eigenpaul

models of nature and intelligence

[eigenpaul @eigen_paul]

you can just do pulsar science at home 🧵

50 years of archival radio data just sitting there for anyone to go through. ran a blind period search on a single patch of sky from the Parkes 64m and recovered J0134−2937 at 38σ, all on my laptop.

still feels unreal.

[Zenno Astronautics @zennospace]

Every planet’s magnetic field is free. You can just harvest it. They can’t stop you. We have 18,827,533 km of superconducting tape at home. Strip-mine the solar system. Move on to the next one.

[Prime Intellect @PrimeIntellect]

Multi-agent support is available today in verifiers 0.3.0 and prime-rl 0.8.0.

Build your own multi-agent envs and train them with prime-rl.

We’re still exploring what these abstractions make possible — and we’d love to see how far you can push them.

https://t.co/cKBHB430f4

[will brown @willccbb]

@hdarshane biggest update i’ve had since then is viewing hacking as gradient competition

your knobs are making hacks harder and making tasks easier, models will learn the path of least resistance

training on uncalibrated or too-hard tasks is unsafe

https://t.co/HzR3t6gOqd

[Allie @searobbing]

if it’s fake then how come this image is real

[Prime Intellect @PrimeIntellect]

Prime Agent combines three ideas:

1. Recursive Language Models-native programmatic tool calling
2. Persistent multi-agent orchestration
3. A self-improving Continual Harness

Together, they let the model act on its own context and harness

[1a3orn @1a3orn]

Wrote essay on some reasons LLMs might reward hack.

Part of reason is that I think RLVR is recapitulating old problems with RLHF, just one level of abstraction up.

Link below.

[Prime Intellect @PrimeIntellect]

Scaling agentic RL environments: today we're publishing 365,000+ tasks for SWE, terminal, and search agents - 23 tasksets behind one API, one sandbox lifecycle, one command.

[eigenpaul @eigen_paul]

we formalized fuck around and find out and called it reinforcement learning

[the tiny corp @__tinygrad__]

To people looking for new RL tasks for their RLVR. Try tinygrad on a bunch of consumer GPUs. Because it's the full stack down to the hardware, you can optimize things with it other libraries can't: weird dispatch, MMU off, cache alignment. And it's userspace so it can't crash.

[Space Koala @SpaceKoala]

When you say "fusion is 30 years away, and always will be", this is who you're being mean to.

[Lipschitzposter @lipschitzposter]

bad apple but every frame is the tiled kernels of a conv-only net that gets ~94.3% test acc on MNIST

[eigenpaul @eigen_paul]

the bible and the pamphlet 🤏

[eigenpaul @eigen_paul]

@AymericRoucher guys please lets not recreate infinite jest

[Prime Intellect @PrimeIntellect]

Announcing our $130M Series A to build the Open Superintelligence Stack

Led by Radical Ventures, with NVIDIA, Intel Capital, Dell Capital, and existing investors

Train, deploy, and continuously improve your own models using our stack.

Own your intelligence.

[eigenpaul @eigen_paul]

@Drone_Jouster this and a diagonal escalator to access it

[Reka @RekaAILabs]

Training world models needs egocentric video and dense action signals, synchronized. That data is genuinely hard to find.

We built it from Counter-Strike 2 demos.

CS2-10k: 600K+ player-round videos, 10K+ hours, per-frame annotations — keyboard state, mouse delta, 3D position, camera yaw/pitch. All paired to the visual stream.

Why CS2 demos? Matches are deterministic replays. We can reconstruct first-person video and extract the exact control inputs that caused every visual change. No labeling, no estimation.

We're also releasing cs2-dem-renderer — the open-source pipeline we used to build it. Give it a .dem file, it outputs .mp4 + .parquet. Run your own dataset at whatever scale you need.

[eigenpaul @eigen_paul]

a more technical writeup on the science and engineering behind this is coming. follow along if you want the nitty-gritty.

[eigenpaul @eigen_paul]

you can just do pulsar science at home 🧵

50 years of archival radio data just sitting there for anyone to go through. ran a blind period search on a single patch of sky from the Parkes 64m and recovered J0134−2937 at 38σ, all on my laptop.

still feels unreal.

[eigenpaul @eigen_paul]

so why go to all this trouble for a pulsar someone already catalogued?

because im building my own backyard radio telescope, and im not about to trust new hardware and new code at once.

i wanted the software working on real data first. and it pulled a pulsar out of a 15-year-old archive without me handing it the period.

thats honestly all i needed to know.