Training World Models @ StealthCo Author @shaderbench Formerly Research Scientist @Canva, @Adobe, @wandb
San Francisco, California
- Tweets 6,669
- Following 689
- Followers 1,116
- Likes 11,301
TLDR;
Throw away pytorch.
Fuse your kernels in CUDA
seems like programming language convergence is crossing the overton window thanks to @dhh. i still stand by below, it hasn’t aged.
some programming languages will die, authors of languages who do not adapt for a changing language consumer demographic (agents). for what is the point adding sugar to a language (operator chaining) to make a language easier for humans to produce code when humans no longer write the code?
in other news; i’ve been smashing the OTP and Rust pretty hard lately. i’ve found my personal happy divide:
rust: cli, anything where memory pressure and perf need explicit guarantees (ie: network daemons or sockets) and for everything else elxir/otp. god damn these agents can attach to an actor process, observe what’s going on and troubleshoot fassst.
70
Opus 5.5 is a refactor Monster
"Take a look at X, describe the data model, types and systems designs. How can we drastically simplify the code base and remove -500 lines"
1
1
61
Some papers age like wine.
Hertzmann, SIGGRAPH '98
for the past few months i've been asking our models to paint. opus 5.5 is very skilled at emulating different styles
every image here is a python program generated pixel by pixel. there is no image model, and no off-the-shelf art software. instead, it's about 7,500 lines of code using standard libraries to emulate different brush styles. the agents don't use any pictures as reference, instead working only from what they know about each painter
1
74
At what point does educating the world's technical talent conflict with preserving opportunities and technical capacity for Americans?
In 2024, temporary-visa students earned:
• 61% of U.S. computer science doctorates
• 54% of engineering doctorates
• 52% of math & statistics doctorates
At MIT, 41% of graduate students are international.
Source: NSF/NCSES; MIT, 2025–26
82
Why does this guy have the weirdest speaking Rhythm...
It's ... Like.. random pauses... that don't.. make sense
BREAKING: I just signed an EXECUTIVE ORDER to dramatically accelerate independent oversight of AI companies and advance the creation of an “AI kill switch.”
With Donald Trump and Congress failing to lead, California is stepping up to keep all Americans safe.
99
The first thing I do in ever new folder is to sim link CLAUDE.md => AGENTS.md
this will save me a whole 10 seconds!
We're adding support for AGENTS.md to Claude Code.
Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md.
You can toggle this behavior in /config.
1
146
You can do this with VLLM + Parallel Decode on top of an LLM without any training.
I used this trick to get a 100x speedup in inference.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
184
Starting to regularly hit 10+ agents working for me at a time. Crucially, they are not just bumping into each other and are coordinating work across different roles.
New version of Unleashd cooking!
(My Open Source Swarm Management software)
64
My hot take on Naiver Stokes:
It's not that impressive, exactly what you'd expect if you scaled up test time compute from Erdos problems to a harder question.
I'll be especially excited when the LLM's can invent new fluid dynamics equations. New physics models, new fields of math.
Stuff like this is a way more exciting way to use LLM's than to find increasingly hard counter examples
nitter.net/EMostaque/status/20966…
---
They proved a counter example to navier stokes, which essentially amounts to finding an edge case in a mathematical equation.
Millennium prize for humans because it requires an insane amount of effort, but is largely a brute force problem that LLM's are very good at.
In fact, MOST of the LLM Erdos problems are of the same variety, finding counter examples.
They are still incredibly capable machines, but for them to really start changing the world they need to start deriving new models of the world themselves, not just applying them or operating within them, which means they need to set their own goals.
The standard model is awesome!
It is also unique 👀
Take all compact Lie algebras & sub-algebras
Filter based on chirality & anomaly cancellation
Only e8 ⊃ e6 ⊕ su(3) → 27 ⊗ 3 survives
No extra dimensions or 10^500 vacua needed..
emad-ii.github.io/chirality-…
4
277