How should AGI be defined?
7 recorded positions from 7 people, first said Jun 30, 2023. They do not agree — the readings below are what each one actually argued.
Also on the record
Edwin Chen · Jul 21, 2025
AGI arrives around 2028 if defined as automating the job of the average engineer, and around 2038 if defined as curing cancer
56:16 Agi timeline depends entirely on definition 2028 for engineering 2038 for curing cancer
Douwe Kiela · Jun 30, 2023 · hedged
On OpenAI and Anthropic's definition of AGI — systems capable of doing the majority of economically valuable human tasks — we are within five to ten years of that economic displacement
Once AGI is defined in capability/economic terms rather than consciousness terms, current systems are not far from that bar
37:10 Economic task based agi definition puts displacement within five to ten years
Richard Socher · Aug 18, 2023
A system that only predicts the next token should not be called generally or super intelligent; genuine intelligence requires having its own goals
Pure token-prediction maximization would be satisfied by everyone only ever saying 'a', which is not an interesting or fulfilling intelligent existence
39:22 Genuine intelligence requires having own goals not just next token prediction
Harry Stebbings · Jul 28, 2025 · hedged
The operative definition of AGI is whatever Sam Altman and Microsoft decide it is
58:01 Agi definition is whatever openai and microsoft contractually decide
Nick Frosst · Sep 1, 2025
We are not at AGI today
People do not treat language models the way they treat people
13:04 People not treating models as people shows we havent reached agi
Demis Hassabis · Apr 7, 2026
AGI should be defined as a system that exhibits all the cognitive capabilities the human mind has
The brain is the only existence proof we have that general intelligence is possible, so it sets the bar
4:46 All human cognitive capabilities is the bar
Ethan Mollick · Jul 31, 2024
The useful version of the question is concrete capability specification — e.g. can it do a full legal review of a document, or write an academic paper end to end from a dataset
Concrete tasks decompose into answerable questions about hallucination rates, token-vs-word handling, precedent, multilingual ability and context window size
42:35 Concrete task specification beats abstract agi debate
Your assistant can query this graph directly — 7 positions here, 19,646 across the corpus. Add 996.fm over MCP.