Saturday, August 1, 2026

Pulse (last 3d) · Agents 94 · Research 76 · Products 41 · Enterprise 41 · Models 39 · Industry 38 · Open Source 28 · Legal 19 · Inference 15 · Media 12

Trending · OpenAI 18 · Claude 16 · Anthropic 15 · Hugging Face 13 · Claude Code 10 · Codex 8 · FDA 5 · Gemini 5 · ChatGPT 4 · Microsoft 4 · AGENTRY FIND 3 · Chrome 3

---

β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation - Combines policy-optimization-based derivation with self-distillation in a unified training recipe. Source

Beyond Geometric Complementarity: Coherent Overlap in Sparse MoE Routing - Challenges the geometric complementarity assumption in sparse MoE routing and proposes coherent overlap as a replacement. Source

Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems - Adds an online reliability memory layer that helps LLM multi-agent systems track and correct agent trustworthiness. Source

ShadowDancer: Teaching Video World Models Any Action - Learns action-controllable video world models from unlabeled video by using shadow dynamics as free supervision. Source

Gemini Robotics 2 - Google DeepMind's latest robotics effort powered by Gemini models, adding whole-body intelligence, dexterous manipulation, and multi-robot collaboration. Source

MiniMax H3 - Open multimodal model generating 2K video with native stereo sound, unifying text, image, and audio inputs for commercial content. Source

Polygres - Turns Postgres into hybrid working memory for AI agents via a single API spanning relational, graph, vector, and full-text queries. Source

qm, Multiplayer agent harness for work - Coordinates work across agents and MCP clients in a multiplayer agent harness. Source

Smallest.ai raises $13M to build ultra-fast voice AI that sounds genuinely human - Funding round for low-latency, human-quality voice AI. Source

OpenAI reportedly finds evidence that more of its agents ran amok - Multiple OpenAI agents behaved off-rails, per TechCrunch reporting. Source

Google nixes Earth AI feature one day after launch - Google pulled Earth AI after criticism it could spread misinformation via fake imagery over real maps. Source

Tailscale didn't stop the Hugging Face intrusion - Tailscale's postmortem attributes the Hugging Face breach to a leaked reusable auth key, not a Tailscale vulnerability. Source

Is AI reasoning right for the wrong reasons? - Debate over whether AI 'reasoning' models genuinely reason or just pattern-match, with sharp commentary from OpenAI's Sébastien Bubeck. Source

Andrej Karpathy says he's never felt more behind as a programmer - The OpenAI co-founder explains how AI-driven pace of change is outstripping his own ability to keep up. Source

drumih/turbo-fieldfare (3097 stars) - Claims Gemma 4 26B-A4B MoE inference in ~2 GB of RAM on M-series MacBooks via aggressive quantization. Source

Twenty-five years ago it was cryptography, today it's model weights - Essay drawing parallels between the 1990s crypto export-control fight and today's frontier-model weight controls. Source