mid-level coding jobs, then at that point there's no culture to change, so staff/principal jobs would be just as at risk.
Sounds like an argument for better hiring practices and planning.
Producing a lot of code isn’t proof of anything.
Yep. Let’s see the projects and more importantly the incremental returns…
Out of curiosity, what is your product?
I suspect he means as a trillion dollar corporation led endeavor.
I trained a small neural net on pics of a cat I had in the 00s (RIP George, you were a good cat).
Mounted a webcam I had gotten for free from somewhere, above the cat door, in the exterior of the house.
If the neural net recognized my cat it switched off an electromagnetic holding the pet door locked. Worked perfectly until I moved out of the rental.
Neural nets are, end of the day, pretty cool. It's the data center business that's the problem. Just more landlords, wannabe oligarchs, claiming ownership over anything they can get the politicians to give them.
The problem is... you're going to deprive yourself of the talent chain in the long run, and so is everyone else who is switching over to AI, both generative like ChatGPT and transformative like the various translation, speech recognition/transcription or data wrangling models.
For now, it works out for companies - but forward to, say, ten years in the future. There won't be new intermediates or seniors any more to replace the ones that age out or quit the industry entirely in frustration of them not being there for actual creativity but to clean up AI slop, simply because there won't have been a pipeline of trainees and juniors for a decade.
But by the time that plus the demographic collapse shows its effects, the people who currently call the shots will be in pension, having long since made their money. And my generation will be left with collapse everywhere and find ways to somehow keep stuff running.
Hell, it's already bad to get qualified human support these days. Large corporations effectively rule with impunity, with the only recourse consumers have being to either shell out immense sums of money for lawyers and court fees or turning to consumer protection/regulatory authorities that are being gutted as we speak both in money and legal protections, or being swamped with AI slop like "legal assistance" AI hallucinating case law.
> Meanwhile, my cofounder is rewriting code we spent millions of salary on in the past by himself in a few weeks.
This is one of those statements that would horrify any halfway competent engineer. A cowboy coder going in, seeing a bunch of code and going 'I should rewrite this' is one of the biggest liabilities to any stable system.
I assume this is because they're already insanely profitable after hitting PMF and are now trying to bring down infra costs?
Right? RIGHT?!
Every professional SWE is going to stare off into the middle distance, as they flashback to some PM or VP deciding to show everyone they still got it.
The "how hard could it be" fallacy claims another!
As someone who is more involved in shaping the product direction rather than engineering what composes the product - I will readily admit many product people are utterly, utterly clueless.
Most people have no clue the craftsmanship, work etc it takes to create a great product. LLMs are not going to change this, in fact they serve as a distraction.
I’m not a SWE so I gain nothing by being bearish on the contributions of LLMs to the real economy ;)
Definitely been in that room multiple times.
I find it a bit odd that people are acting like this stuff is an abject failure because it's not perfect yet.
Generative AI, as we know it, has only existed ~5-6 years, and it has improved substantially, and is likely to keep improving.
Yes, people have probably been deploying it in spots where it's not quite ready but it's myopic to act like it's "not going all that well" when it's pretty clear that it actually is going pretty well, just that we need to work out the kinks. New technology is always buggy for awhile, and eventually it becomes boring.
> Generative AI, as we know it, has only existed ~5-6 years, and it has improved substantially, and is likely to keep improving.
Every 2/3 months we're hearing there's a new model that just blows the last one out of the water for coding. Meanwhile, here I am with Opus and Sonnet for $20/mo and it's regularly failing at basic tasks, antigravity getting stuck in loops and burning credits. We're talking "copy basic examples and don't hallucinate APIs" here, not deep complicated system design topics.
It can one shot a web frontend, just like v0 could in 2023. But that's still about all I've seen it work on.
You’re doing exactly the thing that the parent commenter pointed out: Complaining that they’re not perfect yet as if that’s damning evidence of failure.
We all know LLMs get stuck. We know they hallucinate. We know they get things wrong. We know they get stuck in loops.
There are two types of people: The first group learns to work within these limits and adapt to using them where they’re helpful while writing the code when they’re not.
The second group gets frustrated every time it doesn’t one-shot their prompt and declares it all a big farce. Meanwhile the rest of us are out here having fun with these tools, however limited they are.
Someone else said this perfectly farther down:
> The whole discourse around LLMs is so utterly exhausting. If I say I don't like them for almost any reason, I'm a luddite. If I complain about their shortcomings, I'm just using it wrong. If I try and use it the "right" way and it still gets extremely basic things wrong, then my expectations are too high.
As I’ve said, I use LLMs, and I use tools that are assisted by LLMs. They help. But they don’t work anywhere near as reliably as people talk about them working. And that hasn’t changed in the 18 months since I first promoted v0 to make me a website.
Sure, but think about what it's replacing.
If you hired a human, it will cost you thousands a week. Humans will also fail at basic tasks, get stuck in useless loops, and you still have to pay them for all that time.
For that matter, even if I'm not hiring anyone, I will still get stuck on projects and burn through the finite number of hours I have on this planet trying to figure stuff out and being wrong for a lot of it.
It's not perfect yet, but these coding models, in my mind, have gotten pretty good if you're specific about the requirements, and even if it misfires fairly often, they can still be useful, even if they're not perfect.
I've made this analogy before, but to me they're like really eager-to-please interns; not necessarily perfect, and there's even a fairly high risk you'll have to redo a lot of their work, but they can still be useful.
I am an AI-skeptic but I would agree this looks impressive from certain angles, especially if you're an early startup (maybe) or you are very high up the chain and just want to focus on cutting costs. On the other hand, if you are about to be unemployed, this is less impressive. Can it replace a human? I would say no its still long way to go, but a good salesman can convince executives that it does and thats all that matters.
> On the other hand, if you are about to be unemployed, this is less impressive
> salesman can convince executives that it does
I tend to think that reality will temper this trend as the results develop. Replacing 10 engineers with one engineer using Cursor will result in a vast velocity hit. Replacing 5 engineers with 5 "agents" assigned to autonomously implement features will result in a mess eventually. (With current technology -- I have no idea what even 2027 AI will do). At that point those unemployed engineers will find their phones ringing off the hook to come and clean up the mess.
Not that unlike what happens in many situations where they fire teams and offshore the whole thing to a team of average developers 180 degrees of longitude away who don't have any domain knowledge of the business or connections to the stakeholders. The pendulum swings back in the other direction.
I just think Jevins paradox [1]/Gustafson's Law [2] kind of applies here.
Maybe I shouldn't have used the word "replaced", as I don't really think it's actually going to "replace" people long term. I think it's likely to just lead to higher output as these get better and better .
[1] https://en.wikipedia.org/wiki/Jevons_paradox
[2] https://en.wikipedia.org/wiki/Gustafson%27s_law
Not you, but the word replaced is the being used all the time. Even senior engineers are saying they are using it as a junior engineers while we can easily hire junior engineers (but Execs don't want to). Jevon's paradox wont work in Software because user's wallets and time is limited, and if software becomes too easy to build, it becomes harder to sell. Normal people can have 5 subscriptions, may be 10, but they wont be going to 50 or 100. I would say we would have already exhausted users already, with all the bad practices.
You’ve missed my point here - I agree that gen AI has changed everything and is useful, _but_ I disagree that it’s improved substantially - which is what the comment I replied to claimed.
Anecdotally I’ve seen no difference in model changes in the last year, but going from LLM to Claude code (where we told the LLMs they can use tools on our machines) was a game changer. The improvement there was the agent loop and the support for tools.
In 2023 I asked v0.dev to one shot me a website for a business I was working on and it did it in about 3 minutes. I feel like we’re still stuck there with the models.
In my experience it has gotten considerably better. When I get it to generate C, it often gets the pointer logic correct, which wasn't the case three years ago. Three years ago, ChatGPT would struggle with even fairly straightforward LaTeX, but now I can pretty easily get it to generate pretty elaborate LaTeX and I have even had good success generating LuaTeX. I've been able to fairly successfully have it generate TLA+ spec from existing code now, which didn't work even a year ago when I tried it.
Of course, sample size of one, so if you haven't gotten those results then fair enough, but I've at least observed it getting a lot better.
>Every 2/3 months we're hearing there's a new model that just blows the last one out of the water for coding
I haven't heard that at all. I hear about models that come out and are a bit better. And other people saying they suck.
>Meanwhile, here I am with Opus and Sonnet for $20/mo and it's regularly failing at basic tasks, antigravity getting stuck in loops and burning credits.
Is it bringing you any value? I find it speeds things up a LOT.
> We're talking "copy basic examples and don't hallucinate APIs" here, not deep complicated system design topics.
If your metric is an LLM that can copy/paste without alterations, and never hallucinate APIs, then yeah, you'll always be disappointed with them.
The rest of us learn how to be productive with them despite these problems.
> If your metric is an LLM that can copy/paste without alterations, and never hallucinate APIs, then yeah, you'll always be disappointed with them.
I struggle to take comments like this seriously - yes, it is very reasonable to expect these magical tools to copy and paste something without alterations. How on earth is that an unreasonable ask?
The whole discourse around LLMs is so utterly exhausting. If I say I don't like them for almost any reason, I'm a luddite. If I complain about their shortcomings, I'm just using it wrong. If I try and use it the "right" way and it still gets extremely basic things wrong, then my expectations are too high.
What, precisely, are they good for?
It seems like just such a weird and rigid way to evaluate it? I am a somewhat reasonable human coder, but I can't copy and paste a bunch of code without alterations from memory either. Can someone still find a use for me?
I think what they're best at right now is the initial scaffolding work of projects. A lot of the annoying bootstrap shit that I hate doing is actually generally handled really well by Codex.
I agree that there's definitely some overhype to them right now. At least for the stuff I've done they have gotten considerably better though, to a point where the code it generates is often usable, if sub-optimal.
For example, about three years ago, I was trying to get ChatGPT to write me a C program to do a fairly basic ZeroMQ program. It generated something that looked correct, but it would crash pretty much immediately, because it kept trying to use a pointer after free.
I tried the same thing again with Codex about a week ago, and it worked out of the box, and I was even able to get it to do more stuff.
I think it USED to be true that you couldn't really use an LLM on a large, existing codebase. Our codebase is about 2 million LOC, and a year ago you couldn't use an LLM on it for anything but occasional small tasks. Now, probably 90% of the code I commit each week was written by Claude (and reviewed by me and other humans - and also by Copilot and ZeroPath).
For a long time, I've wanted to write a blog post on why programmers don't understand the utility of LLMs[1], whereas non-programmers easily see it. But I struggle to articulate it well.
The gist is this: Programmers view computers as deterministic. They can't tolerate a tool that behaves differently from run to run. They have a very binary view of the world: If it can't satisfy this "basic" requirement, it's crap.
Programmers have made their career (and possibly life) being experts at solving problems that greatly benefit from determinism. A problem that doesn't - well either that needs to be solved by sophisticated machine learning, or by a human. They're trained on essentially ignoring those problems - it's not their expertise.
And so they get really thrown off when people use computers in a nondeterministic way to solve a deterministic problem.
For everyone else, the world, and its solutions, are mostly non-deterministic. When they solve a problem, or when they pay people to solve a problem, the guarantees are much lower. They don't expect perfection every time.
When a normal human asks a programmer to make a change, they understand that communication is lossy, and even if it isn't, programmers make mistakes.
Using a tool like an LLM is like any other tool. Or like asking any other human to do something.
For programmers, it's a cardinal sin if the tool is unpredictable. So they dismiss it. For everyone else, it's just another tool. They embrace it.
[1] This, of course, is changing as they become better at coding.
I’m perfectly happy for my tooling to not be deterministic. I’m not happy for it to make up solutions that don’t exist, and get stuck in loops because of that.
I use LLMs, I code with a mix of antigravity and Claude code depending on the task, but I feel like I’m living in a different reality when the code I get out of these tools _regularly just doesn’t work, at all_. And to the parents point, I’m doing something wrong for noticing that?
If it were terrible, you wouldn't use them, right? Isn't the fact that you continue to use AI coding tools a sign that you find them a net positive? Or is it being imposed on you?
> And to the parents point, I’m doing something wrong for noticing that?
There's nothing wrong pointing out your experience. What the OP was implying was he expects them to be able to copy/paste reliably almost 100% of the time, and not hallucinate. I was merely pointing out that he'll never get that with LLMs, and that their inability to do so isn't a barrier to getting productive use out of them.
> What, precisely, are they good for?
scamming people
Its strong enough to replace humans at their jobs and weak enough that it cant do basic things. Its a paradox. Just learn to be productive with them. Pay $200/month and work around with its little quirks. /s
There’s a subtle point a moment when you HAVE to take the driver wheel from the AI. All issues I see are from people insisting to use far beyond the point it stops being useful.
It is a helper, a partner, it is still not ready go the last mile
It's funny how many people don't get that. It's like adding a pretty great senior or staff level engineer to sit on-call next to every developer and assist them, for basically free (I've never used any of the expensive stuff yet. Just things like Copilot, Grok Code in JetBrains, just asking Gemini to write bits of code for me).
If you hired a staff engineer to sit next to me, and I just had him/her write 100% of the code and never tried to understand it, that would be an unwise decision on my part and I'd have little room to complain about the times he made mistakes.
As someone else said in this thread:
> The whole discourse around LLMs is so utterly exhausting. If I say I don't like them for almost any reason, I'm a luddite. If I complain about their shortcomings, I'm just using it wrong. If I try and use it the "right" way and it still gets extremely basic things wrong, then my expectations are too high.
I’m perfectly happy to write code, to use these tools. I do use them, and sometimes they work (well). Other times they have catastrophic failures. But apparently it’s my failure for not understanding the tool or expecting too much of the tool, while others are screaming from the rooftops about how this new model changes everything (which happens every 3 months at this point)
There's no silver bullet. I’m not a researcher, but I’ve done my best to understand how these systems work—through books, video courses, and even taking underpaid hourly work at a company that creates datasets for RLHF. I spent my days fixing bugs step-by-step, writing notes like, “Hmm… this version of the library doesn’t support protocol Y version 4423123423. We need to update it, then refactor the code so we instantiate ‘blah’ and pass it to ‘foo’ before we can connect.”
That experience gave me a deep appreciation for how incredible LLMs are and the amazing software they can power—but it also completely demystified them. So by all means, let’s use them. But let’s also understand there are no miracles here. Go back to Shannon’s papers from the ’60s, and you'll understand that what seems to you like "emerging behaviors" are quite explainable from an information theory background. Learn how these models are built. Keep up with the latests research papers. If you do, you’ll recognize their limitations before those limitations catch you by surprise.
There is no silver bullet. And if you think you’ve found one, you’re in for a world of pain. Worse still, you’ll never realize the full potential of these tools, because you won’t understand their constraints, their limits, or their pitfalls.
>Generative AI, as we know it, has only existed ~5-6 years
Probably less than that, practically speaking. ChatGPT's initial release date was November 2022. It's closer to 3 years, in terms of any significant amount of people using them.
We implement pretty cool workflows at work using "GenAI" and the users of our software are really appreciative. It's like saying a hammer sucks because it breaks most things you hit with it.
> Generative AI, as we know it, has only existed ~5-6 years, and it has improved substantially, and is likely to keep improving.
I think the big problem is that the pace of improvement was UNBELIEVABLE for about 4 years, and it appears to have plateaued to almost nothing.
ChatGPT has barely improved in, what, 6 months or so.
They are driving costs down incredibly, which is not nothing.
But, here's the thing, they're not cutting costs because they have to. Google has deep enough pockets.
They're cutting costs because - at least with the current known paradigm - the cost is not worth it to make material improvements.
So unless there's a paradigm shift, we're not seeing MASSIVE improvements in output like we did in the previous years.
You could see costs go down to 1/100th over 3 years, seriously.
But they need to make money, so it's possible non of that will be passed on.
I think that even if it never improves, its current state is already pretty useful. I do think it's going to improve though I don't think AGI is going to happen any time soon.
I have no idea what this is called, but it feels like a lot of people assume that progress will continue at a linear pace for forever for things, when I think that generally progress is closer to a "staircase" shape. A new invention or discovery will lead to a lot of really cool new inventions and discoveries in a very short period of time, eventually people will exhaust the low-to-middle-hanging fruit, and progress kind of levels out.
I suspect it will be the same way with AI; I don't now if we've reached the top of our current plateau, but if not I think we're getting fairly close.
They are focused on reducing costs in order to survive. Pure and simple.
Alphabet / Google doesn’t have that issue. OAI and other money losing firms do.
>and is likely to keep improving.
I'm not trying to be pedantic, but how did you arrive at 'keep improving' as a conclusion? Nobody is really sure how this stuff actually works. That's why AI safety was such a big deal a few years ago.
Totally reasonable question, and I only am making an assumption based on observed progress. AI generated code, at least in my personal experience, has gotten a lot better, and while I don't think that will go to infinity, I do think that there's still more room for improvement that could happen.
I will acknowledge that I don't have any evidence of this claim, so maybe the word "likely" was unwise, as that suggests probability. Feel free to replace "is "likely to" with "it feels like it will".
Because the likes of Altman have set short term expectations unrealistically high.
I mean that's every tech company.
I made a joke once after the first time I watched one of those Apple announcement shows in 2018, where I said "it's kind of sad, because there won't be any problems for us to solve because the iPhone XS Max is going to solve all of them".
The US economy is pretty much a big vibes-based Ponzi scheme now, so I don't think we can single-out AI, I think we have to blame the fact that the CEOs running these things face no negative consequences for lying or embellishing and they do get rewarded for it because it will often bump the stock price.
Is Tesla really worth more than every other car company combined in any kind of objective sense? I don't think so, I think people really like it when Elon lies to them about stuff that will come out "next year", and they feel no need to punish him economically.
"Ponzi" requires records fraud and is popularly misused, sort of like if people started describing every software bug as "a stack overflow."
I'd rather characterize it as extremes of Greater Fool Theory.
https://en.wikipedia.org/wiki/Greater_fool_theory
I maintain that most anti-AI sentiment is actually anti-lying-tech-CEO sentiment misattributed.
The technology is neat, the people selling it are ghouls.
Exactly: the technology is useful but because the executive class is hyping it as close to AGI because their buddies are slavering for layoffs. If that “when do you get fired?” tone wasn’t behind the conversation, I think a lot of people would be interested in applying LLMs to the smaller subset of things they actually perform well at.
Maybe CEOs should face consequences for going on the stage and outwardly lying. Instead they're rewarded by a bump in stock price because people appear to have amnesia.
This is how I felt about Bitcoin.
I think that the wider industry is living right now what was coding and software engineering around 1 year or so ago.
Yeah you could ask ChatGPT or Claude to write code, but it wasn't really there.
It needs a while to adopt the model AND the UI. As in software are the first one because we are both makers and users.
This feels like a pretty low effort post that plays heavily to superficial reader's cognitive biases.
I work commercializing AI in some very specific use cases where it extremely valuable. Where people are being lead astray is layering generalizations: general use cases (copilots) deployed across general populations and generally not doing very well. But that's PMF stuff, not a failure of the underlying tech.
I think both sides of this debate are conflating the tech and the market. First of all, there were forms of "AI" before modern Gen AI (machine learning, NLP, computer vision, predictive algorithms, etc) that were and are very valuable for specific use cases. Not much has changed there AFAICT, so it's fair that the broader conversation about Gen AI is focused on general use cases deployed across general populations. After all, Microsoft thinks it's a copilot company, so it's fair to talk about how copilots are doing.
On the pro-AI side, people are conflating technology success with product success. Look at crypto -- the technology supports decentralization, anonymity, and use as a currency; but in the marketplace it is centralized, subject to KYC, and used for speculation instead of transactions. The potential of the tech does not always align with the way the world decides to use it.
On the other side of the aisle, people are conflating the problematic socio-economics of AI with the state of the technology. I think you're correct to call it a failure of PMF, and that's a problem worth writing articles about. It just shouldn't be so hard to talk about the success of the technology and its failure in the marketplace in the same breath.
> This feels like a pretty low effort post that plays heavily to superficial reader's cognitive biases.
I haven’t followed this author but the few times he’s come up his writings have been exactly this.
A year ago I would have agreed wholeheartedly and I was a self confessed skeptic.
Then Gemini got good (around 2.5?), like I-turned-my-head good. I started to use it every week-ish, not to write code. But more like a tool (as you would a calculator).
More recently Opus 4.5 was released and now I'm using it every day to assist in code. It is regularly helping me take tasks that would have taken 6-12 hours down to 15-30 minutes with some minor prompting and hand holding.
I've not yet reached the point where I feel letting is loose and do the entire PR for me. But it's getting there.
> I was a self confessed skeptic.
I think that's the key. Healthy skepticism is always appropriate. It's the outright cynicism that gets me. "AI will never be able to [...]", when I've been sitting here at work doing 2/3rds of those supposedly impossible things. Flawlessly? No, of course not! But I don't do those things flawlessly on the first pass, either.
Skepticism is good. I have no time or patience for cynics who dismiss the whole technology as impossible.
I would strongly recommend this podcast episode with Andrej Karpathy. I will poorly summarize it by saying his main point is that AI will spread like any other technology. It’s not going to be a sudden flash and everything is done by AI. It will be a slow rollout where each year it automates more and more manual work, until one day we realize it’s everywhere and has become indispensable.
It sounds like what you are seeing lines up with his predictions. Each model generation is able to take on a little more of the responsibilities of a software engineer, but it’s not as if we suddenly don’t need the engineer anymore.
https://www.dwarkesh.com/p/andrej-karpathy
Though I think it's a very steep sigmoid that we're still far on the bottom half of.
For math it just did its first "almost independent" Erdos problem. In a couple months it'll probably do another, then maybe one each month for a while, then one morning we'll wake up and find whoom it solved 20 overnight and is spitting them out by the hour.
For software it's been "curiosity ... curiosity ... curiosity ... occasionally useful assistant ... slightly more capable assistant" up to now, and it'll probably continue like that for a while. The inflection point will be when OpenAI/Anthropic/Google releases an e2e platform meant to be driven primarily by the product team, with engineering just being co-drivers. It probably starts out buggy and needing a lot of hand-holding (and grumbling) from engineering, but slowly but surely becomes more independently capable. Then at some point, product will become more confident in that platform than their own engineering team, and begin pushing out features based on that alone. Once that process starts (probably first at OpenAI/Anthropic/Google themselves, but spreading like wildfire across the industry), then it's just a matter of time until leadership declares that all feature development goes through that platform, and retains only as many engineers as is required to support the platform itself.
AI first of all is not a technology.
Can people get their words straight before typing?
Is LLM a technology? Are you complaining about the use of AI to mean LLM? Because I think that ship has sailed
Gary Marcus (probably): "Hey this LLM isn't smarter than Einstein yet, it's not going all that well"
The goalposts keep getting pushed further and further every month. How many math and coding Olympiads and other benchmarks will LLMs need to dominate before people will actually admit that in some domains it's really quite good.
Sure, if you're a Nobel prize winner or PhD then LLMs aren't as good as you yet, but for 99% of the people in the world, LLMs are better than you at Math, Science, Coding, and every language probably except your native language, and it's probably better at you at that too...
This post is literally just 4 screenshots of articles, not even its own commentary or discussion.
Don’t be too harsh, it’s the most effort Gary has put into his criticism in a while </s>
I appreciate good critique but this is not it
Ignoring the actual poor quality of this write-up, I think we don't know how well GenAI is going to be honest. I feel we've not been able to properly measure or assess it's actual impact yet.
Even as I use it, and I use it everyday, I can't really assess its true impact. Am I more productive or less overall? I'm not too sure. Do I do higher quality work or lower quality work overall? I'm not too sure.
All I know, it's pretty cool, and using it is super easy. I probably use it too much, in a way, that it actually slows things down sometimes, when I use it for trivial things for example.
At least when it comes to productivity/quality I feel we don't really know yet.
But there are definite cool use-cases for it, I mean, I can edit photos/videos in ways I simply could not before, or generate a logo for a birthday party, I couldn't do that before. I can make a tune that I like, even if it's not the best song in the world, but it can have the lyrics I want. I can have it extract whatever from a PDF. I can have it tell me what to watch out for in a gigantic lease agreement I would not have bothered reading otherwise.
I can have it fix my tests, or write my tests, not sure if it saves me time, but I hate doing that, so it definitely makes it more fun and I can kind of just watch videos at the same time, what I couldn't before. Coding quality of life improvements are there too, I want to generate a sample JSON out of a JSONSchema, and so on. If I want, I can write the a method using English prompts instead of the code itself, might not truly be faster or not, not sure, but sometimes it's less mentally taxing, depending on my mood, it can be more fun or less fun, etc.
All those are pretty awesome wins and a sign that for sure those things will remain and I will happily pay for them. So maybe it depends on what you expected.
And what do you think investors in OAI et al are expecting?
Meanwhile $employer is continuing to migrate individual tasks to in-house AI tooling, and has licensed an off-the-shelf coding agent for all of us developers to put in our IDEs.
You're absolutely right!
The irony of a five sentence article making giant claims isn't lost on me. Don't get me wrong: I'm amenable to the idea; but, y'know, my kids wrote longer essays in 4th grade.
Should have used an LLM to proofread.. LLMs can still cannot be trusted?
How dare you accuse Gary-Marcus-5.2-2025-12-11 of being an LLM??
a historic moron. Marcus will make Krugman's internet==fax machine look like a good prediction
What a joke this guy is. I can sit down and crank out a real, complex feature in a couple hours that would have previously taken days and ship it to the users of our AI platform who can then respond to RFQs in minutes where they would have previously spent hours matching descriptions to part numbers manually.
...and yet we still see these articles claiming LLMs are dying/overhyped/major issues/whatever.
Cool man, I'll just be over here building my AI based business with AI and solving real problems in the very real manufacturing sector.
LLMs help me read code 10x faster - I’ll take the win and say thanks
I see stuff like this and think of these two things:
1) https://en.wikipedia.org/wiki/Gartner_hype_cycle
or
2) "First they ignore you, then they laugh at you, then they fight you, then you win."
or maybe originally:
"First they ignore you. Then they ridicule you. And then they attack you and want to burn you. And then they build monuments to you"
Guessing this isn’t going to be popular here, but he’s right. AI has some use cases, but isn’t the world-changing paradigm shift it’s marketed as. It’s becoming clear the tech is ultimately just a tool, not a precursor to AGI.
Is that the claim the OP is making?
If AGI is ever going to happen, then it's definitionally a precursor to it.
So I'm not really sure how to parse your statement.
I’m not sure I follow. What if LLMs are helpful but not useful to AGI, but some other technology is? Seems likely.
The comment wasn't referencing LLMs, but generative AI.
Even then, given the deep impact of LLMs and how many people are using them already, it's a stretch to say LLMs will have no effect on the development of AGI.
I think it's pretty obvious that AGI requires something more than LLMs, but I think it's equally obvious LLMs will have been involved in its development somewhere, even if just a stepping stone. So, a "precursor".
not YET.
First of all, popping in a few screenshots of articles and papers is not proper analysis.
Second of all, GenAI is going well or not depending on how we frame it.
In terms of saving time, money and effort when coding, writing, analysing, researching, etc. It’s extremely successful.
In terms of leading us to AGI… GenAI alone won’t reach that. Current ROI is plateauing, and we need to start investing more somewhere else.
I keep reading comments that claim GenAI's positive traits, but this usually amounts to some toy PoC that very eerily mirrors work found in code bootcamps. You want an app that has logins and comments and upvotes? GenAI is going to look amazing setting up a non-relational db to your node backend.
> LLMs can still cannot be trusted
But can they write grammatically correct statements?
Haters gonna hate.
Meanwhile I'm over here reducing my ADO ticket time estimates by 75%.
I’m starting to think this take is legitimately insane.
As said in the article, a conservative estimate is that Gen AI can currently do 2.5% of all jobs in the entire economy. A technology that is really only a couple of years old. This is supposed to be _disappointing_? That’s millions of jobs _today_, in a totally nascent form.
I mean I understand skepticism, I’m not exactly in love with AI myself, but the world has literally been transformed.
Odds this was AI generated?
It's literally just four screenshots paired with this sentence.
> Trying to orient our economy and geopolitical policy around such shoddy technology — particularly on the unproven hopes that it will dramatically improve– is a mistake.
The screenshots are screenshots of real articles. The sentence is shorter than a typical prompt.
It's more about how you use it. It should be a source of inspo. Not the end all be all.
How on Earth do people keep taking Gary Marcus seriously?
He's such a joke that even LLMs make fun of him. The Gemini-generated Hacker News frontpage for December 9 2035 contains an article by Gary Marcus: "AI progress is stalling": https://dosaygo-studio.github.io/hn-front-page-2035/news
As if the articles he’s linked were written by him
It's going well for coding. I just knocked out a mapping project that would have been a week+ of work (with docs and stackoverflow opened in the background) in a few hours.
And yes, I do understand the code and what is happening and did have to make a couple of adjustments manually.
I don't know that reducing coding work justifies the current valuations, but I wouldn't say it's "not going all that well".
Holy moving goal posts batman!
I hate generative AI, but its inarguable what we have now would have been considered pure magic 5 years ago.
Download models you can find now and forever. The guardrails will only get worse, or models banned entirely. Whether it's because of "hurts people's health" or some other moral panic, it will kill this tech off.
gpt-oss isn't bad, but even models you cannot run are worth getting since you may be able to run them in the future.
I'm hedging against models being so nerfed they are useless. (This is unlikely, but drives are cheap and data is expensive.)
This entire take is nonsense.
I just used ChatGPT to diagnose a very serious but ultimately not-dangerous health situation last week and it was perfect. It literally guided me perfectly without making me panic and helped me understand what was going on.
We use ChatGPT at work to do things that we have literally laid people off for, because we don't need them anymore. This included fixing bugs at a level that is at least E5/senior software engineer. Sometimes it does something really bad but it definitely saves times and helps avoid adding headcount.
Generative AI is years beyond what I would have expected even 1 year ago. This guy doesn't know what he's talking about, he's just picking and choosing one-off articles that make it seem like it's supporting his points.
I wholeheartedly agree. Shitty companies steal art and then put out shitty products that shitty people use to spam us with slop.
The same goes for code as well.
I’ve explored Claude code/antigravity/etc, found them mostly useless, tried a more interactive approach with copilot/local models/ tried less interactive “agents”/etc. it’s largely all slop.
My coworkers who claim they’re shipping at warp speed using generative AI are almost categorically our worst developers by a mile.