Viewing profile — vatsachak

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

Tesla is a trillion dollar company because it's a meme stock. It has a 300 P/E

Yeah. But this kinda feels like a bandaid. Eventually someone will have to solve compute in memory at scale.

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…

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 ha…

Because pattern recognition and deduction is automatable and LLMs are superhuman at low depth high depth problems. Next

The LLM knows how to solve problems in any way you want. This is not a good thing

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

Well the difficult part is finding a good intermediate representation, which they did

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…

Awesome! Confirms what we know; LLMs are superhuman at short term reasoning and breadth

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.

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 n…

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 m…

Have you tried giving them the types you want and then have them auto complete? They are pretty good at a auto completing abstractions

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…

You're right, forgot a factor of 10 whoops

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 s…

Yeah computer proof writing involves choosing good abstractions at every turn. LLMs aren't great at that yet

I doubt anything other than an LLM auto complete could add something like a tensor reshape

Yep! I do use Codex CLI

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 …

I hate saying this but unironically opus 5/gpt 5.6 are writing maintainable code on the order of 500-1000 lines

Fair