vatsachak

6 ms·

1. 181.▲
   by vatsachak 3mo ago

   Vibe coded
2. 182.▲
   by vatsachak 3mo ago

   Something like Windows can't even keep it's cool on my lunar lake 32 GB laptop while NixOS with Niri works great on an 8GB guy
3. 183.▲
   by vatsachak 3mo ago

   Tesla is a trillion dollar company because it's a meme stock. It has a 300 P/E
4. 184.▲
   by vatsachak 3mo ago

   Yeah. But this kinda feels like a bandaid. Eventually someone will have to solve compute in memory at scale.
5. 185.▲
   by vatsachak 3mo ago

   Okay then why do software engineers still exist. And if companies are making an error keeping people on instead of LLMs, why isn't there an all agent company making money off of it?
6. 186.▲
   by vatsachak 3mo ago

   Okay here's something LLMs can't. They can't solve problems that are longer than ~10 pages of math. They also can't maintain codebases without supervision. It's because they are have no memory and use various tricks
7. 187.▲
   by vatsachak 3mo ago

   Because pattern recognition and deduction is automatable and LLMs are superhuman at low depth high depth problems. Next
8. 188.▲
   by vatsachak 3mo ago

   The LLM knows how to solve problems in any way you want. This is not a good thing
9. 189.▲
   by vatsachak 3mo ago

   What makes you think that they will get good at driving themselves? Can they now train on their own outputs? Doesn't seem like it
10. 190.▲
    by vatsachak 3mo ago

    Well the difficult part is finding a good intermediate representation, which they did
11. 191.▲
    by vatsachak 3mo ago

    And everyone who's been into this stuff for a while has had their prediction come through. If AI is water, Lean is the pipe and collatz is a clog on one end, then surely we'll find the cracks.
12. 192.▲
    by vatsachak 3mo ago

    Awesome! Confirms what we know; LLMs are superhuman at short term reasoning and breadth
13. 193.▲
    by vatsachak 3mo ago

    I think that a gpt 5.6 level model might sustain the growth. We're seeing talented users squeeze non trivial math and code out these tiers of models.
14. 194.▲
    by vatsachak 3mo ago

    Hi, used to be actual mathematician here (lecturer, publication in JEMS and all that) For starters he's talking about how peer review is bad because of politics. That's fair. But no one is saying that's the case with AI math.
15. 195.▲
    by vatsachak 3mo ago

    I need to make the same post when someone finally RLs an LLM to get really good at performance engineering. Many "serious" performance engineers are aghast "Nooo the r00b0t can't make llvm faster and not tell us how!!! i
16. 196.▲
    by vatsachak 3mo ago

    Have you tried giving them the types you want and then have them auto complete? They are pretty good at a auto completing abstractions
17. 197.▲
    by vatsachak 3mo ago

    Video models can certainly predict what stirring cream into coffee does. Most LLMs these days are limited by context, and within context they are super human at reasoning tasks/per unit time. I agree with the author in one sense in tha
18. 198.▲
    by vatsachak 3mo ago

    You're right, forgot a factor of 10 whoops
19. 199.▲
    by vatsachak 3mo ago

    Worse than 0.01% of humans means that there are 8,000,000 people better than it. I know that's being pedantic I understand what you're saying. But every time I use Codex unless I specifically give it the abstractions it writes cod
20. 200.▲
    by vatsachak 3mo ago

    Yeah computer proof writing involves choosing good abstractions at every turn. LLMs aren't great at that yet
21. 201.▲
    by vatsachak 3mo ago

    I doubt anything other than an LLM auto complete could add something like a tensor reshape
22. 202.▲
    by vatsachak 3mo ago

    Yep! I do use Codex CLI
23. 203.▲
    by vatsachak 3mo ago

    I disagree. I speak in code to the LLMs. It's just that LLMs are really good at reinventing the wheel that you were supposed to in your codebase. Recent example. struct TensorView<T>{ body: Arc<[T]>, shape: [usize], stri
24. 204.▲
    by vatsachak 3mo ago

    I hate saying this but unironically opus 5/gpt 5.6 are writing maintainable code on the order of 500-1000 lines
25. 205.▲
    by vatsachak 3mo ago

    Fair
26. 206.▲
    by vatsachak 3mo ago

    Yeah copyright is kinda just like "if we say it's so then it's so"
27. 207.▲
    by vatsachak 3mo ago

    I mean movies are up on YouTube slightly pitch shifted lol.
28. 208.▲
    by vatsachak 3mo ago

    Well maybe they should find better training methods or architectures
29. 209.▲
    by vatsachak 3mo ago

    This is all a moot point because there's no way you can copyright a model because you can just randomly perturb weights and still be fine.
30. 210.▲
    by vatsachak 3mo ago

    That's why competition is good though