Isn't that the point of vibe coding? You don't even look at the code. Just trust the llm to take the wheel. https://x.com/karpathy/status/1886192184808149383?lang=en
Account
@karpathy
- Ideas
- 19
- Recommendations
- 41
- HN citations
- 148
AI16Tools1Internet1Learning1
View @karpathy on XI don't think there is evidence in either direction (takeoff impossible vs inevitable). But the "convex hull" idea always came across as wishful thinking and hand-wavy pseudo-science to me. Here's a rebuttal [0] of it from Tim Gowers (uses another math metaphor, but still clear and the whole thread is worth a read): >...the popularity of the idea that AI "merely explores the convex hull" mystifies me. Why wouldn't it be more like the subgroup generated by existing mathematical knowledge? (That's also a metaphor -- I'm not saying there's a group structure.) >The convex hull metaphor suggests that AI isn't having radically new ideas and will eventually run out of things to do, whereas the group metaphor suggests the possibility of building up in small steps to something well beyond what you started with, with no obvious limit. [0] https://twitter.com/wtgowers/status/2108090920864268679
Relatedly, Karpathy had an insightful post awhile back on perils of tokenization https://twitter.com/karpathy/status/1657949234535211009 (I cited it in my Excel tutorial on BPE https://www.youtube.com/watch?v=PvZN3-WqAOI&t=1916s )
op - https://x.com/karpathy/status/1915581920022585597 cool tips there
At this point, anyone claiming that LLMs are "just" language models aren't arguing in good faith. LLMs are a general purpose computing paradigm. LLMs are circuit builders, the converged parameters define pathways through the architecture that pick out specific programs. Or as Karpathy puts it, LLMs are a differentiable computer[1]. Training LLMs discovers programs that well reproduce the input sequence. Tokens can represent anything, not just words. Roughly the same architecture can generate passable images, music, or even video. [1] https://x.com/karpathy/status/1582807367988654081
Karpathy offers a more concise (and whimsical) explanation https://x.com/karpathy/status/1733299213503787018
For those who are interested in how Transformers seem more prevalent, please read this thread by Karpathy where he talks about a consolidation in ML: https://twitter.com/karpathy/status/1468370605229547522 And of course one of the early classic papers in the field, as a bonus: https://papers.nips.cc/paper/2017/file/3f5ee243547dee91fbd05... (The paper is mentioned in the article)
Back when I was looking for a new PKM system, I remember a "competitor" (now irrelevant) of obsidian making fun via Quote tweet of the Files over app manifesto, saying something like "nobody cares about the way the notes are stored". Ironically, that mean-spirited tweet made me discover Obsidian, and I've adopted it since then as I strongly agree with the philosophy of sovereignty over the content by them being plain files. I do care about how files are stored, felt trapped with OneNote proprietary format, and seeing what happens to Evernote, now care even more. More recently, Andrej Karparthy succinctly captured what's great about Obsidian: https://x.com/karpathy/status/1761467904737067456 And with the new LLM powered, I find having tight control over files to be extremely powerful Thank you kepano!
Thanks for the comment. The point I am trying to make is similar to what Karpathy says in this tweet - https://twitter.com/karpathy/status/1617979122625712128 - but applied more broadly than just programming. I don't think this is an exaggeration or overselling.
https://twitter.com/karpathy/status/1435827240286109702?s=46... there’s a good example of 25 seconds of constant bombardment for InfoWorld visitors. I cant count the popups.
Something from Andrej Karpathy on learning that stuck with me [0]: > Learning is not supposed to be fun. It doesn't have to be actively not fun either, but the primary feeling should be that of effort. It should look a lot less like that "10 minute full body" workout from your local digital media creator and a lot more like a serious session at the gym. You want the mental equivalent of sweating. It's not that the quickie doesn't do anything, it's just that it is wildly suboptimal if you actually care to learn. [0] https://x.com/karpathy/status/1756380066580455557?lang=en
Interestingly this point was indicated by Karpathy last summer that RLHF is barely RL. He said it would be very difficult to apply pure reinforcement learning on open-domains. This is why RLHF are a shortcut to fill this gap but still because the reward model is trained on human vibes checks the LLM could easily game the RM by giving out misleading responses or gaming the system. Importantly the barrier is that open domains are too complex and too undefined to have a clear reward function. But if someone cracks that — meaning they create a way for AI to self-optimize in these messy, subjective spaces — it'll completely revolutionize LLMs through pure RL. Here's the link of the tweet: https://x.com/karpathy/status/1821277264996352246
The thing is, it's not gibberish. A sufficiently small language model can be understood by humans: https://twitter.com/karpathy/status/1645115622517542913 The explanation is perfectly sensical, just too complex for humans to understand as the model scales up. The thing you're looking for - a reductive explanation of the weights of a ANN that's easy to fit in your head, does not exist. If it were simple enough to satisfy your demands, it wouldn't work at all.
Andrej Karpathy made an interesting comment on the same paper: https://x.com/karpathy/status/1980397031542989305
He's building Eureka Labs[1], an AI-first education company (can't wait to use it). He's both a strong researcher[2] and an unusually gifted technical communicator. His recent videos[3] are excellent educational material. More broadly though: someone with his track record sharing firsthand observations about agentic coding shouldn't need to justify it by listing current projects. The observations either hold up or they don't. [1] https://x.com/EurekaLabsAI [2] PhD in DL, early OpenAI, founding head of AI at Tesla [3] https://www.youtube.com/@AndrejKarpathy/videos
Will Smith spaghetti was garbage a couple of years ago and now AI videos are becoming close to indistinguishable from real videos in many cases. Spaghetti 2026: https://x.com/dreamingtulpa/status/2083304533829066873 https://xcancel.com/dreamingtulpa/status/2083304533829066873
The idea that LLMs are "running out of training data" is pretty widespread, but does it actually hold up? Has anyone seen representatives of the cutting-edge AI labs - OpenAI, Anthropic, Meta, Mistral etc - express concern about this? The impression I've been getting is that quality turns out to be much more important than quality. This tweet from Andrej Karpathy for example: https://twitter.com/karpathy/status/1797313173449764933 > Turns out that LLMs learn a lot better and faster from educational content as well. This is partly because the average Common Crawl article (internet pages) is not of very high value and distracts the training, packing in too much irrelevant information. The average webpage on the internet is so random and terrible it's not even clear how prior LLMs learn anything at all. You'd think it's random articles but it's not, it's weird data dumps, ad spam and SEO, terabytes of stock ticker updates, etc. And then there are diamonds mixed in there, the challenge is pick them out.
That's called sleeper agent problem, and is extremely actual (and I don't think solvable): https://x.com/karpathy/status/1745921205020799433?s=46&t=Hpf...
He addressed this well in this post: https://x.com/karpathy/status/1778153659106533806?s=46&t=QIb...