Rephrase "mission impossible" into an actual science question, get an answer, and then I'll build follow-ups from that.

Rephrasing "mission impossible" into science

Here's a concrete reframe: Can current AI systems solve unpublished, research-level physics problems at the frontier?

The answer

Not yet—and it's a significant gap [AG-2025.09-1139].

Researchers created CritPt, a benchmark of 71 unpublished research challenges spanning 12 fields (condensed matter, quantum physics, astrophysics, high-energy physics, and others), all designed by active physicists to simulate entry-level research projects. The result: even the best current AI models solve only ~6% of full research-scale challenges [AG-2025.09-1139]. When broken into 190 simpler checkpoint tasks, performance is higher, but the gap between AI and what physicists actually need for real research remains large.

Why this matters: it reveals that general reasoning ability—which LLMs have demonstrated in competition math and coding—*does not automatically transfer* to the open-ended, multi-step, guess-resistant problems that define frontier science. The frontier demands something beyond pattern-matching to textbook solutions.

Why you should care about the follow-ups

This opens natural questions: What kinds of reasoning do physicists specifically want AI to help with? Where does AI fail most badly—at conceptual understanding, mathematical derivation, or creative problem-solving? And can we design AI systems that genuinely augment human research rather than merely mimicking it?

What would you like to dig into next?