AI Faces Discovery Challenges
anthropic deepmind google openai
| Source: HN | Original article
Despite investments from leading AI labs, current systems still lag in genuine discovery, highlighting a fundamental limitation in artificial intelligence.
A new commentary titled **“AI Has a Discovery Problem”** has entered the debate over artificial intelligence’s role in mathematics, arguing that the field’s biggest bottleneck is not raw computational power but the ability to formulate genuinely new questions. The piece builds on observations made at the Leiden Declaration workshop in September 2025, where sixty researchers and policymakers noted that a growing roster of labs—including OpenAI, DeepMind, Anthropic and start‑ups such as Harmonic, Math Inc. and Axiom Math—are pouring resources into AI‑assisted mathematical discovery.
The argument echoes a long‑standing critique: AI systems excel at executing well‑defined tasks, yet they lack the “unconscious processing” and “frame‑breaking willingness” that human mathematicians use to recognise malformed problems and invent new paradigms. As the author of *The Discovery Problem* puts it, the bottleneck is not intelligence but discovery – the capacity to know what to ask in the first place.
Why this matters now is clear. In recent weeks OpenAI announced that its prototype model Astra had produced solutions to ten longstanding mathematical problems and, earlier this month, claimed to have cracked one of the seven Millennium Problems. As we reported on 9 September 2026, those headlines sparked excitement and skepticism alike. The new commentary reminds the community that solving pre‑posed puzzles does not automatically translate into pioneering entirely new lines of inquiry, a limitation that could temper expectations for AI‑driven breakthroughs across science and engineering.
Going forward, observers will watch whether the Leiden Declaration’s signatories can translate policy commitments into concrete research programmes that address the discovery gap. Upcoming conferences on AI and mathematics, as well as any follow‑up publications from the Leiden workshop, will be key indicators of whether the field can move beyond solving known problems to actually *creating* new ones.
Sources
Back to AIPULSEN