new preprint. my thesis. comments very welcome. abstract below. link in comment.

Why is no one talking about this arxiv.org/abs/2505.11821…

🔥 Excited to share our work "Efficient Agent Training for Computer Use" Q: Do computer use agents need massive data or complex RL to excel? A: No, with just 312 high-quality trajectories, Qwen2.5-VL can outperform Claude 3.7, setting a new SOTA for Windows computer use. 1/6

AGI achieved internally. Using DSPy and Keras 3, I built a system where an LLM can self-reflect and modify its own keras source code based on iterative performance. See its reasoning as it changes its neural network architecture after an iteration. Open source soon.

🌐 From Matching to Reasoning — Retrieval just grew a brain. LightOn introduces Reason-ModernColBERT, a State-of-the-Art multi-vector model purpose-built for the era of Deep Research — where matching isn’t enough, and true insight demands reasoning. Built on #ModernBERT and

Cursor expert blows my mind with his vibe marketing process in this 60 minute masterclass

1/N Introducing SkyRL-SQL, a simple, data-efficient RL pipeline for Text-to-SQL that trains LLMs to interactively probe, refine, and verify SQL queries with a real database. 🚀 Early Result: trained on just ~600 samples, SkyRL-SQL-7B outperforms GPT-4o, o4-mini, and SFT model

Talk to unmute.sh 🔊, the most modular voice AI around. Empower any text LLM with voice, instantly, by wrapping it with our new speech-to-text and text-to-speech. Any personality, any voice. Interruptible, smart turn-taking. We’ll open-source everything within the

The French Ministry of Culture released 175k high-quality arena-style preferences It's exactly the type of data LMSYS stopped releasing. They created their own chatbot arena with 55 models and open-sourced everything. Incredible work! 🤗 Dataset: huggingface.co/datasets/minis…

I just generated a 5:30 min Multi-Speaker Podcast on Agentic Patterns using Gemini 2.5 Flash and our new Text-to-speech (TTS) Model! At I/O we launched native controllable Audio Generation for Gemini 2.5 Pro & Flash. > Controllable style, accent, pace, tone. > single and

We want to give more visibility to the whole AI community! So everyone can now share community blogposts on Hugging Face. Wether you want to share about your latest science breakthrough, the model, dataset or space that you build or just your opinion on the latest AI dramas, you

Strong recommend for this book and the JAX/TPU docs, even if you are using Torch / GPUs. Clean notation and mental model for some challenging ideas. github.com/jax-ml/scaling… github.com/jax-ml/scaling… docs.jax.dev/en/latest/note…

🚨Wouldn’t it be nice if your agentic search system could reason over all your docs? ✨Introducing Rank-K, a listwise reranker that benefits from test-time compute and long-context! Rank-K sets a new SoTA for reasoning-based reranking, without reasoning chains from other models.

Samuel Colvin Pydantic LangChain Done! If you’d like to collaborate on a framework-agnostic solution let us know. a code interpreter isn't going to be the deciding feature when choosing between agent frameworks. Many use cases will require a full container based sandbox anyway...

Hey fal Bagel 7b is a good candidate to join your cloud! 🙏

build fast guys and drink water 🙏

Love this! What GRPO does with this is amplify behavior that produces code aligned with this. Another, far more sample efficient way to amplify that behavior is to grab the whole trajectory that worked best and stick it into the prompt(s). That gives you dspy.BootstrapFewShot,

DSPy simplifies prompt tuning for optimal LLM responses. We fine-tune prompts based on input/output analysis, addressing incorrect or inappropriate LLM behavior.

12 types of JEPA (Joint-Embedding Predictive Architecture) ▪️ I-JEPA ▪️ MC-JEPA ▪️ V-JEPA ▪️ UI-JEPA ▪️ A-JEPA (Audio-based JEPA) ▪️ S-JEPA ▪️ TI-JEPA ▪️ T-JEPA ▪️ ACT-JEPA ▪️ Brain-JEPA ▪️ 3D-JEPA ▪️ Point-JEPA Save the list and check this out for the links and more info:

I think people don't really use LLM for complex code/projects, here claude4 couldn't solve anything, everything remains the same...