AI math breakthroughs keep proving earlier limits wrong, researcher says
herbiebradley · x · 2026-07-24
AI researcher Herbie Bradley argues that most AI math breakthroughs so far have actually been counterexamples to earlier assumptions.
- He says many people hope RSI will come from a 2x–3x jump in sample efficiency, but RL is still not very sample-efficient on long-horizon tasks.
- In his view, current systems still look like transformer-based pattern learners with variations, not something that truly models scientific discovery.
- He adds that models remain noticeably worse at novel discovery, intuition-based R&D, and innovation than human researchers.
- The math breakthroughs we’ve seen are often counterexamples to claims about what models supposedly cannot do.
More from AGI Musings
- Which jobs fall to AI first? The rule: any layer whose rules can be written as code — FinanceYF5 · 2026-10-26
- Anthropic S-1: expert-level AI 'within a few years', 90% of code written by Claude — kimmonismus · 2026-10-26
- The identity crisis in software engineering: multiagent systems demand a new problem set — rakyll · 2026-10-26
- Frontier LLMs negotiate contracts poorly and frequently renege for selfish gain, study finds — xuanalogue · 2026-10-26
- Study of 133 patent lawyers: AI augments veterans but causes cognitive offloading in novices — xiaohu · 2026-10-26
- OpenAI computer-use agents near superhuman speed, say the team building them — a16z · 2026-10-26