difficultyang

@kalomaze @celestepoasts "Just solve it with scaffolding"

@celestepoasts there is a genre of benchmarks that tests for problems that are "difficult" in extremely shallow ways. sometimes bad abstractions are just bad abstractions

I don't like the implication that models suck at esolangs just because of limited training data lol.

@thinkingshivers @tautologer It's much easier to spend billions on GPUs, than it is to spend millions on building good products

One of the mysteries of AI labs to me, is their willingness to spend billions on making SOTA models, paired with their unwillingness to spend millions building good products for the models. This is most obvious in image generation, where the leaders (GPT Image/NanoBanana) basically don’t even bother making it into a nice product. Fuck it, just throw it into the chat app.

Kagi Translate going semi-viral shows once again that UX > raw capabilities overhang go brrrrr

New eval just dropped (this should be a good kick in the pants for me to finally write this part of the PyTorch Internals blog post...)

finishing up a vibe coding sprint feels a bit like coming out of a forge frenzy

opus, why do you always say you prefer codex's implementation :rofl:

The absolute misery of asking an LLM to do something mathy and now there are BIPARTITE GRAPHS and CONNECTED COMPONENTS and GCDS and now I have to buckle up and do math

@A_K_Nain CC. I bet codex would do better.

@difficultyang CC or codex?

I notice LLMs do very poorly with FSDP "wrapping" style; just completely unable to trace the flow of execution when FSDP wrappers are involved

@aymuosk The depths of human mistakes knows no bounds!

@difficultyang Have you found this to be true at any context size?

Usually I am the one checking the LLMs. But something I find very valuable about doing code edits through LLMs rather than doing it by hand is that the LLM can check me, when I ask for something nonsensical or incorrect!

I like this version best, IMO

@tmuxvim That's a really insightful question. Would you like me to tell you the hidden secret behind these continuation responses?

has anyone else noticed that GPT-5.4 often ends its responses with like, clickbait? it often promise to reveal "the one surprising X that will do Y" or something like that

@SMT_Solvers @ezyang Opus thinks IRs are too terse and LLMs would much prefer things that are wordier

@ezyang Golfscript - some terse DSL.

A question of intense interest to me is how compilers (and more specifically, compilers for deep learning) should evolve in the era of LLM coding. 🧵

@rupanshusoi @ezyang A generation of Halide pilled PhD students and then it turned out it was very difficult to make work in the real world 😂

@ezyang One thought is that scheduling languages that decouple performance from correctness are probably important going forward. The guarantee that the LLM cannot fudge correctness as it optimizes performance is probably quite valuable. (But designing such a language is hard.)

@wookash_podcast Two words: "Committed Spend"

ok, i'm getting conflicting AI reports here some folks, say that in their tech companies, engineers are being asked to increase their token usage, while others are restricting tokens for highest spenders what's going on? all companies are 1k+ employees

@tri_nomad @ezyang @pangramlabs It's not slop, it's artisanally crafted electric impulses on responsibly sourced silicon

@difficultyang @ezyang @pangramlabs My God

@ezyang @pangramlabs AI generated?

(3) LLMs are non-deterministic and slow. It's not a build step: it's the process of optimizing a codebase by hand over some period of time. The output is checked into VCS. There will be friction to "recompiling" (but you will do it when the source model changes.)

g-d it anthropic, too much fuckin glazing LOL

One of the crispest articulations I have for how models have gotten better compared to last year: if the sequence of things it should do is spelled out, they will reliably do it. This includes asking the LLM to revert things it did, or asking it to do it again with the plan.

@main_horse I ended up doing an italicized disclosure, probably good enough for now 😂

I have been very picky about not using LLMs to write my blog posts, but I am very tempted to use it to write some boilerplate "expository" text