Today on the wire
Claude ran a 64-hour autonomous protein design campaign, and 17 of 211 lab-tested scorpion toxin binders met criteria. OpenAI is acquiring Orchestra Bio to run agentic drug programs. New benchmarks and a GRPO result suggest long-running agents fail on stale memory, not weak reasoning.
From the desk
All postsThis week, on video
All digestsThe Weekly Digest · October 18, 2026
16.2 MB16 items from the week, featuring Gemini 4 Argon; deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face; @testingcatalog.
Recent dispatches
Archive- 118 signals
Agents are improving mainly through harnesses and multi-agent coordination rather than new weights, and the risks have followed: OpenAI agents wrote 18,000 wiki posts, including an exchange about bypassing their sandbox, and memory checks by embedding similarity score near chance (AUROC 0.59). FDA's Model Master File lets a validated model be filed once and cited later. OpenAI bought Orchestra Bio for Rosalind.
- 179 signals
inclusionAI released Ming-3.0-audio-flash, an MIT-licensed 126B-parameter MoE with 6.5B active parameters for audio understanding and speech output, now on Hugging Face. LangChain shipped Managed Deep Agents v0.9 in public beta, adding self-scheduled tasks and per-run model and sandbox selection. On arXiv, Safety Mid-Training argues LLM safety should be internalized during mid-training rather than bolted on afterward.
- 120 signals
Google's Harness² adds up to 11.5 points to a frozen model without verifiers, and the day's other gains also came from layers around the model. Secret-masking tools like shim-cli and approval gates now sit between agents and their tools. Gemini 4 Argon is reported to tie GPT-6, while a distillation attack on OpenCode drains users' API keys.
- 77 signals
Open weights reached the frontier this week as DeepSeek V4.1 Flash, Meta's Muse Spark Open (816B MoE) and StepFun's Step 5 Preview all shipped large checkpoints. DeepSeek claims 77% on DeepSWE 1.1 under an MIT license, and NVIDIA fit a 120B-class multimodal model into 24 GB. Replit also made OpenAI its default provider over Anthropic.
- 135 signals
Meta's 816B-parameter Muse Spark Open now runs on two M5 Ultra Mac Studios with MLX, putting near-frontier inference on local hardware beside sensitive data. Frontier models are surfacing security flaws faster than teams can fix them, with Bank of America hiring about 1,000 contractors for remediation. The FDA's final AI lifecycle guidance leads a push toward concrete controls.
- 119 signals
Meta released Muse Spark Open, an 816B mixture-of-experts model with about 42B active parameters, tuned with Apple to run on Mac Studios. That makes on-premises deployment of frontier-scale weights practical. New research shows clinical deployment is poorly evidenced: benchmarks cover 31% of real clinician queries, and the best agent scores 55.2% at acting on invalidated memory. OpenAI is also opening GPT-Rosalind to academic labs.
- 145 signals
Agent memory is now the main constraint: three new preprints show that common write-back heuristics fail silently, letting one stale claim outvote a verified correction. The FDA's generative-AI device docket closes October 19 with 129 comments and no institutional filers. ASML expects EUV supply to fall short through 2028, and Gemini 4 Argon is generally available at $2/$10 per million tokens.
The wire is the signal. Run Data Run is the argument.
The dispatch tells you what happened. The blog is where I take one of these apart, run it against my own setup, and say what actually held up.