Gary Marcus says today’s systems are not close to AGI
GaryMarcus · x · 2026-07-24
Gary Marcus says we have not reached AGI and argues we are not even close.
He points back to his December 30, 2024 bet with Miles Brundage and says a true AGI would need to be able to do nearly all of the listed tasks:
- Follow complex movie and novel plots without hallucinating
- Write reliable biographies and obituaries grounded in sources
- Learn new video games quickly and solve original puzzles
- Draft persuasive legal briefs without inventing cases
- Produce 10,000+ lines of bug-free code from natural language
- Write Pulitzer-, Oscar-, and Nobel-caliber work with little human help
- Convert math proofs into symbolic form for verification
His claim is that nobody has systematically shown success on even one of these in the way a real AGI would require.
More from AGI Musings
- NeurIPS 2026 workshop will spotlight failure modes of AI in biology — anshulkundaje · 2026-07-24
- Metascience could reshape the scientific core, not replace it — anshulkundaje · 2026-07-24
- Frontier models still depend on stable university research and global talent — anshulkundaje · 2026-07-24
- AI industry pulls professors out of academia as research funding shrinks — ruthstarkman · 2026-07-24
- Token efficiency comes from search and validation, not just implementation — yunta_tsai · 2026-07-24
- Open-weight frontier models could become dangerous if they can be jailbroken and used anonymously — Afinetheorem · 2026-07-24