What's this about?
Scientific discovery often means optimizing expensive-to-evaluate objectives over vast, structured, open-ended hypothesis spaces — molecules, protein sequences, computer programs. This paper (52 upvotes) proposes Large Discovery Models: generative models built for empirically-grounded, model-based open-ended search rather than for answering questions.
What exactly is an "expensive-to-evaluate objective"?
It means finding out whether a candidate is good costs real time and money. Evaluating a chess move is nearly free; evaluating a candidate drug molecule means synthesizing and testing it. When each evaluation is expensive, the search strategy matters more than search speed — you must choose candidates that are informative, not merely plausible. That's why this is different from generation: a model that produces a thousand reasonable-looking molecules isn't useful if testing them costs a year. The value is in proposing the few worth the expense.
Why is this worth noticing?
- It's directly connected to today's Anthropic protein result — 14 of 15 binders working is exactly the hit-rate improvement this framing is about.
- "Open-ended" is the important qualifier. Most optimization assumes a fixed space of options; discovery means the space itself is unbounded and partly unknown, which breaks standard methods.
- It joins Apodex and the AutoResearch diagnostic in this month's cluster treating discovery as a distinct capability with its own architecture, rather than a downstream benefit of general intelligence.