Andrej Karpathy

I did that and see karpathy has a bearblog https://karpathy.bearblog.dev/ wow ref: https://news.ycombinator.com/item?id=46095250

Thank you. It was a great one-shot and I didn't end up doing any updates. Thrilled to see how it inspired work from: Thomas Morford (CSE @ UC Merced, thomasm6m6) who did the amazing article/thread generation (in < 100 lines of PY!): https://sw.vtom.net/hn35/news.html ; and also from Andrej Karpathy (ex-OpenAI, now Eureak Labs, karpathy) who did an interesting analysis of prescience quality of threads/commenters inspired by a reply linking the page of 10 years past to compare: https://karpathy.bearblog.dev/auto-grade-hn/ This was wonderful. 3000 points? I mean, fuck. Among the biggest posts of all time, and definitely of Show HN. Funny for me is that all the work I've done in the last 10 years, probably 100 Show HN's all different, this was by far the hugest. Could be months of work, no interest. And this thing, which dropped into my mind, probably 30 minutes, demolished them all. It's hilarious that it even beat out legitimate AI posts, and contaminated search results with future stories. One of the funniest things for me was hearing how people tabbed away from the page, only to come back and momentarily feel it was the actual HN page. Hahahahaha! :) All I can say is, I love you all. Watching it stay at the top for 24 hours...it felt like it wasn't something I made at first. But it was. Cool

I think there's an opportunity right now to reinvent public forum moderation with LLMs. Karpathy's recent post is going in the right direction I think: https://karpathy.bearblog.dev/auto-grade-hn/

The idea of HN being dismissive of impactful technology is as old as HN. And indeed, the crowd often appears stuck in the past with hindsight. That said, HN discussions aren't homogeneous, and as demonstrated by Karpathy in his recent blogpost "Auto-grading decade-old Hacker News", at least some commenters have impressive foresight: https://karpathy.bearblog.dev/auto-grade-hn/

This caused me to delete the HN password on my main account. It was already disconcerting to be reminded by Karpathy (I cannot help but see this as bragging) in his recent post that we must be good because LLMs are watching: https://karpathy.bearblog.dev/auto-grade-hn/ LLMs will be used to deanonymize all internet users. The AI "visionaries" are already bragging about this. You have been warned.

Karpathy himself has used "vibe coding" to describe "usage of LLMs for coding," so it's fair to say the definition has expanded. https://karpathy.bearblog.dev/year-in-review-2025/

Here's the source for the jagged spiky intelligence diagram: https://x.com/colin_fraser/status/1994235521812328695 https://karpathy.bearblog.dev/the-space-of-minds/

I like "ghosts" as a simple metaphor for what you chat with when you chat with AI. Usually we chat with Casper the Friendly Ghost, but there are a lot of other ghosts that can be conjured up. Some people are obsessed with chatting with ghosts. It seems like a rational adult couldn't be seriously harmed by chatting with a ghost, but there are news reports showing that some people get possessed. It's a better metaphor than parrots, anyway. more: https://karpathy.bearblog.dev/animals-vs-ghosts/

Inspired by Andrej Karpathy's project: https://karpathy.bearblog.dev/auto-grade-hn/

The blog post will do: https://karpathy.bearblog.dev/auto-grade-hn/ (https://news.ycombinator.com/item?id=46220540)

What is this "clerk" library he used at this timestamp to tell him what to do? https://youtu.be/LCEmiRjPEtQ?si=XaC-oOMUxXp0DRU0&t=1991 Gemini found it via screenshot or context: https://clerk.com/ This is what he used for login on MenuGen: https://karpathy.bearblog.dev/vibe-coding-menugen/

By Karpathy: https://karpathy.bearblog.dev/the-append-and-review-note/

See also https://karpathy.bearblog.dev/the-append-and-review-note/

The essence of it is that after the "read the whole internet and predict the next token" pre-training step (and the chat fine-tuning), SotA LLMs now have a training step where they solve huge numbers of tasks that have verifiable answers (especially programming and math). The model therefore gets the very broad general knowledge and natural language abilities from pre-training and gets good at solving actual problems (problems that can't be bullshitted or hallucinated through because they have some verifiable right answer) from the RL step. In ways that still aren't really understood, it develops internal models of mathematics and coding that allow it to generalize to solve things it hasn't seen before. That is why LLMs got so much better at coding in 2025; the success of tools like Claude Code (to pick just one example) is built upon it. Of course, the LLMs still have a lot of limitations (the internal models are not perfect and aren't like how humans think at all), but RL has taken us pretty far. Unfortunately the really interesting details of this are mostly secret sauce stuff locked up inside the big AI labs. But there are still people who know far more than I do who do post about it, e.g. Andrej Karpathy discusses RL a bit in his 2025 LLMs Year in Review: https://karpathy.bearblog.dev/year-in-review-2025/

He has a new blog at https://karpathy.bearblog.dev/blog/