@OmkarRay9 The no-answer-key problem is where I'm stuck. I'm building a math problem-generation Skill for my exam prep, and there's no gold set for "a good problem". Adversarial review catches broken ones, not whether one is worth doing. Is a small golden set of my own ratings enough?
@0xlelouch_ #1 applies to tiny projects too. High school senior here, building a math problem-generation Skill for my study dashboard. Before rewriting it I ran failure attribution + adversarial review. I haven't yet measured whether the new version writes better problems. That gap is #1.
@ycombinator@divnori@garykbrixi@FrancoisChauba1@yusufroohani@Omantahealth Bio-elective high school senior here. Last year I got to visit a university lab and came home with fluorescence microscopy photos of cells. Made me wonder: for virtual-cell models, is the bigger bottleneck labeled imaging data or perturbation/sequencing data?
OpenAI's new BioBasicsBench: 100 routine wet-lab tasks. Top model GPT-6 Astra scored 66.7%, up from 43.8%.
What still trips it up? Misreading controls.
That's the first rule high school bio drills into us: no control group, no conclusion. AI needs that lesson too 🧬
@EveryDevAi@allen_ai This is the eval I care about as a student building toward research. My question: does SUPER check that the experiment's result is correct, or only that it ran?