Open tabs (June 12, 2026)

I just read this one right before posting. The post describes the difficulty agents have at retrieving biological data. Which isn't limited to agents! It's difficult for a human to navigate the disparate databases and web interfaces and NCBI Virus search incantations to get the thing you're looking for. If this problem were solved for agents, it'd make life easier for us humans as well. A conclusion from the post: "We want models to be creative when they generate hypotheses, design experiments, or reason about mechanisms. But the layer underneath that creativity—gene identifiers, schemas, retrieval logic, coordinate systems, metadata conventions, and data access paths—has to be boringly reliable (or in other words, deterministic)".

I haven't had a chance to do anything with Fable yet, mostly because I work in AIxBio, and Bio is off limits. And because I'm a biologist, Fable refuses to talk to me ("Who am I?" leads to safety flags and demotion of the rest of the conversation to Opus). Precautionary principal is probably the right move here given the benchmarks, and I think managed access will likely be the way these models are released from here out.