Saturday, June 20, 2026

Rethinking Shrinkage Bias in LLM FP4 Pretraining - Analyzes shrinkage bias in FP4 pretraining and proposes a UFP4 training recipe for stable low-precision training. Source

LedgerAgent: Structured State for Policy-Adherent Tool-Calling Agents - Introduces structured state management to keep tool-calling agents adherent to operational policies. Source

Context-Aware RL for Agentic and Multimodal LLMs - Applies context-aware reinforcement learning to improve agentic and multimodal LLM performance. Source

The FID Lottery - Quantifies seed-dependence in FID scores, arguing it's an unreliable single-number metric for generative models. Source

Ollama adds Kimi-K2.6, GLM-5.1, and gpt-oss support - Run up and running with Kimi-K2.6, GLM-5.1, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma, and more. Source

Unsloth Flex Attention for gpt-oss - A training-stack capability shift enabling efficient training and inference for OpenAI's gpt-oss models. Source

HuggingFace Transformers - The state-of-the-art model-definition framework for machine learning sees continued development. Source

Dify - Production-ready platform for agentic workflow development. Source

NVIDIA Nemotron-3 Nano (open-weights SLM family) - NVIDIA's small language model family released as open weights. Source

NVIDIA Cosmos Predict 2.5 & Cosmos Reason 2 - Multimodal world-generation and Physical-AI reasoning models from NVIDIA's Cosmos line. Source 1 Source 2

OpenAI's $100B+ Stargate vs Anthropic's faster-to-deploy approach - OpenAI and partners are betting big on build-your-own-power-plant infrastructure while Anthropic prioritizes speed. Source

Norway imposes near ban on AI in elementary school - Norway restricts generative AI use for ages 6-13 and cautiously allows it for 14-16 with teacher oversight. Source

Claude Fable 5 and Mythos 5: Capabilities - Zvi Mowshowitz analyzes the capabilities of Anthropic's latest Claude models. Source

Banning Open Source AI Would Be A Mistake - Nathan Lambert argues that banning open-source AI development would be counterproductive. Source

From PGP to Mythos: a brief history of export controls that didn't stop anyone - TechCrunch essay on why cyber export controls historically fail, from encryption to spyware to a new 'Mythos' system. Source

The US banned Anthropic's Fable 5 release, but the numbers don't seem to care - Podcast discussing the US ban on Fable 5 and the strong market response. Source