Researcher: no evidence frontier models can write good neural nets from scratch or read weights
paul_cal · x · 2026-11-05
Researcher paulcal raises two skeptical points about the current RLVR-driven agenda:
- There's no evidence frontier models can write good neural nets "from the void" that could stand in for typical neural net use cases.
- Nor is there evidence yet of zero-shot reading, writing, or understanding of "giant inscrutable matrices" (model weights).
His conclusion is sardonic: the field keeps RLVR-ing anyway. A representative skeptic take on limits of model self-improvement and automated research.
More from AGI Musings
- Knights of Guinevere plays 'do AI deserve moral respect?' straight as a 2026 plot point — 1a3orn · 2026-11-05
- Agentic commerce splits two ways: network tokens vs. agents with their own stablecoin wallets — schwentker · 2026-11-05
- LeCun joins MAIN 2026 world model panel debating whether dorsal/ventral separation needs architectural biases — soniajoseph_ · 2026-11-05
- Joke: low-information voters found an anti-AI strategy more effective than terrorism — menhguin · 2026-11-05
- You can't stop AI: data centers will just move to Malaysia, India or Kazakhstan — KyeGomezB · 2026-11-05
- Ex-Microsoft CVP: AI hides who caught the flawed assumptions—keep a record of your influence — BenBajarin · 2026-11-05