What Is AGI, and How Would We Know We Built It?
In August 2026, OpenAI's own chief research officer said the company was about eighty per cent of the way to AGI. Its chief executive said 'not quite yet.' Across town, Anthropic's chief executive avoids the term altogether. Three leaders, inside weeks of each other, could not agree on one word.
This video traces that word from its origin: a 1997 security paper that used 'artificial general intelligence' to describe a machine that could replace a soldier in any role, through a 2002 research claim by Shane Legg and Ben Goertzel that gave it a different meaning entirely.
Then it covers the three competing modern tests. OpenAI's charter defines AGI as a labour-market threshold, judged by its own board. A clause tying that definition to billions of dollars in Microsoft's contract was redefined in October 2025 and dropped entirely on 27 April 2026. Google DeepMind's 'Levels of AGI' paper proposes five levels, from Emerging to Superhuman, scored on depth and breadth. The ARC Prize Foundation's ARC-AGI-3 benchmark, launched 25 March 2026, had every leading AI model scoring under one per cent at launch, while humans scored 100%.
Run one single capable system against all three tests, and it gets three different answers. The video closes on Anthropic's $11.5 billion Q2 2026 revenue overtaking OpenAI's $6.7 billion, and its planned IPO, before asking the question the title promises: is the industry converging on one measurable finish line, or keeping the word elastic enough that everyone can say they have already crossed it?
Every figure is on screen with its source and date.
In these topics
Tags
Chapters
- A word with no agreed meaning
- 1997: a soldier's word for a battlefield problem
- 2002: the researchers claim the word
- OpenAI's legal definition: economically valuable work
- The clause that turned a definition into money
- DeepMind's ladder: depth and breadth
- The ARC Prize: can it learn, not just recall
- Three tests, three different winners
- August-September 2026: the claims collide
- Anthropic's rise and the stakes of the label
- Why no one wants to be the referee
- Finish line or moving target
Sources and credits
Photo credits (Wikimedia Commons)
- Sam Altman (OpenAI CEO) and Dario Amodei (Anthropic CEO) together at a White House AI meeting, 4 May 2023: The White House, Public domain - https://commons.wikimedia.org/wiki/File%3AV20230504LJ-0199-2.jpg
- Sam Altman, OpenAI CEO, White House meeting, May 2023: The White House, Public domain - https://commons.wikimedia.org/wiki/File%3APresident_Joe_Biden_drops_by_a_meeting_with_Vice_President_Kamala_Harris_and_AI_CEO%E2%80%99s_on_May_4%2C_2023_in_the_Roosevelt_Room_of_the_White_House_-_P20230504AS-0693.jpg
- Sam Altman, OpenAI CEO, White House dinner for tech leaders, 4 September 2025: The White House, Public domain - https://commons.wikimedia.org/wiki/File%3AF20250904AH-2634_%2854778707580%29.jpg
Primary sources
- Mark Gubrud, 'Nanotechnology and International Security,' Fifth Foresight Conference on Molecular Nanotechnology, Palo Alto, Nov 1997, via Wikipedia's 'Artificial general intelligence' history section and a corroborating secondary account.
- Wikipedia, 'Artificial general intelligence' (history section) and 'Shane Legg' - Legg and Goertzel credited with popularising 'AGI' from 2002.
- Morris et al., 'Levels of AGI: Operationalizing Progress on the Path to AGI,' arXiv:2311.02462 (Google DeepMind, Nov 2023).
- ARC Prize Foundation, arcprize.org, ARC-AGI-3 technical report - benchmark concept published 2019 by Francois Chollet; ARC-AGI-3 launched 25 Mar 2026.
- Microsoft, 'The next chapter of the Microsoft-OpenAI partnership,' blogs.microsoft.com, 28 Oct 2025.
- Multiple reports, 27 Apr 2026 (aitoolly.com; letsdatascience.com; simonwillison.net's dated timeline), on the AGI clause being dropped from the Microsoft-OpenAI contract entirely.
- TIME, 'Inside OpenAI's Reboot,' 26 Aug 2026 - Mark Chen's '80% of the way' estimate and Sam Altman's 'not quite yet' statement, corroborated by multiple outlets quoting identical wording (TIME itself paywalled).
- Dario Amodei, 'Machines of Loving Grace' essay, Oct 2024, and Anthropic's OSTP submission - 'powerful AI' defined as 'a country of geniuses in a datacenter,' timeline 'as early as 2026 or 2027.'
- Company disclosures reported Aug 2026 (Fortune, 15 Aug 2026; Axios, 17 Aug 2026; Yahoo Finance) - Anthropic Q2 2026 revenue $11.5bn vs OpenAI $6.7bn.
- Anthropic S-1 filing, 1 Jun 2026, and reporting through 21 Sep 2026 on a possible slip from an October to a November 2026 Nasdaq listing.
Not regulated financial advice.