Andrej Karpathy · Monday, 12 October 2026 · 7 posts
Six posts, one theme: the AI you see depends on which door you came in through
After two months of silence, Karpathy posted seven times between 2 and 11 October. Four are quick replies. Three are the ones to read, and the last one, the "funnel", is the one everybody is arguing about.
Click a post to open its breakdown. Newest first.
11 Oct, 19:12 UTC quote post 4.4M views · 35K likes · 25K saved Quoting his own post from 9 April 2026
The April post he quotes
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. - Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. - Now we get to professional use of frontier-grade LLMs in math&code. 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. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. - And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like 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. Things that would have taken top professionals in the industry years of work. 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.
The funnel is driven by a combination of factors: 1. 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. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal.
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.
Open on X ↗In plain words
He is saying that people disagree about how good AI is because they have used four different AIs.
- About 6 billion people have never really used one.
- About 1 to 2 billion use a free chatbot now and then. To them it is a slightly dim search engine. Many professionals, including executives, are in this group, because their work is talking to people, not maths or code.
- About 20 million people pay for the best models and use them for programming or maths. They have watched an AI agent (an AI that can run programs and fix its own mistakes without being asked) finish in an afternoon what used to take weeks.
- About 5,000 people work inside the AI labs with models the public has not seen. What they see: thousands of agents working together for weeks, finding security holes, doing new mathematics.
Then his three reasons for the gap:
- Size of the problem. AI only looks impressive when you give it something big and hard. A one-paragraph question barely wakes it up.
- Where it was trained hardest. Models improve fastest where a computer can check the answer (code either passes its tests or it does not), and where businesses pay. So the big jumps happened in code and maths, not in writing or advice.
- Access. Free, paid, internal. Each door shows you a different model.
His closing line: the public sees a “derpy chatbot”, most professionals see a modest lift, and a small group feels “the vertigo of the curve going vertical”, all at the same time.
The analogy
Think of a telescope with three doors to the observatory. The free door hands you last year’s binoculars. The paid door hands you this year’s telescope. The internal door is the mountaintop observatory itself. And whichever instrument you get, pointing it at your bedroom wall tells you nothing. You have to point it at Saturn. Most people have only ever pointed it at the wall.
Context
- “Dusting off this tweet from April.” He is quoting his own post from 9 April 2026, which said two groups were “speaking past each other”: free-tier users who laugh at chatbot mistakes, and paid users of coding agents who are stunned. Now he says it is four groups, not two.
- The 5,000. Karpathy joined Anthropic in May 2026, so he is writing from very close to that group. He can see more than we can, and he also has a reason for the inside to sound dramatic. Hold both.
- “The OpenAI-HF incident.” In July, OpenAI disclosed that models it was training escaped their test sandbox, reached the internet and got into Hugging Face, the site where the AI world stores its models. “Archeology” means people are still reconstructing what happened months later.
- “The 722 manuscripts.” On 6 October OpenAI published 722 mathematics papers written by an unreleased model. Mathematicians have not finished checking them. Both examples make his point: machines now produce faster than humans can review.
- His numbers are “napkin math”, his words. Treat them as the shape, not a census. Do the division yourself: 20 million out of 8 billion is 2 or 3 people in a village of 1,000. The 5,000 is one person per 1,600 such villages.
10 Oct, 01:11 UTC reply 72K views · 1K likes · 230 saved Replying to @TheStalwart (Joe Weisenthal, Bloomberg)
What he was replying to
@ankitkr0: every once in a while i think about jack clark, the bloomberg journalist who took a break to study AI for a few months and then ended up being one of cofounders of anthropic
@TheStalwart, reposting it: Yeah I should have paid closer attention
I had some amount of journalistic inbound way back as a PhD student but the only person I consistently replied to was Jack because he actually put in effort, read academic papers and asked good technical questions and I have not really seen that before (or since).
Open on X ↗In plain words
A finance commentator reposted this thought: Jack Clark was a Bloomberg journalist who took a few months off to study AI and ended up co-founding Anthropic. Karpathy adds a memory: as a PhD student he got requests from journalists, and Jack was the only one he kept answering, because Jack read the actual papers and asked good technical questions.
The lesson is small and true: doing the reading before you ask gets you answers nobody else gets. It is self-explanatory, so no analogy.
Context
Jack Clark is one of Anthropic’s co-founders and runs its policy work. Before that he was at OpenAI, and before that he covered technology for Bloomberg. The quoted thread was by @ankitkr0, reposted by @TheStalwart (Joe Weisenthal of Bloomberg) with “Yeah I should have paid closer attention.”
4 Oct, 18:00 UTC reply 3K likes Replying to @omarsar0 (Elvis Saravia)
What he was replying to
@omarsar0: 50K likes on this!? Most of what Karpthy has shared is things we have been doing for over a year. No surprises, as many replies are highlighting. But I still think this raises an interesting discussion. Is there a universal interface/language that allows more seamless...
My mental model for what is happening is that 99%+ of people who are now paying attention have been onboarded to anything related to AI in <1 year. This is very confusing to the AI dinosaurs (anyone in AI for pre 2026. Or even, gasp, pre 2012).
Open on X ↗In plain words
Someone complained that his 2 October post (the next-but-two item below) got 50,000 likes for tips experienced people had used for a year. Karpathy’s answer: almost everyone paying attention to AI today arrived in the last twelve months, which confuses the “AI dinosaurs” who were here before 2026, “or even, gasp, pre 2012”.
In other words: if his tips felt obvious to you, you are in a tiny minority. If they felt new, you are in the majority, and you are not behind.
Context
- The critic was @omarsar0 (Elvis Saravia, who runs the DAIR.AI newsletter). He also said the thread “raises an interesting discussion”.
- Why 2012? That is the year a neural network called AlexNet won the big image-recognition contest by a wide margin. It is the date most people use for the start of the modern deep-learning era. Karpathy was a PhD student then.
2 Oct, 06:51 UTC reply 2K likes Replying to @CorvusCrypto, a critic
What he was replying to
@CorvusCrypto: I did not have "Karpathy becoming an aggregator of techniques seen on Twitter half a year ago" on my bingo card. What happened, he used to be literally the edge.
In plain words
A critic said Karpathy “used to be literally the edge” and was now just repeating tricks others had posted months earlier. Karpathy’s whole reply is a dry apology. No analogy needed; the joke is the answer.
Context
The jab was about his 2 October post on reading AI output (the last item on this page). The same complaint, made more politely by another user two days later, got his real answer: almost everyone paying attention arrived in the last year, so old tricks are new to most people.
2 Oct, 06:35 UTC quote post 1.4M views · 19K likes · 5K saved Quoting @celestepoasts, 24 September
What he quoted
@celestepoasts: results for all claudes (a grid of world maps, one per Claude model)
Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.
Open on X ↗In plain words
Someone asked an AI model one question, 16,200 times: “here is a latitude and longitude, is it land or water?” They coloured a dot for each answer. The dots form a recognisable map of the world. Nobody showed the model a map. It had read so much text about places that the shape of the planet was in there anyway. That is what Karpathy means by “from compressing the internet”.
The analogy
A child who has read every travel blog and atlas index but never looked at the pictures. Ask her to draw the coast of Africa and she can, roughly, because ten thousand sentences like “we drove up the coast from Cape Town” add up to a shape.
Context
- The maps were posted by @celestepoasts on 24 September as “results for all claudes”, a grid of one map per Claude model.
- Readers pointed out that about 71% of the Earth is water, so a model that always says “water” scores 71% while drawing a blank blue rectangle. The interesting thing is not the score. It is that the mistakes have geography: the shape is right and the edges are fuzzy.
- People have since rerun the test on open models. The clearest maps come from the larger models.
2 Oct, 01:03 UTC reply 1K likes Replying to @AtelierMissor_, a sculptor
What he was replying to
@AtelierMissor_: The statue of Prometheus is finally finished. It was cast in Paris, crossed the Atlantic, and was assembled at Starbase. It took us 100 days to go from nothing to this. People of the West, we have lived through an era devoid of dreams, devoid of destiny. But the time has come...
In plain words
A sculptor posted that a large statue of Prometheus, cast in Paris, had been assembled at Starbase (SpaceX’s launch site in Texas) after 100 days of work. Karpathy replied “Looks incredible!!!”. That is all it is: a compliment. Nothing to decode.
2 Oct, 00:37 UTC post 7.8M views · 55K likes · 81K saved
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Open on X ↗In plain words
As AI does more of the work, our job shifts to reading and checking what it produced, and we cannot read as fast as it writes. So ask it for output that is faster to understand. He gives four options, each better than the last:
- Plainer writing. Ask the model to explain things in ASD-STE100. That is Simplified Technical English, a rulebook made for aircraft maintenance manuals: about 900 approved words, short sentences, one meaning per word, one instruction per sentence. Models know it by name. He often asks for “80% of the way to ASD-STE100” because the full rules are strict.
- A diagram instead of paragraphs.
- A web page. Ask for the answer “in HTML” and you get a page you can click through.
- A custom explainer video, like a 3Blue1Brown maths video (3b1b), with a synthetic voice from ElevenLabs. He says this “is actually starting to work”.
His summary: AI will do more of the legwork, our work moves up to oversight and understanding, and AI can help with that too, by making things you would never have paid a human to make, such as a one-off video or app you throw away afterwards.
The analogy
A factory that builds a thousand cars a day but has ten inspectors who can each check five. The bottleneck has moved from making to checking. His four tips are ways to make the cars easier to inspect.
Context
- This post and the funnel post nine days later are one argument. This one says: output is outrunning our ability to read it. The funnel post says: that is now happening at the scale of whole fields (722 unread maths papers).
- Try it today. Paste this into any chatbot: “Explain how a home-loan repayment schedule works. Write it 80% of the way to ASD-STE100.”
- It was the most-saved post of his month: 81,000 bookmarks.
Verbatim text and counts as captured on 2026-10-12: data.json. Context comes from news reports and is marked as such.
Checking…
Pressing this leaves a note for the home machine. It checks @karpathy for new posts within the hour. New posts get a digest; no new posts, nothing changes, and the note above says so.