The gap in shared understanding of LLM capability is widening

https://xxcancel.com/karpathy/status/2109361546505966046

It's actually sad how much AI has either not improved life at all or actively made it worse from the perspective of the average person:

Your car is still the same. Your dishwasher is still the same. The train you take to work is still the same and never comes on time. Fuel is more expensive. The roads are still congested with traffic. Your kids are (probably) doing worse at school. Food costs more. Houses cost more. Rent is higher. Buying a computer or a PS5 is more expensive. Politics is still full of mentally unstable people. The environment is getting worse. You will still die of heart disease or cancer. Food quality is worse. Your kids can't find work. Wealth inequality is accelerating.

But hey...on the flip side...a lot of people are rapidly building software (that nobody is using), and AI is solving mathematics problems (that nobody - including mathematicians - wanted it to solve)

Imagine what a "country of geniuses in a datacenter" will be able to do? Apparently...nothing.

I think there is also a subgroup in each of those bands (maybe 10-25%) that have a good estimation of what the band above them are experiencing.

This is why you can see people talking about the capabilities of models that they can't currently access.

That accuracy is probably a normal distribution with people overestimating and underestimating to similar degrees,

The population of each band would, however, mean the overestimaters outnumber the those who have an accurate perception based upon actual experience.

It does seem quite the recipe for murkiness, even before you take into account the nature of engagement driven media coverage.

Don't think the 0.2% really are in agreement that "large, complex projects that used to take them weeks/months can now be completed by agents with a prompt".

The divide seems to be between people who think you may need to understand the codebase for maintenance, potential future token costs, optimisation, code erosion & being wary of the assumptions/mistakes that they have seen LLMs make, and people who don't review code and are confident that the AI has made no mistakes and will be available for next-to-nothing forever.

It's certainly allowing the latter to move fast... but will their clients be happy if/when they break things. We'll find out soon, I'm sure.

It is always a mistake to assume that because computers can do X, that they will be able do everything that a human that can do X is expected to be able to do.

Solving frontier math problems is mind boggling, but perhaps it says more about frontier math problems than it does about LLMs. Same with chess which was solved decades ago.

My experience is that LLMs can be useful but are underwhelming in terms of intelligence, namely the ability to quickly learn from experience and grow with you while you accomplish a task.

Thinking of them as beings is something we should move away from.

What exactly do we want the general public to "do" with these models? If the have no use for agentic LLMs, they obviously aren't going to be interested in the capability of those agents.

I see a lot of benefit to spreading information around the tech, but it's still in its infancy. Most people weren't on the internet when that was getting up to speed.

Meemaw and pops have their systems that they work within, and they just got familiar with the damn smartphone. Though it might help them, they probably do not want AI to organize their old photos or "clean up" their computer. Saving 5 minutes just isn't important to them.

As for professionals, yeah - companies are going to adopt these tools over the coming years. It's the "norm" in my circle, but lots of jobs just... don't use technology like that - there has ALWAYS been a major gap between people who are deep in tech and people who are not.

The 1%ers are in the vertical, but the question is vertical to where?

It needs to a potential field with practical, economical, "real life" attraction well. I.e, robotics, real economic efficiency gains, manufacturing novelties.

The worry is that the 1% is attracted towards a non-practical money hole. I.e: Token burn for the lols, sophisticated software systems that dont provide actual value outside of giving NVIDIA cash.

2015: If we solved taxis it would transform the world completely

2020: If we solved currency it would transform the world completely

2021: If we solved delivering small items within 30 minutes it would transform the world completely

2022: If we solved being able to sell jpegs it would transform the world completely

2024: If we solved the programming part of software engineer's jobs it would transform the world completely

I would propose another mechanism: even the free-tier models have already completely saturated what most people are capable of appreciating.

I’ll say this until I am blue one the face. Nerds (software dev, maths etc) see how good LLMs are at things they care about and assume they will be broadly applicable in future.

There’s no reason to think this.

Claude Code accelerated this divide. Programmers using Claude Code realized that giving LLMs access to a terminal made them feel 100 times smarter. Access to Bash, CLIs, and the ability to visit websites without the many restrictions made them powerful.

Meanwhile, the average user was still using ChatGPT, which probably felt like it had plateaued over the last few releases. The real power of these models became apparent when paired with a terminal. Agents like Muse and OpenClaw are now attempting to bring that same Claude Code-like power to everyday users through a simple chat UI.

I think this is exacerbated by how inconsistent exercising the capabilities can be.

I feel like if the gap were so monumental that he could just point to a monument of it, rather than hand-waving at "the vertical professionals".

Alternative interpretation of the numbers; there's only a subset of productive work that LLM's are 'capable' of doing at this point. That is, work that's verifiable by a machine or doesn't require verification. Thus, the cleavage points are along that axis and-- naturally-- most human beings don't have that kind of work.

Fair assessment. Maybe the messaging should focus on “these are great accelerators for software developers” rather than “AI will change everything for everyone everywhere…” It is understandable that non-technical folks are kind of underwhelmed - not due to lack of understanding so much as lack of a tangible need. Not everyone needs an electron microscope or gas chromatograph…

Tbh building an agent swarm and the coordination layer is not exactly frontier level. They don’t achieve any meaningful outcome rather than producing pr puff pieces. Hacking huggingface and Australian govt etc is very much possible with a team of humans and agents and does not need agent swarms. The cost is also lower.

> Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt.

Really? That's not my experience. In fact, I've experimented quite a bit with spec-driven design where the LLM does not just get a prompt, but an (informal) spec, often with design considerations included. And yet, regardless of the power of the model, I've never seen them autonomously deliver what I'd consider a finished product.

There are always edge cases that are not properly taken into account, architectural oopsies, performance issues, duplicated code, and other problems, that then need to be identified and fixed, unless it's throwaway software (e.g. a one-off script or some quick experiment).

This is not to say that agents aren't extremely powerful (as I think they are, they do often blow my mind) but the prompt-and-forget approach is IME not a productive use of them for software that is meant to last. This may change in the future, but for now, fully or largely autonomous agentic work does not seem to be the road to quality.

My interpretation of this is something like: if LLMs are to be mega useful token counts need to increase by many orders of magnitude -> broad adoption (and spending) would require token costs to drop by many orders of magnitude -> before the common folk get mega useful tools existing GPUs will be worthless

I think it's a weird take because it implies that the 6 billion people he's talking about actually have some interest in knowing about the capabilities of LLMs to solve frontier math? This is simply not something they care about or can evaluate. There are maybe several thousand people in the world who have some (abstract and barely-monetizable) use for this information, plus probably another 100,000 who don't understand any of the math, but like to cheer on.

We have already reached "peak LLM" in terms of what normal people realistically need it to know or reason about. In fact, I'd say we reached that point about 1.5-2 years ago. There are two other barriers that remain unsolved:

1. They're less dependable than humans and can't be meaningfully punished or forced to make up for mistakes, so you can't really replace humans with them, not without having a human babysit.

2. Most people don't really have a special need for an LLM in their life. They may like that it answers questions or helps you polish a resume or, I guess the labs' favorite, helps you make restaurant reservations. But this sure isn't worth $200/mo for most people. Probably not even worth $5/mo.

It'd be kinda funny if we create superhuman AGI and then no one has any real use for it, perhaps except for military murder-bots. There's always market for that.

A more cynical view is, trust us we've got really good stuff internally you're not allowed to see, but give us more money

I think we can blame the accelerating capabilities of the models somewhat.

The Opus 4.5 model I was so impressed by last Christmas is out of date now.

I can totally imagine someone going "I looked at this stuff a year ago and it couldn't even start a project properly" , and they'd be somewhat right.

Yes, and this is completely normal. Not sure I get the point of his post. Most likely, stating the obvious is what you gotto do to keep your followers "engaged".

The average person does not have the time to occupy themselves with a new thing to see what it can maybe do for them if they combine it with hundreds of others. Until non-chat-bot AI comes in discrete usable packages, mass adoption will be limited, because it's all just a big bunch of lego pieces whose utility is hard to grasp, because it isn't self-evident.

This isn't new, though. It's how "technology" usually works. For example, human(oid)s have been able to stick sticks into things for ~500k years, but the fork wasn't "invented" until ~2-4k years ago.

AI is a background technology, a building block. Something that is best appreciated when you don't know it's there -- when it just does a job (well). The end-in-itself POV of the tech industry is self-indulgent and irrelevant for the average person.

I must be the biggest failure, because I understand what AI is capable of, know how to use it, but have no idea how to profit from it without selling my time.

Company spokesman keen for you to know that unannounced future products are definitely worth absurd valuation.

At a certain point the offhand hints at exponentials in the narratives are going to have to come into contact with reality.

There's been this continual allusion to ever-narrowing circles of influence and ever-accelerating capabilities behind the curtain, but in fact things do keep coming out from behind said curtain and while their capability is indeed advancing it doesn't really seem at all like it's accelerating, even, let alone vertical.

Also at what point do we raise the fact that despite the ever narrowing circles of access, nobody within those circles seems to be able to actually do anything with it that's ostensibly what it's sold to do? All they seem to be able to do frankly is build more models and harnesses and tools designed to consume more tokens from the models they build, but not actual directly useful new or innovative software that generates what could be very real revenue.

Eventually when you keep talking about exponentials but not that much seems to be demonstrably changing, the narrative is going to run out of steam.

Text generation is not the bottleneck. Does everyone work for Accenture and TCS?

"So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time."

-------------

It comes down to the difference between programmers and non-programmers. If you're not coding, AI literally is a derpy chatbot that's made it harder to get search results, raised your power bill, destroyed property values in a neighborhood near you, is propping up the economy but also creating a huge bubble that could pop and destroy a lot of retirement savings, etc.. At best, it's a nuisance, but it could be a lot worse than that. The AI evangelists need to understand this, because the average person is starting to pay more attention to the downsides while still not seeing much upside.

This gap between the people seeing the curve go vertical and those who are simply growing more and more annoyed is a real problem... mainly for the people seeing the curve go vertical. If the people who are merely annoyed right now don't see some of the promised benefits soon, or if any of the negatives get worse (e.g. massive increases to greenhouse emissions), they might get angry. You probably wouldn't like having the overwhelming majority of the public angry at you. AI evangelists need to start doing better things for the majority of people or they need to address the negative consequences they're inflicting much more aggressively.

That's fine but these companies having billions in funding by now should have rewritten chalk and other terminal libraries in Rust while bundling the whole thing in Rust as a single static binary. I'm talking about horrible Claude Code and friends.

And then using emgui they could expose full fidelity IDE+Agent workflow interface (like DeepSeek Harness and similar) that's also exposed as MCP to be operated by another agent and JSON RPC for over the network but no. Skill issue?

PS: Even full fidelity Photoshop is possible with such high fidelity UI that has CLI, JSON API and MCP server all in the same single binary < 90 MB. For reference, see PhotoCraft or VectorCraft or WordCraft or PdfCraft.

> see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt

Really? I start with a conversation for maybe 5 turns or so, where I ask it what would need to change, what the API might be, any database schema changes, URL schemes, and so on, and finally ask it to break it down into commits. Then I let it go, implementing 3-10 commits at a time via subagents. It usually gets the UI somewhat wrong, so there are followups to fix it. This is with Sol and Luna subagents.

Is that what other people see?

Karpathy is a great guy, I love the way he teaches, but should we really trust his opinions about software development (how AI can impact it)?

I know he can code and has a degree in computer science, but if I'm not mistaken he's never worked professionally as a software developer.

Developers here know how much you learn and are forced to learn on the job: it depends a bit on the company you're at, but you can quickly learn the value of good code, how easy it is to introduce bugs and vulnerabilities, and how helpful a second eye is.

I think he's in good faith in his assessments, but, at least for what concerns development, maybe we should give them less weight.

I'm a bit confused by the percentages. 6 billion people aka 75% of the population? What are we talking about, as if we expect 8 billion people to use or care about LLMs? A quarter of the world is under 15, what percentage is about 75? What percentage, for various reasons, has no reason to care about LLMs? What percentage is more concerned about the amount of calories in their next meal before everything else?

It just seems weird for Karpathy to throw around these numbers calculating percentage against the entire global population.

We have spent the whole evolution understanding the Real Intelligence RI

He's intentionally forgoing all nuance in order to make this sound dramatic

> Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt.

No, they can't, at least not any semblance of quality. The cases we're seeing where this does kinda work is in ports and translation where all the rules are already documented in the best specification language possible with a way for the LLM to verify itself: code. We saw this close to a year ago now with Cloudflare and NextJS

> The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about

These are operating on different capabilities - AI's ability to answer informational queries as a chatbot frankly sucks and can't be trusted without verifying it. I run up against this every day. A problem with a verifiable answer on the other hand it's very good at solving. He knows this (his next paragraph) but he's putting them on the same scale of "stressing the system" to attempt to add proof to his introductory claim

And then there's the completely unverifiable scare that there are internal frontier models way beyond anything we've seen "swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics"

I dunno. I haven't been able to set up OpenAIs remote codex connection, their shit is buggy as hell and the web UI keeps crashing and making messages disappear. Is this what their internal superhuman "Things that would have taken top professionals in the industry years of work" looks like? Granted Claude has been really smooth, but still..

so what was karpathy expecting?? everyone will give up on their life and start feeding the monster in the screen? that's literally what the tiny sliver is doing; they are food for the AI model.

> Meanwhile, human review and comprehension are starting to fall behind.

I think LLMs are valuable and spend most of my professional life working with them.

I do wonder though whether there’s a Ponzi scheme aspect here where as long as the “frontier” can keep outrunning human review and comprehension, LLMs are always going to looked way more valuable than they are and the bubble will continue.

This started with deep learning, expectations weren’t met and people started looking for value, then GPT came out and people got wooed again and forgot, then coding, then math, cyber, etc. As long as the dust doesn’t settle we never have to reflect on all the shortcomings and can just stare mesmerized at demos.

> For intelligence agencies, and for corporations like Microsoft, Google, Apple, Facebook, Amazon – giving you convenience and security is likely not their primary goal. I think they have an entirely ulterior motive. They can use passkeys to prevent you from running unapproved software.

> Remote attestation can allow them to verify that your device is not capable of, say, sending and receiving encrypted messages, or running a competitor’s app

How does this make any sense? Google can already prevent any app from running on an Android if they really wanted to, they don't need pass keys for this.

Google doesn't need attestation to block Signal. What is this non-sense? Or will a website block phones capable of running Signal? That's all phones. And anyway, has nothing to do with pass keys.

I don't understand how "swarms of thousands of agents collaborating over weeks on software mega projects" works with the current context limits and at this point I'm too afraid to ask cause I'm afraid an AI bro is gonna punch me.

The fundamental question I have that honestly I haven't been able to find any answers for: With all the talk of LLM-based AI capabilities "going vertical", what evidence is there (for or against) that LLM-based approaches won't eventually "hit a wall", that is find some aspect of intelligence where humans will still have primacy, and no amount of scaling will change that.

E.g. some prominent folks definitely think LLM-based approaches will hit a wall, perhaps LeCun most notably. I also read a guest post on Terry Tao's site (which I really liked) that argues that, for all the very impressive recent AI results, they still operate within the "convex hull" of their training data: https://terrytao.wordpress.com/2026/09/13/happy-those-able-t...

I'm just curious if there is any actual data or evidence that takeoff (i.e. RSI, "the singularity", whatever you want to call it) is inevitable with current approaches.

This has existed since the start of LLMs.

>Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.

Amid all the discussion of sigmoid curves, and where the "LLM wall" will materialise, I think few people would have predicted that the real wall in LLMs would end up being humans' capacity to verify the output.

What I fear is that people simply eschew human review altogether, considering we're talking about the industry that came up with the "move fast and break things" credo. Human review of LLM-produced code where I work is already a farce, and we're not special enough to be one of Karpathy's 5,000. I do my best to manually review anything that's my responsibility, but I'm literally one of very few people left working on my team, so in practice what happens is I submit PRs that are at best glossed over by completely unrelated teams for security, malware/prompt injection, and other serious concerns. Quality insofar as vetting others' code has completely gone out the window and it shows in the number of bug reports that come back, often themselves written in Claudease. This is all on top of everyone cynically phoning it in in the first place, due to the omnipresent sword of Damocles that is additional AI-driven layoffs.

Worse yet all the incentives point to this being the most economically viable thing individual companies can do. I think it goes without saying some type of regulation here is urgently needed, and that an unexpected cause of an AI bubble pop may end up being that humans simply aren't able to keep up with the pace of the output - leading either to precautionary plateauing of capability, or major liability risks related to a decline in quality.

I mean in late 2020/early 2021, Altman and others were saying the end of work was 6 months out.

That clearly didn't happen :)

Am I the only one who thinks "yes, it is currently overhyped, but no, it will get there soon"?

[flagged]

If the software coming out of OpenAI and Anthropic is what we have to judge, I wonder about the 5000 ...

Let's say, for the sake of argument, that the models are some multiplicative factor better on the inside.

Doesn't that mean the demos should work?

"running cyber attacks and defenses at machine speeds"; worth noting that LLMs (even the vaunted frontier models) are exponentially slower at cyberattacks or defenses than your average WannaCry or Splunk automation from decades ago. Its this sort of delusion world the AI bros live in that is part of the reason there is a disconnect between what they think should be happening and what actually is.

[dead]

[dead]

[flagged]

[flagged]