will brown

create your own environments. train your own models. be your own lab.

@HanchungLee @FireworksAI_HQ they used fireworks for inference during RL, trainer was in-house

so, how was the @FireworksAI_HQ rl infra compare to thinky or pi? or did cursor home brewed?

I personally find this quite inspiring — Cursor is much more sophisticated than what most companies can do today, but compounding your learnings into a model that excels at your use case is the ultimate way to build an AI product.

@TheAhmadOsman optimizing local stacks just for bs=1 was never the right call personal agents are becoming multi-agents real quick

DGX Spark uses unified memory > 273 GB/s RTX PRO 6000 delivers > 1.8 TB/s (1792 GB/s) If someone told you they’re comparable, they’re wrong And this is exactly why llama.cpp isn’t the right tool here Try vLLM or SGLang on a GPU and you’ll see very different results

@yacineMTB yeah

prediction: someone is going to get a coding AI like codex to automate turning existing steam video games into harnesses, come up with architecture to parallelize the games themselves in a manner that is conducive for RL training, and train an RL demigod model

@celinehalioua “☹️” 💀

went on a date with a gen z guy and he told me i tweet like a boomer ☹️

@vikhyatk yc startup for automating signing up for all the other yc startups that automate all your ops stuff who’s building this

software generation is no longer the bottleneck. it's operations trillion dollar opportunity for whoever solves it

My “I don’t have LLM psychosis" hoodie has people asking a lot of questions already answered by my hoodie.

it's really exciting to see openai contributing $1M in compute and a new speedrun environment. this is also a nice benchmark for frontier model

underrated product: @PrimeIntellect’s liquid compute. we're going to earn money on one of our clusters this month.

Big fan of Luke and what they’re doing at workshop labs. It’s also great tactical example of why we built this feature, no lab should be forced to pay for idle compute when there is an abundance of demand. Tbh it’s insane it’s not an option with any other providers.

it was a good event sir

At Prime Intellect, we’re building that stack end to end: - agentic RL training and inference on frontier open models - RL sandboxes - open-source libraries like verifiers + prime-rl Giving everyone access to frontier lab infrastructure

@lateinteraction @tom_doerr re*search

@tom_doerr Ah, I can’t unsee that extrapolation ..

Folks worked on multi-hop open-domain question answering since 2019. Led to powerful systems like GoldEn, IRRR, Baleen, and STORM. Then it got rebranded “deep research”. A small fraction worked on LLMs as optimizers. Now that got rebranded “autoresearch”. 🤔 *-research it is!

Great kickoff at NVIDIA GTC - starting with a fun panel on agentic AI this morning with @steipete @SGRodriques @hwchase17 @saranormous @Alfred_Lin

Today, we’re sharing how our collaboration with @nvidia helps power the open superintelligence stack. The next frontier of AI infrastructure is building systems for agentic models that can reason for hours, use tools, execute code, and learn from outcomes at scale. primeintellect.ai/blog/nvidia-co…

Letting a provider see all your data is the price of admission for AI. We're changing that. Introducing Silo, the first private post-training and inference stack for frontier models, with hardware-level guarantees that we can’t see your data. Privacy without compromises. 🧵

first time in whichever of the San’s is in south bay i always forget

Stitched up smth fun over the weekend, prototype of an autoresearch RLM environment inspired by @karpathy, using @PrimeIntellect infra. Haven’t run full evals yet, but the setup looks like this: The model can: - modify training file - run experiments inside a sandbox - parse logs for the metric (val_bpb) - iterate to improve the score So the model does the full research loop: code, experiment, logs, hypothesis, patch, repeat Essentially turning autoresearch loop into an RLM training environment, producing trajectories of autonomous research behavior. The interesting bit would be generalizing this to: - any repo - any metric - any experiment harness - envs where model can optimize on specific pieces in a big codebase. Most importantly, this produces trajectories of autonomous research behavior. From those we can identify failure modes and iteratively improve the environment itself.👀🧋🎋 Github: github.com/stochi0/athena… Environments Hub: app.primeintellect.ai/dashboard/envi…

@LottoLabs qwen3-4b-instruct-2507

I like my models small, chinese, dense and not thinking.