Aug 12, 2026

Cactus releases a 45M-parameter agentic LLM: a single 14MB binary running in 28MB of RAM, built narrowly for tool calling, device control, and structured extraction. Their Simple Attention Network compressed to CQ2-bit, aimed at phones, wearables, and robots, with an in-browser WebAssembly demo.

Why it matters: A bet against cloud-first agents: drop general chat, keep tool execution, and tens of megabytes suffice. That existence proof is what set off the on-device crowd.

Researchers from the Tübingen AI Center, MPI and others show that hidden reasoning traces behind proprietary LLM APIs can be reconstructed with two API calls, and demonstrate extracting secrets from the recovered reasoning plus misuse-direction experiments.

Why it matters: It breaks the assumption that hidden reasoning is private by design. The thread grew as a warning against putting secrets or instructions inside the hidden trace.

Anthropic signs the EU AI Act Article 50(2) transparency Code of Practice: models launched in the EU on or after Aug 2, 2026 embed text watermarks and signed provenance metadata from day one, applied worldwide across the API and Claude Code. Older models are in a transition period.

Why it matters: Two flashpoints: the marks apply worldwide, not just in the EU, and they follow content into downstream use like pasted code and deliverables. Promised third-party detection tooling reopened the detectability debate.