Researchers Clash Over Whether AI Has Already Solved 'Intelligence'

On October 28, a debate erupted in the AI research community over whether frontier models have substantially solved the "intelligence problem." It was sparked by one developer's bold claim: based on a recent series of results, OpenAI's internal models are already smart enough to solve the vast majority of problems people care about—the "intelligence" part is done, and what remains is just engineering work like proper integration and data collection.

Confirmed

• fast.ai founder Jeremy Howard responded that current models fall short in his day-to-day work: whenever he tries to build something novel and unlike anything before, the models' designs are "always exactly the wrong answer"
• Anthropic researcher Blanche Minerva said in the same discussion that currently available models cannot solve math problems at the level of ETH or the Quasi-Riemann hypothesis; she noted the models capable of such problems are unreleased, and even the available models Howard referred to in his work can't do it. She also said that in her own research, models are useless for anything but writing code
• Researcher François Fleuret responded to the tweet with sarcasm: if the "intelligence problem is solved," we should celebrate with fireworks and music, and asked whether its "score" could beat the World Cup

Why it matters

The debate reflects the polarized assessments of frontier model capabilities in today's AI community: one side believes model intelligence has crossed a critical threshold and the rest is engineering; the other, drawing on hands-on research and development experience, argues models still systematically fail on genuinely novel designs and frontier mathematics. Notably, both sides cite the performance of "internal/unreleased models" as evidence, showing that the capability gap between publicly available and frontier internal models is itself a point of contention. For judging progress toward AGI, such negative feedback from practitioners deserves as much attention as optimists' claims.

2026-10-28 ~ 2026-10-28 · 5 related posts

• [source] "The intelligence part is solved": dev argues OpenAI's internal model effectively is AGI — BlancheMinerva · 2026-10-28
• [source] Jeremy Howard: Current Models Reliably Give the Wrong Design for Anything New I Want to Build — jeremyphoward · 2026-10-28
• Is AGI Already Here? Researchers Spar Over 'Intelligence Is Solved' — francoisfleuret · 2026-10-28
• Unreleased model reportedly solves ETH and Quasi-Riemann problems that current models can't — BlancheMinerva · 2026-10-28
• [source] Anthropic researcher: current models can't solve ETH or Quasi-Riemann, useless beyond coding — BlancheMinerva · 2026-10-28