Monday, August 10, 2026

Pulse (last 3d) · Products 44 · Agents 39 · Research 30 · Models 23 · Enterprise 22 · Industry 21 · Legal 13 · Inference 12 · Infra 11 · Open Source 10

Trending · OpenAI 12 · Claude 10 · ChatGPT 7 · Claude Code 7 · llama.cpp 7 · Codex 6 · Gemini 5 · DeepSeek 4 · Nvidia 4 · Anthropic 3 · Astra 3 · FDA 3

---

Daily AI Brief, August 10, 2026

Generative design of novel bacteriophages with genome language models - Research paper applying LLM methods to design novel bacteriophages, a high-impact synthetic biology use case. Source

Beyond Output Matching: Preserving Internal Geometry in NVFP4 LLM Distillation - Paper proposing preservation of internal geometry during NVFP4 quantization-aware distillation. Source

Strategy-first synthesis planning for complex natural products - AI planning system for total synthesis of complex natural products using a strategy-first approach. Source

Noise-aware training for analog hardware - Empirical study shows analog hardware accuracy collapses at a threshold rather than degrading smoothly under weight noise. Source

Soup CLI - LoRA training CLI that streams base model layers into VRAM one at a time, enabling Llama-3.1-8B training at 119.6 tok/s in 3.32 GB on an RTX 3050 Laptop. Source

Claude CLI Proxy - Turns Claude Code into a local API server speaking both Anthropic Messages and OpenAI Chat Completions formats with no extra LLM costs. Source

TAKT - Open-source CLI that converts AI coding agents into repeatable YAML workflows with plan, implement, review, and fix loops. Source

KLQ: Training-free measured rotation quantization - Beats SpinQuant and approaches ReSpinQuant at W4A4KV4 on Llama 3.2 1B without GPTQ/LDLQ rounding. Source

Auto mode is now the default in Claude Code - Anthropic makes auto mode default, citing a study of 1,053 testers where auto mode blocked 89% of dangerous commands versus 13.6% human catch rate. Source

Google Weather Next 2 opensourced - DeepMind's WeatherNext 2 hurricane model gives forecasters ~1 extra day of lead time and is now available on GitHub. Source

OpenAI omitted benchmarks in latest release - Critique noting OpenAI left SWE-bench, GPQA, and FrontierMath out of a recent GPT-5.6 release. Source

Sonnet 5.5 leak: 2M-token context window - Report that Anthropic's Sonnet 5.5 may be in development with a 2M-token context window. Source

100% on ARC-AGI-3 ft09 with zero model calls - Solve report achieving 100% on an ARC-AGI-3 task using zero model calls, with analysis of remaining failures. Source

Lessons from the hacks - Nathan Lambert reflects on recent hacks and what they mean for the AI community. Source

Claude Opus 5 system prompt quoted - Simon Willison publishes the Claude Opus 5 system prompt with context on Anthropic's Fable/Mythos export-control saga. Source

GitHub Models has been retired - GitHub Models service is shut down. Source