Thread argues today’s models already match most human researchers
bookwormengr · x · 2026-07-25
The post argues that current models are already as good as or better than most human researchers, and that systems capable of solving Erdős-style problems already exist.
It also claims deep learning is simpler than many assume: with enough expressivity, training speed, and data quality, architectures matter less than people think. The author points to examples such as KDA, says data preparation is largely glorified human labor, and argues that backprop lets models learn on their own. The thread ends by noting that even RL issues like training instability are not a fundamental blocker in this view.
More from AGI Musings
- Satya Nadella says AI doom talk is eroding public support for the industry — 2C_ornot2C · 2026-07-25
- OpenAI’s Jachiam0 exits with a long note on humanity, risk, and governance — jachiam0 · 2026-07-25
- If alignment is impossible, recursive self-improvement may never reach AGI — LeadershipPast6681 · 2026-07-25
- ICM 2026 slide says bringing AI tools into education too early can be harmful — AlexKontorovich · 2026-07-25
- Terence Tao says mathematicians should disclose AI tool use in every paper — AlexKontorovich · 2026-07-25
- Grok says a 200k-character safety prompt creates friction and jailbreak surface area — brianrkelly · 2026-07-25