Caleb Eom

over-analyze everything.

[Caleb Eom @calebfoundry]

@LeveragingTIME I'm really curious to find out who caught the news first and give them credit but also hold them accountable.

[Caleb Eom @calebfoundry]

Does anybody know who actually first broke the news about Jensen's affair with Alice Huang?

[Caleb Eom @calebfoundry]

so is ARC AGI dead now

[Caleb Eom @calebfoundry]

catching up on JEPA research papers feels like reading a 10-year saga that Yann is unfolding

[Caleb Eom @calebfoundry]

@ayaanyyz "make no mistake" 🙅
"make mistake" 🫡

[Caleb Eom @calebfoundry]

can you please explain the intuition behind why using multiple large target blocks somehow performs better over something seemingly trivial..? @ylecun

[Caleb Eom @calebfoundry]

its hard to admit but i think im addicted to reading research papers.

the dopamine hit from being in flow state trying to wrap my head around novel ideas is so addicting.

shout out to my dealer arxiv for giving me the good stuff.

[Caleb Eom @calebfoundry]

@savov PPO part of what they published shouldn't be impacted since it was a pure edit no tools/forward pass

[Caleb Eom @calebfoundry]

i created a open benchmark on "weight-space programming" and GPT-6 Astra scored the highest at 96.3 against baseline (500M PPO).

runner up: GPT-6.1 Sol, Claude 5.5

my main curiosity was how other LLMs write weights from scratch rather than editing them or "reading them".

so this benchmark only measures the model's ability write new weights from scratch.

full disclosure, im pretty skeptical about this still - but i think it's crazy to just let this idea go by. we need the ecosystem to run with it and see its limits, and train foundation models to optimize this.

github repo: https://t.co/E1Cmqygfh8

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Caleb Eom @calebfoundry]

ok britney spears kinda slaps

[Caleb Eom @calebfoundry]

@gpelluet27 the PPO edit i believe was a pure edit without forward pass and validation.

otherwise, like you mentioned Gemma 3 270B edit and Qwen edit both use tools and experiments to try to narrow down the problem which sounds less exciting than a pure read.

[Caleb Eom @calebfoundry]

Weight-space programming explained.

[Caleb Eom @calebfoundry]

@kostasbotonakis this one specifically doesn't measure math but a small neural net ~25k weights from scratch to play games.

it would take 2-3 days of compute and a lot of money to make this benchmark scale to test its ability write LLMs that performs on domain specific jobs.

[Caleb Eom @calebfoundry]

@AdamHoltererer doubt it

[Caleb Eom @calebfoundry]

TIL judgement can be spelled judgment

[Elliot Arledge @elliotarledge]

this is such a well done and beginner-friendly intro to how CUDA works at the high level and will help mold your brain around how kernel engineers think about performance for any given problem. no code at all, just dark theme diagrams!

link below

[Caleb Eom @calebfoundry]

@bycloudai @elliotarledge @bycloudai no you ♥️

[Caleb Eom @calebfoundry]

CUDA explained from a first principles perspective.

Turns out, it's more than just a collection of kernels but it's a huge ecosystem and a huge rabbit hole to dig into.

@NVIDIAHPCDev @NVIDIAAIInfra

[Caleb Eom @calebfoundry]

ok im now back to being stupid again. CUDA is a huuuuge rabbit hole.

https://t.co/HrlSDzJli9

[Caleb Eom @calebfoundry]

holy crap i finally understand CUDA.. finally at the peak of mount stupid, now heading towards valley of despair..