Thursday, August 20, 2026

Pulse (last 3d) · Research 103 · Products 76 · Agents 62 · Models 24 · Industry 21 · Legal 21 · Open Source 20 · Enterprise 19 · Inference 18 · Media 18

Trending · OpenAI 24 · Claude 22 · Anthropic 14 · ChatGPT 12 · Claude Code 6 · Cursor 6 · Google 6 · Codex 5 · OpenRouter 5 · Alibaba 4 · Gemini 4 · Nvidia 4

---

Same GRPO recipe on three from-scratch LLMs gave three different outcomes. Author trained 353M/316M/672M models with identical GRPO and found RL post-training hurt the two larger variants, with no clean relationship to scale. Source

Looped Language Models Improve Compositional Tool Calling. Iterative computation reuse enables looped LMs to handle compositional tool calls more effectively. Source

Grouping the Stochastic Machine: Precision, Not Capability, as the Frontier Metric for AI Systems. Position paper arguing that precision, not mean capability, should be the frontier metric for evaluating AI systems. Source

Continuous-Time Reinforcement Learning for Controlled Hawkes Jump-Diffusions. A new RL algorithm designed specifically for controlled Hawkes jump-diffusion processes. Source

Microsoft AI launches MAI: 7 new AI models. Suite spans reasoning, coding, image, voice, transcription, cybersecurity, and enterprise personalization. Source

Origin by Cursor. A new code hosting platform built for human-agent collaboration, with integrated Git repos, PR reviews, and GitHub sync. Source

CrewCode. Free, open-source control center for running multiple AI coding agents across real repositories with isolated worktrees and unified diff review. Source

LineBreak Gate. A git-only CI gate that blocks CVEs (fail-closed) and serves human-approved acceptance criteria to coding agents over MCP. Source

Open-weight model closes gap on autonomous cyber offense. Irregular reports Kimi K3 is the first open-weight model to pass CyScenarioBench for autonomous cyber offense. Source

Pacing model development in an era of cyber-critical capabilities. HN debate on whether open-weight models matching closed models on cyberbench should trigger alarm. Source

FDA CDRH discussion paper on regulating generative AI-enabled medical devices. Released Tue 2026-08-18; seeks public feedback on regulatory approach (docket FDA-2026-N-7874). Source

B7-H3 ADC Landscape: First Phase 3 OS Benefit, Multi-Program Race. Regulatory update on B7-H3 antibody-drug conjugates showing first Phase 3 overall survival benefit. Source

Independent AI users converge on identical file-based memory architectures. Eighteen Reddit users described nearly the same memory setup for LLM sessions, suggesting a common pattern is emerging organically. Source

Mathematics in the age of AI. Discussion of Terry Tao's essay on mathematics and LLM-generated proofs. Source

Feature Request: Support AGENTS.md. Community request for a standardized agent instruction file, analogous to README.md. Source

Extensible Software in the age of LLMs. Essay arguing LLMs enable personalized 'software for one,' reshaping extension ecosystems. Source