Wednesday, July 8, 2026

Universal layer geometry across 38 open-source LLMs. Prime Intellect's Elie Bakouch uses Jacobian-lens similarity to show hidden layers align across unrelated models. Source

TurnOPD: Turn-aware on-policy distillation. Improves efficiency of long-horizon agent training by making distillation aware of turn structure. Source

Hierarchical Sparse Attention for near-infinite context. A new sparse attention design aimed at scaling context windows without linear cost. Source

DFlash speculative decoding in llama.cpp. Block-diffusion speculative decoding delivers 4.44x speedup at 36K context on Qwen 3.6 27B. Source

Liquid AI releases Antidoom. Open-source method that cuts reasoning-model doom loops from ~10–23% down to ~1% by retraining overfit transition tokens. Source

Goodfire Block-Sparse Featurizers. Unsupervised mapping of multidimensional geometric concepts in neural network activations. Source

Meta's Muse Image goes live. Alexandr Wang's team ships an agentic image generator with pre-planning, alongside an early Muse Video preview. Source

mistral.rs v0.9.0. Claims up to 1.8x faster CPU decoding than llama.cpp on x86 (AVX2/512) and ARM (NEON), with full repro scripts. Source

Gepard streaming TTS. 0.6B-parameter Apache-2.0 model with ~50ms time-to-first-audio and vLLM-native serving for real-time dialogue. Source

China weighs restricting foreign access to Alibaba and ByteDance frontier models. Proposed rules would mirror US export controls on critical technology. Source

Claude Cowork expands to mobile and web. Anthropic extends its agentic assistant for cross-device task handoff, rolling out to Claude Max subscribers. Source

Cloudflare opens Monetization Gateway waitlist. Stablecoin-based billing for APIs, datasets, and MCP tools over the open x402 protocol. Source

Google expands Gemini Managed Agents. Adds background tasks, remote MCP integration, and the Antigravity agent for asynchronous execution. Source

Lilian Weng on harness engineering for AI self-improvement. New post examining how to design training harnesses for self-improving systems. Source

Learning FlashAttention the Hard Way (Part 1). Algebraic walkthrough of FlashAttention's core math, built from first principles. Source

AI finds 7 bugs in Cloudflare's CIRCL crypto library. Discussion of Cloudflare using AI to surface vulnerabilities in their own cryptography code. Source