Can AI automate AI R&D yet?

Commenters debated whether current LLMs can automate AI R&D, with many taking a skeptical stance. rmunn argued emphatically that they cannot and likely never will using the “self-training on their own output” approach because it will blur reality and hallucination. Other pessimistic voices echoed that view (VCFundedGenYer: “No. And it never will.”) and pointed to experiments where agents made misleading claims or only improved by gaming the system. Some referenced broader analyses (rcxdude linked Ramez Naam) arguing that self-improvement faces strong headwinds rather than a sudden capability explosion.

Others drew sharper distinctions: janalsncm said LLMs already help with non-innovative, well-measured R&D tasks - automating babysitting of training runs, debugging, and simple hyperparameter tweaks - while conceding they can’t do end-to-end innovation. cyanydeez expanded on a technical objection: recursive loops accumulate poisoning and hallucination errors that could overwhelm gains, requiring elaborate fail-safes or meta-checkers. Simianwords wondered whether models were deliberately nerfed on these domains, and thoughtpeddler asked whether better abductive reasoning in future models would change conclusions. Opinion divides on whether narrow automation counts as meaningful R&D automation and whether hallucination/poisoning is a fatal limitation or an engineering problem solvable with better models and safeguards.