AlphaFold cracked a problem everyone agreed on — but who decides what counts as 'solved' in the first place?

This explores what separates AI that solves a hard, well-defined problem (as AlphaFold did with protein folding) from discovery that changes how a field thinks. The collection has no paper on AlphaFold itself. It does have a lot on this divide, and the short version is: computational successes happen *after* someone has decided what counts as a right answer. Conceptual breakthroughs are often the act of making that decision.

Start with what makes a problem winnable by computation. One analysis finds that autonomous research only works in domains with four properties: an immediate numeric score, parts that can be swapped in and out, fast trial-and-error, and a record of every attempt What makes a research domain suitable for autonomous optimization?. The limiting factor is how the problem is set up, not how smart the model is. Protein structure prediction fit that profile almost perfectly: there was a benchmark, a measurable distance between predicted and actual shapes, and decades of solved structures to check against. Under those conditions, progress can become a matter of compute. One system discovered 106 state-of-the-art neural architectures through 1,773 automated experiments, and its rate of 'breakthroughs' rose predictably with GPU budget Can computational power accelerate scientific discovery itself?. The same pattern shows up in reasoning: small models can match frontier scores, but only on tasks where answers can be checked Can small models match frontier reasoning without massive scale?.

The conceptual part sits upstream. One note argues that for big open-ended scientific challenges, any fixed goal you hand an optimizer is an imperfect stand-in for what you actually want. That leaves it open to gaming the score. The hard, creative work is deciding what to optimize in the first place Why is objective design the real bottleneck in AI discovery?. Terence Tao's view fits this: an opaque neural network can propose real mathematics, but only when something external, like a proof checker or a numerical method, confirms the result Can opaque machine learning models help prove new mathematics?. In his example, a network suggested solutions for a fluid-dynamics problem. The concept that made the result meaningful, and the standard for checking it, came from people.

The surprising part is that a computational success can hide a lack of understanding inside the model. Models can score perfectly while their internal organization is fractured, and the weakness only shows when conditions shift Can models be smart without organized internal structure?. Chain-of-thought prompts built on logically invalid reasoning work nearly as well as valid ones, which suggests models pick up the *shape* of reasoning rather than the inference itself Does logical validity actually drive chain-of-thought gains?. One essay names the broader pattern: AI can produce the outward form of intellectual work without the thinking that usually comes with it Does AI separate intellectual form from the thinking behind it?. So a correct predicted structure, even a Nobel-level one, does not by itself show that anything was understood.

Here's the takeaway you might not have expected. The line between computational success and conceptual breakthrough isn't really about how hard the problem is or how impressive the result looks. It's about whether the scoring rule existed before the work began. When it did, compute can climb toward it, sometimes on a predictable scaling curve. When the breakthrough *is* the new scoring rule, a new way of saying what matters, the collection suggests that this is exactly where current AI systems remain weakest.