Wednesday, August 12, 2026

Pulse (last 3d) · Research 95 · Agents 64 · Models 62 · Products 45 · Enterprise 32 · Open Source 26 · Industry 24 · Legal 22 · Inference 19 · Infra 11

Trending · OpenAI 20 · Claude Code 11 · Anthropic 10 · Codex 8 · Claude 7 · GitHub 7 · Hugging Face 6 · Nvidia 6 · Meta 5 · Caris 3 · DeepSeek 3 · Mark Zuckerberg 3

---

Stealing Reasoning Traces from Proprietary LLM APIs - Researchers demonstrate they can extract reasoning traces from a proprietary LLM API by replaying them through a weaker sibling model. stolen-thoughts.com

Context-Induced Activation Drift - Benign long context (100–3000 tokens) causes large activation drift and breaks RLHF refusal behavior in Gemma-3-1b without any adversarial prompt. Reddit

Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding - Paper argues agentic coding READMEs grow without bound because deleting instructions risks correctness regressions, costing exponential verification effort. arXiv

Encrypted reasoning from ClosedAI et al 100% recoverable - Claim that encrypted-reasoning features from major closed AI providers are fully recoverable by a third party. Reddit

Daybreak models are now available on AWS - OpenAI's Daybreak cyber models (Daybreak Blue/Red), including GPT-5.6 Sol, are now available on AWS Bedrock for authorized defensive security work. OpenAI

Nvidia Nemotron 3.5 Lightning + NeMo Switchyard - Nvidia shipped a 30B MoE (3B active) model explicitly built as the execution tier for always-on agent systems, plus an open-source Rust routing proxy. NVIDIA Developer

Tracely - Open-source agent tracing tool that grades every trace, clusters failures into issues, freezes bad runs into hermetic replayable test cases, and blocks PRs that would ship them again. Product Hunt

HyperSAE - Open-source interpretability tool mapping LLM features into hyperbolic tree structures. Reddit

Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral sign EU Code of Practice on Transparency of AI-Generated Content - All six major AI providers have signed the EU's voluntary transparency code for AI-generated content. Reddit

Claude will apply invisible watermarks to AI text and images - Anthropic will embed invisible watermarks in text and images produced by Claude, with early users already reporting false positives. The Verge

NVIDIA building 1T+ parameter Nemotron 4 family - NVIDIA is building a next-gen Nemotron 4 family with at least 1 trillion parameters to compete with leading Chinese open models. Reddit

Bristol-Myers Squibb to build AI Factory with NVIDIA - BMS announces partnership to build what it calls the most powerful AI factory in life sciences using NVIDIA infrastructure. BMS News

We quantized DeepSeek V4 0731 and benchmarked it - Community quantization of DeepSeek V4 (0731) benchmarked against other quants on an 8× RTX 5090 rig. Reddit

Meta AI can now connect to email and calendars - Meta AI update adds email/calendar access, slide generation, and recurring tasks. Reddit

Apple could help you prove your iPhone photos aren't deepfakes - Apple is developing a system to cryptographically prove iPhone photos are not deepfakes. The Verge

OpenAI's head of ethics leaves less than a year after joining - Departure reignites debate about whether ethics teams can function inside frontier labs. Financial Times