After the Turing Test (blog.judicata.com)

The piece argues that while large language models have effectively “passed” the original Turing Test by mimicking human conversation, that milestone doesn’t equate to genuine, generative, genius‑level intelligence. Using Fermat’s Last Theorem and Andrew Wiles’s centuries‑late proof as a lens, it shows the vital distinction between an AI asserting a novel, high‑stakes mathematical claim and an AI producing a convincing, verifiable proof that convinces expert skeptics. Truth in science and math depends not on declarative statements alone but on reproducible evidence and arguments that human peers can evaluate; absent that, an AI’s novel claim is as untrusted as any unproven conjecture. To bridge that gap the author proposes a raised bar — a variant of the Imitation Game tailored to elite intellectual tasks (an “Andrew Wiles Test”): can an AI convincingly persuade domain experts of a novel, complex claim by supplying rigorous, inspectable evidence? For ML researchers this implies shifting benchmarks toward verifiability and novelty detection: formal proofs, mechanized theorem‑proving, provenance and training‑data transparency, expert‑in‑the‑loop evaluation, and reproducible verification pipelines. The upshot: measuring AI by ordinary human performance won’t suffice; we need tests that demand demonstrable, auditable contributions at the level of human genius.