Senior in Python and MLE projects? Ping me, we are hiring!

¿Quieres aprender sobre aplicaciones prácticas de la #IA en empresas y Administración Pública? 📢 ¡Matricúlate en el curso de verano organizado por S.E. Digitalización e Inteligencia Artificial en la Universidad Menéndez Pelayo UIMP! 💡 1.5 ETCS 📅 16-17-18 de julio Plazas limitadas 👉 uimp.es/agenda-link.ht…

Overclocking LLM Reasoning: Monitoring and Controlling Thinking Path Lengths in LLMs Introduces a method to monitor and control structured reasoning in LLMs by extracting and manipulating a “thinking progress vector” (TPV) from hidden states. 𝖯𝖠𝖯𝖤𝖱 𝖨𝖭 𝖠𝖫𝖳

LLMs Context Ops... That's the field, the topic, the science

Just saying to the LLM in the system prompt "...you must reason in <lang>" doesn't seem to work well when lang!=English. Luckily, a few RL steps (tested using GRPO) can help a lot.

BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning "we introduce BREAD: a GRPO variant that unifies the SFT and RL stages via partial expert guidance and branched rollouts. When self-generated traces fail, BREAD adaptively inserts short expert

Kind reminder! We're still looking for senior Python engineers with 2–3 years of experience working with GenAI.

My passion for how Operating Systems work helped me realize that a limited context window isn't a problem—as long as you keep the necessary information in context at each step. This insight was also key in developing Maisa’s KPU

We have finally released the 📝paper for 🥂FineWeb2, our large multilingual pre-training dataset. Along with general (and exhaustive) multilingual work, we introduce a concept that can also improve English performance: deduplication-based upsampling, which we call rehydration.

When using LLMs' structured outputs (JSON mode) feature, you may note that the values of the resulting schema are not very long. Fortunately, if you are using an open-source LLM, reinforcement learning (RL) can help you there!

AI researchers are the new rock stars 👏