bruno

señor htmx ceo

[bruno @spoonrider]

@superalesha Now the loop is going to be LLM grinding to build games from those videos as objective, until the juge can't tell them appart, the gap will narrow quickly

[bruno @spoonrider]

@NomadsVagabonds @PromptCentralX Could you share the soundtrack ?

[bruno @spoonrider]

@gibkun1 Where is it available ?

[bruno @spoonrider]

@juliendubois Interested to peek into it, could you share it with me ? https://t.co/OBxfNZKV1o

[bruno @spoonrider]

@juliendubois Have you tried running https://t.co/NCAYvqkzCt loop on your objective ? with proper targets and tests it could do magic 🤌

[bruno @spoonrider]

@VictorTaelin Sft LORAs not the hard part : once you stop relying on ootb inference endpoints, you need good infra to serve the model, handle Kv cache, prefill, spéculative deciding etc .. hard to reach similar t/sec

[bruno @spoonrider]

@jeremyphoward Searxng as local container

[bruno @spoonrider]

@a1zhang @lateinteraction @zli11010 Using gepa/other prompt optimization over RLM trajectories could improve this further @lateinteraction ? As dspy is already used here, maybe worth looking at the ROMA paper for this

[Aksel @akseljoonas]

Introducing ml-intern, the agent that just automated the post-training team @huggingface

It's an open-source implementation of the real research loop that our ML researchers do every day. You give it a prompt, it researches papers, goes through citations, implements ideas in GPU sandboxes, iterates and builds deeply research-backed models for any use case. All built on the Hugging Face ecosystem.

It can pull off crazy things:

We made it train the best model for scientific reasoning. It went through citations from the official benchmark paper. Found OpenScience and NemoTron-CrossThink, added 7 difficulty-filtered dataset variants from ARC/SciQ/MMLU, and ran 12 SFT runs on Qwen3-1.7B. This pushed the score 10% → 32% on GPQA in under 10h. Claude Code's best: 22.99%.

In healthcare settings it inspected available datasets, concluded they were too low quality, and wrote a script to generate 1100 synthetic data points from scratch for emergencies, hedging, multilingual etc. Then upsampled 50x for training. Beat Codex on HealthBench by 60%.

For competitive mathematics, it wrote a full GRPO script, launched training with A100 GPUs on https://t.co/udm7xGpNzR, watched rewards claim and then collapse, and ran ablations until it succeeded. All fully backed by papers, autonomously.

How it works?

ml-intern makes full use of the HF ecosystem:
- finds papers on arxiv and https://t.co/brvCC7fLPa, reads them fully, walks citation graphs, pulls datasets referenced in methodology sections and on https://t.co/hrJuRkRyzi
- browses the Hub, reads recent docs, inspects datasets and reformats them before training so it doesn't waste GPU hours on bad data
- launches training jobs on HF Jobs if no local GPUs are available, monitors runs, reads its own eval outputs, diagnoses failures, retrains

ml-intern deeply embodies how researchers work and think. It knows how data should look like and what good models feel like.

Releasing it today as a CLI and a web app you can use from your phone/desktop.
CLI: https://t.co/l3K1PslZ1n
Web + mobile: https://t.co/orko5srL4H

And the best part? We also provisioned 1k$ GPU resources and Anthropic credits for the quickest among you to use.

[bruno @spoonrider]

@8teAPi For ref. Jensen recently layed out his view of the future in this long form interview (not 100% matching what post says ) https://t.co/TyHzjxLtjv

[bruno @spoonrider]

@dnhkng Seems a bit like an unrolled TRM, there could be a smooth path between the two

[bruno @spoonrider]

@dnhkng nice!

[Mo @atmoio]

I was a 10x engineer. Now I'm useless.

[bruno @spoonrider]

@MaximeRivest Why not using boolean return type ?

[bruno @spoonrider]

@UnslothAI @Zai_org That's one concept of "locally"

[bruno @spoonrider]

@jeroendee @badlogicgames I had my pi openclaw build it as skill to test it , added a rolling window to avoid same hash for identical lines problem , if you want a minimal standalone starting point : https://t.co/70cPbBwrzA

[bruno @spoonrider]

@springrod @RyeBrye Any pointers at what good codebases for fine-tuning one ? Apart from Spring kotlin codebase

[bruno @spoonrider]

@andrewlamb1111 @ApacheParquet @MOVNTDQ Now we need this for huggingface xet format , to avoid big recomputing when data changes

[bruno @spoonrider]

@benjaminvrbk Are you saying the prompt is hand crafted ? Did you abandoned prompt optimization / dspy?