Top models may already be smart enough for RSI, but data and compute still block self-improvement
bookwormengr · x · 2026-07-26
The post argues that today’s top models may already be smart enough for RSI, and that the real bottlenecks are not intelligence itself but data and compute.
- The author claims current models are at least as intelligent as the best young AI researchers.
- He says many of the field’s biggest ideas have come from very young researchers, not veterans.
- He argues there is no deep theory of deep learning; the field is mostly an experimental collection of tricks.
- In his view, models are not improving themselves because better models require more and better data plus more compute, and both face real-world constraints.
- The quoted reply frames the strategic implication: if RSI is close, intelligence becomes economically and geopolitically decisive.
Related event: Engineering Path and Bottlenecks of AI Recursive Self-Improvement(3 posts)→
More from AGI Musings
- U.S. traffic deaths hit 36,640 last year as Waymo-style vehicles promise further declines — SuzKP · 2026-07-26
- Why one AI writer says the uncanny valley helps us read machine-made text — HamelHusain · 2026-07-26
- Oxford study finds AI can out-persuade world-champion debaters in controlled tests — Justgototheeffinmoon · 2026-07-26
- Power outage knocks 3 GW of data centers offline as China’s AI progress pressures US spending — emmanuelvivier · 2026-07-26
- Brown in-person exam scores fell 50%, pushing schools to rethink assessment in the AI era — emmanuelvivier · 2026-07-26
- Musk says AI may beat all human intelligence in 5 years and humans could lose control in 10 — theomitsa · 2026-07-26