vatsachak

8 ms·

1. 271.▲
   by vatsachak 4mo ago

   Even if you had to submit a picture of your driver's license, you can send someone else's
2. 272.▲
   by vatsachak 4mo ago

   The other day some sucker told me, "don't throw that trash over there it's littering" and I told that sucker "there's no way anyone could enforce it" lol. Some people get up in their jimmies about this but
3. 273.▲
   by vatsachak 4mo ago

   Yeah it's a hype slop piece
4. 274.▲
   by vatsachak 4mo ago

   JEPA in chess leads to interpretable chess boards: https://arxiv.org/abs/2606.11860 JEPA in image classification leads to interpretable image latents https://arxiv.org/abs/2508.10104 Easy intro to
5. 275.▲
   by vatsachak 4mo ago

   Look at VLM mechanistic interpretability papers vs just pca on JEPA trained weights. JEPA gives you interpretability for free. I have not personally inspected them and my view is maybe a more exaggerated/dramatic claim of those working
6. 276.▲
   by vatsachak 4mo ago

   All of these LLMs are getting better at being at an LLM But GPT-5.5 is as useful an LLM can be; it has solved lemmas I've thought about for a year, it can implement typed STLCs in Rust when I give it a formal grammar, it can help me an
7. 277.▲
   by vatsachak 4mo ago

   An anonymous record type in a language whose type system is an enum LangType in Rust is just HashMap<String, Box<LangType>> My pet project requires an STLC with anonymous record types and type inference for anonymous records was
8. 278.▲
   by vatsachak 4mo ago

   Isn't this trivial? What's more interesting is as to why double descent happens
9. 279.▲
   by vatsachak 4mo ago

   This is why you still work on open source and do coursework for fun. The same reason you go for a jog
10. 280.▲
    by vatsachak 4mo ago

    Academia is stuck in the past in a lot of ways.
11. 281.▲
    by vatsachak 4mo ago

    Because it turns pdf generation into a programming language, hence you can automatically generate invoices with little hassle.
12. 282.▲
    by vatsachak 4mo ago

    Typst killed the invoice industry
13. 283.▲
    by vatsachak 4mo ago

    Beautiful software. Thanks!
14. 284.▲
    by vatsachak 5mo ago

    What stops Anthropic of American energy and infrastructure costs. And those AI ASIC companies just got bought by OpenAI and you guessed it... NVIDIA
15. 285.▲
    by vatsachak 5mo ago

    Yeah. Models haven't really improved much from last year to this year. I really love LLMs for debugging and rubber ducking, but I kinda want to write all my code. LLMs tend to have a hard time understanding composition.
16. 286.▲
    by vatsachak 5mo ago

    Alright, let me explain what's happening this Q Chinese providers realized that LLMs have peaked and have started trying to reduce the price per token. Deepseek pro v4 can easily add tests to my complicated code and costs cents for a m
17. 287.▲
    by vatsachak 5mo ago

    And if a pig had wings it could fly
18. 288.▲
    by vatsachak 5mo ago

    Deepseek is realistic about this and is offering great prices. I think that once companies are more realistic about token demand they can start making a profit
19. 289.▲
    by vatsachak 5mo ago

    I'm using both csv and aho-corasick on my project! I wish you the best and I'm sure us Rustaceans are happy to help with anything
20. 290.▲
    by vatsachak 5mo ago

    Docker moment
21. 291.▲
    by vatsachak 5mo ago

    Postgres doesn't cost any extra lol
22. 292.▲
    by vatsachak 5mo ago

    Sorry, I meant the code in the repo
23. 293.▲
    by vatsachak 5mo ago

    Looking at the commenting pattern, it seems like AI unfortunately
24. 294.▲
    by vatsachak 5mo ago

    I agree with this take. But this take also means that actual productive token use is not as high as people currently make it out to be. AI is an excellent rubber duck and test writer. Maybe I sniff my farts too much but I like my code just
25. 295.▲
    by vatsachak 5mo ago

    That's true. I should clarify by saying I meant that a human playing on par with Kasparov obviously has the planning ability to manage a McDonald's
26. 296.▲
    by vatsachak 5mo ago

    It might just be that we didn't have enough compute till now. StockFish definitely has superior intuition
27. 297.▲
    by vatsachak 5mo ago

    They are most likely using them in training. I doubt their IMO team are show ponies
28. 298.▲
    by vatsachak 5mo ago

    They most definitely threw in rl with formal verification somewhere between GPT 4 and now. The models are better at not hallucinating. I don't think their IMO team are only show ponies...
29. 299.▲
    by vatsachak 5mo ago

    My claim is that we haven't even witnessed the move 37 of math yet. I am claiming that math AI is going to get even better
30. 300.▲
    by vatsachak 5mo ago

    Fair. Let's see in a year. I'm willing to bet that nothing happens.