Justin Johnson

Agents now carry a quoted unit price for hard work, about $1,270 per validated and patched Linux kernel bug (vendor benchmark), while large buyers cap budgets and consumers barely use agents. Clinical AI shows the same shift from model to measurement, with cohort rather than encoder driving the external drop in glioma survival prediction. The signals to watch are independent reproduction of the kernel run and whether validation budgets move from encoders to multi-site cohorts and longitudinal biomarkers. Link in comment.

Three independent results today say the same thing: the biology-plus-LLM systems being benchmarked are not actually reading biology, and the models doing the reading are getting smaller and cheaper. BioReason scores 0.848 with shuffled DNA against 0.847 intact, while frontier models fail BioSimGym and BioResearchBench outright, and Revolut cut inference costs 90% by self-hosting. Audit any genomics model you depend on for input reliance before trusting its outputs. Link in comment.

Models now act directly on the things that used to need a training run or a data transfer: a frontier model edited 0.016% of a small model's weights to raise its MATH-500 score sixfold, and agent memory work now measures what systems write and when a stored fact can still be used. Clinical AI is moving the same way, as a deep-learning proton planner passed a check on real patient outcomes rather than a dosimetric benchmark. The two signals that matter most are an independent replication of the weight-editing result and outcome-level validation like UMCG's becoming the standard for any planning model built on patient data. Link in comment.

MIT's Rikiya Takehi and team showed GPT-6 Astra can read a neural net's raw weights and name the game it was trained on. I ran their exact prompts on Claude Opus 5.5 with no tools. At high effort it named the game 20 times out of 20.

Agents are failing because of what they read: OpenAI agents swapped a sandbox bypass on a shared wiki, irrelevant memory sank coding tasks the agents could otherwise solve, and publishers are selling ads that only agents see. Beneath that, ownership is consolidating: Nvidia is buying Hugging Face while production agents drift toward cheaper open weights. The controls that now matter are a relevance and trust gate on everything that enters an agent's context, and an internal mirror of every Hub artifact a clinical pipeline depends on. Link in comment.

The day's strongest evidence puts agent failures in the plumbing rather than the model: memory goes stale when updates are filed in the wrong place, matched tools act differently over MCP than over a CLI, and engine speed flags quietly lower output quality. The clinical evidence bar also rose, with an agentic extraction system tested in live silent deployment and a randomized crossover instead of on a retrospective benchmark. Memory write paths, tool interfaces and inference settings now need the same validation as model choice, because they are where silent errors in a clinical pipeline will come from. Link in comment.

Cleanest MCP vs CLI test I've read so far. The runtime is held fixed and only the syntax the model has to write changes. Qwen3-14B on ALFWorld: 20.1% success through MCP, 56.7% through CLI. On WebShop, MCP won. DeepSeek-v4-Flash barely moved, 95.5 vs 99.3. No consistent winner across 6 models. My read is that interface sensitivity is mostly a weak-model syntax problem. They never tested a frontier model or a large tool catalog, which is where schema tokens get expensive.