What is AGI?
AGI
AGI is a concept referring to artificial intelligence that matches or exceeds human-level performance across a wide range of cognitive tasks, as discussed in the context of benchmarks like ARC-AGI and debates about models like OpenAI’s o3. The term has been applied to specific models, but definitions and timelines vary among experts, with recent debate focusing on whether large language models can lead to general intelligence.
How it developed
- Dec 2025 – Tyler Cowen called OpenAI’s O3 model AGI, noting it can’t do every job but meets the old definition of being as good as human experts.
- Dec 2025 – Arvind Jain stated, “I think we have AGI. I think we have artificial general intelligence. We really have it.”
- Dec 2025 – Mike Israetel predicted AGI would arrive in 2029, 2030, or 2031.
- Apr 2026 – Demis Hassabis estimated the number of AGIs at “20 or 30” and said his betting is “about 50/50” on whether more than one or two exist.
- Jul 2026 – François Chollet affirmed that AGI is possible.
- Jul 2026 – An unnamed speaker noted that a year ago most disagreed with the idea that LLMs don’t lead to general intelligence, but now a good chunk agree they don’t.
- Aug 2026 – Garry Tan said, “AGI isn’t arriving as an event. It’s arriving diffused as your agent running on your context doing your work.”
In the evidence
Every line below is attributed to a named speaker.
Tyler Cowen declares the O3 model qualifies as AGI under the longstanding benchmark definition.
“Maybe the biggest surprise in AI was how good the O3 model was, which I called AGI at the time. Maybe tongue-in-cheek, but I also meant it, that it can't do every job. The old definition of AGI, that it's just as good as human experts, is largely true.”Tyler Cowen · 23 Dec 2025
Nathan Lambert on OpenAI's strategic pivot away from AGI premiumization.
“It's like if they're really going for price as the differentiator here. It's like it's it's a very different game than saying you're going to build AGI and charge $20,000 a month for it.”Nathan Lambert · 7 Aug 2025
Demis Hassabis estimates AGI is 20 to 30 years away, a notably longer horizon than most public AI lab timelines.
“Mine's like 20 30 or something like this.”Demis Hassabis · 29 Apr 2026
Garry Tan on AGI arriving not as an event but as a personal agent
“AGI isn't arriving as an event. It's arriving diffused as your agent running on your context doing your work.”Garry Tan · 6 Aug 2026
Chollet argues it is possible to score well on ARC-AGI-3 without being any closer to AGI, undermining the benchmark's framing as a reliable AGI proxy.
“Yes, I do think it is possible.”François Chollet · 1 Jul 2026
Demis Hassabis puts the probability at roughly 50/50 that existing AI techniques can scale to AGI without any new fundamental ideas, with only one or two such ideas remaining if new ones are needed.
“I don't think it's more than one or two if there are out there. And I think, you know, my betting is about 50/50 if that's the case.”Demis Hassabis · 29 Apr 2026
Arvind Jain asserts we already have AGI today.
“I think we have AGI. I think we have artificial general intelligence. We really have it.”Arvind Jain · 23 Dec 2025
A year ago most AI practitioners believed LLMs could lead to general intelligence. Now a significant portion have shifted to disagreeing with that view.
“A year ago when I when we were saying the LLMs are nice and powerful but they're not AGI. They don't lead to most people disagreed now I think most people agree that LLMs don't lead maybe not most people but a good chunk of people agree that LLMs don't lead to a general intelligence.”<UNKNOWN> · 25 Jul 2026
Mike Israetel argues ASI will arrive before AGI, predicting ASI by 2026 to 2027 and AGI only by 2029 to 2031, inverting the conventional ordering.
“I think ASI is coming in 2627 and AGI is coming in 29, 30, maybe 31.”Mike Israetel · 24 Dec 2025