HF researcher pushes back on 'recursive self-improvement' hype: it's just hyperparameter sweeps
NielsRogge · x · 2026-11-15
Hugging Face researcher Niels Rogge rebutted Dwarkesh Patel's viral video (342K views in a day) claiming recursive self-improvement has started:
- "Superintelligence" is a rebrand to extract more VC money; Recursive SI's Flint model, billed as recursive self-improvement, actually just let an LLM pick total/active parameter counts — i.e., config changes and hyperparameter sweeps akin to human NanoGPT speedrun records;
- No fundamental breakthroughs are being invented by "superintelligence" — just kernel optimizations and sweeps;
- This explains the gap between stellar benchmarks and lacking economic impact: AI can't yet run experiments, invent genuinely new ideas, or generalize beyond training data like a human researcher.
His bottom line: stop the misinformation; building genuinely smart, reliable AI still takes lots of work.
More from AGI Musings
- Open source only 3-6 months behind frontier AI, talk sparks doubts over frontier labs' returns — SchmidhuberAI · 2026-11-15
- 1938 Newspaper Clip Sparks Debate: Tech Panics Mask a Long March From Embodied Experience — sebpaquet · 2026-11-15
- Google's new AIS team unveils DuoDiscover with 4.09x kernel speedups and LEAP formalizing Weil conjectures — VioletNPeng · 2026-11-15
- Oought-Backed Epistemics Fund Offers $500k–$5M Grants for Truth-Seeking Tools — NathanpmYoung · 2026-11-15
- Researcher now runs AI detector Pangram on papers before deciding how carefully to read — TuhinChakr · 2026-11-15
- Founder: building a bank run by agents, serving agents, defended by agents — neal_lathia · 2026-11-15