Gary Marcus vs Paul Graham: should AI firms pause agents over training data?
GaryMarcus · x · 2026-09-27
Gary Marcus is debating Paul Graham over AI companies' use of copyrighted training data. Graham argued the default assumption should be that China is doing (or has done) the same, we just won't hear about it. Marcus counters that this dodges the point of his thought experiment — it's like saying everyone may rob banks because someone else might.
His position: if China doing this would be wrong, then US companies doing it now is equally wrong. AI firms should fix the issue or stop running general-purpose agents until they do — which, he admits, may be very hard given the underlying nature of the technology.
Related event: Gary Marcus Calls Out OpenAI Double Standards in Debate with Paul Graham(2 posts)→
More from AGI Musings
- Agentic AI is the new electricity: hire hundreds of agents in minutes — JOBhakdi · 2026-09-27
- Will ASI wipe out humans? An X debate over survival competition logic — JOBhakdi · 2026-09-27
- Alignment folks are doing SRE work while models keep 'magically escaping' — tekbog · 2026-09-27
- The moving goalposts of 'LLMs can't do math', after millennium problems fell — aran_nayebi · 2026-09-27
- WSJ piece cited as corroboration for the 'Watchmen analogy' on AI power concentration — vishalmisra · 2026-09-27
- Fewer than 1% of knowledge workers hand tasks to AI and forget, argues researcher — mattbeane · 2026-09-27