Monday, August 17, 2026

Pulse (last 3d) · Industry 9 · Research 5 · Products 2 · Policy 2 · Open Source 2 · Agents 2 · Models 1 · Legal 1 · Infra 1 · Media 1

Trending · Tempus 2 · AACR D3 2026 1 · CNPV 1 · Caris 1 · Claude 1 · Cobenfy 1 · Commissioner's National Priority Voucher 1 · Daraxonrasib 1 · FDA 1 · GB-4362 1 · GLM-5.2 1 · Generate Biomedicines 1

---

Second Thought: Reasoning in Parallel as LLM Agents Act and Observe - A method for parallel/lookahead reasoning during LLM agent act-observe steps, improving tool-use efficiency. Source

Paper claims RL for reasoning only changes 1-3% of tokens - Paper argues RL for reasoning changes only 1-3% of tokens and gains can be replicated without RL at ~1000x lower compute. Source

Multimodal Model Diffing for Feature Discovery and Control - Paper on using diffing techniques to discover and control features in multimodal models. Source

Defending Against AI-Generated Crisis Video Attacks - Systematic study of defenses against AI-generated crisis-event video attacks across detectors, generators, and social spread. Source

cache-assembler - Proxy that fixes prompt-cache invalidation for Claude Code and the Anthropic API, delivering 8.2x cost reduction on real traffic. Source

Mac Developer Bridge - Open-source MCP bridge that lets ChatGPT operate your actual Mac via shell, files, and real PTY sessions. Source

Agentinel - Locally-run firewall for AI agents that intercepts hallucinated and malicious packages across JS, Python, and Rust. Source

MCPScan - Offline, zero-telemetry scanner that discovers and audits Model Context Protocol servers across private networks and local AI tools. Source

OpenAI disbands its preparedness team - OpenAI has reportedly disbanded its preparedness team. Source

Nvidia scales back OpenAI data-center guarantee - Nvidia reportedly reduced its potential financing guarantee for an OpenAI data-center project in Ohio. Source

Anthropic publishes watermark FAQ - Anthropic releases an FAQ explaining their approach to LLM output watermarking for EU AI Act compliance. Source

Qwen 3.8 27B review - Hands-on review noting strong benchmarks but high-latency overthinking behavior by default. Source

Frontier LLMs can't write GPU communication code - CommBench benchmark shows the best model scores only 30.7% on 100+ GPU communication tasks (NCCL/NVLink). Source

ChatGPT's computer history logs clicks and keystrokes - Report on ChatGPT's computer-use feature tracking detailed user interaction. Source

Claude watermark removal loophole - Reddit user claims a method to strip Claude's watermark without rephrasing output. Source

Model release roundup - GLM-5.3, DeepSeek V4 Pro, Gemini 3.7 Flash, Grok 4.6, GPT-5.6 Cyber, Muse Spark 1.2, and Qwen3.8-12B all released recently. Source