AI researcher at CHAI explains philosophy to building LLMs / AI systems.

KV Cache Steering for Inducing Reasoning in Small Language Models.

Interesting way to steer LLMs via modifying KV cache.- One-shot modification to KV cache.- Improved stability, compatibility, and efficiency at inference time ⏳

❤️

Test-Time Scaling with Reflective Generative Model Hugging Face:

MetaStone-S1-32B introduced a new form that shares the backbone of the policy model and process reward model, and uses self-supervised training to eliminate the reliance on process-level supervision. Reduces the inference cost of PRMs by 99%! 💰

DeepSeek vs. Kimi K2 difference in architecture

Github:

Kimi K2 is SOTA MoE LLM with 32b activated parameters anf 1T total parameters. Exceptional performance across frontier knowledge, reasoning, and coding tasks.

Hiring. Engineers in Palo Alto. hardcore team . ships lots of code . comp ≥ 2x last role.

Thanks for the kind words of appreciation ❤️

SingLoRA: Low Rank Adaptation Using a Single Matrix.

Instead of 2 low-rank matrices (as in LoRA), SingLoRA uses one low-rank matrix and its transpose.

Small change to LoRA’s formula, big difference. 👏.Stabler optimization + halve the parameter count.

❤️

👀.

Mercury: Ultra-Fast Language Models Based on Diffusion. Website:

Claim to be insanely fast - 10x the throughput.

SmolLM3: smol, multilingual, long-context reasoner. Base model: ***.Instruct and reasoning model: ***.

SmolLM3.- 3B model trained on 11T tokens, SoTA at the 3B scale and competitive with 4B models.- Instruct model with dual mode reasoning, supporting think/no_think modes.- Multilingual support for 6 languages: English, French, Spanish, German, Italian, and Portuguese.- Long