FULFILLEDStated December 25, 2022
Gary MarcusAI researcher; author of “Marcus on AI”
Seven predictions about GPT-4: it would still hallucinate, remain unreliable at physical/psychological/mathematical reasoning, not give reliable medical advice, and not be safely hooked up to downstream programs.
LLM limitsreliability
- Verdict
- Came true
- Deadline in claim
- On GPT-4's release (2023)
Assessment
GPT-4 and its successors still hallucinate, remain unreliable on hard reasoning, are not trusted for unsupervised medical advice, and need guardrails downstream. Widely judged correct. Scored fulfilled — a skeptical call that held.
Put this through The Wringer →Primary sources
Links open the original publications. Quotations are reproduced for the purpose of commentary and criticism. Spot an error? Corrections policy.