What Is
AGI?
Artificial general intelligence: why every major AI lab is racing towards it, where current AI actually falls short, and what AGI would mean for every industry on earth. The hype stripped away.
00 — The definition
AGI — artificial general intelligence — is AI that can perform any intellectual task a human can perform, without being specifically trained for each task. The key word is "general." Everything we have today is narrow.
Current AI systems are impressive but deeply specialised. GPT-4o can write essays, debug code, and analyse images. But it can't drive a car, perform surgery, or adapt to a task it's never encountered in any form. AlphaGo defeated the world's best Go player — but it can't play chess. AlphaFold predicts protein structures with extraordinary accuracy — but it can't do anything else. Every AI system in production today is narrow AI: expert at its specific domain, useless outside it.
AGI is different in kind, not degree. It's not "a better ChatGPT." It's a system capable of generalising across arbitrary tasks — learning something in one domain and applying that learning to a completely different domain, the way humans do constantly without thinking about it.
The clearest definition comes from Anthropic: AGI would be "an AI system that is broadly as cognitively capable as a human across all intellectual domains." ANTH OpenAI defines it as "AI systems that are generally smarter than humans." OAI Both definitions are contested — there's no agreed technical benchmark for AGI, which is itself a significant problem for the field.
01 — Where current AI falls short of AGI
Understanding the gap isn't pessimism — it's necessary context for interpreting the claims coming out of AI labs every month.
| Capability | Human performance | Current best AI | AGI threshold |
|---|---|---|---|
| Task transfer | Trivially transfers learning across domains | Requires domain-specific training; limited transfer | Not reached |
| Causal reasoning | Builds causal mental models natively | Pattern matching, not causal understanding | Not reached |
| Long-horizon planning | Plans weeks, months, years ahead | Limited to context window; degrades quickly | Partially reached |
| Common sense | Implicit, vast, and reliable | Frequently fails on edge cases | Not reached |
| Self-directed learning | Humans seek information, update beliefs | Requires re-training; can't self-update reliably | Not reached |
| Narrow task performance | Variable; exceptional in trained domains | Superhuman in specific benchmarks | Exceeded in many areas |
| Physical world interaction | Full embodied cognition | Robotic systems improving but brittle | Not reached |
OpenAI claimed in January 2025 that o3 had achieved "human-level performance across cognitive tasks" on several benchmarks. OAI But benchmark performance on curated test sets isn't the same as general intelligence. ARC-AGI benchmark scores improved dramatically — yet the tests themselves measure specific reasoning patterns, not the full breadth of human cognition. The lab's own researchers caution against interpreting these as AGI evidence.
02 — What OpenAI, Anthropic, and DeepMind actually say
Stripping the press releases: here's what the lab founders and researchers say in their own words.
"We may be only a few thousand days away from AGI... we are going to live in a world with fantastically more intelligence available."
Blog post, February 2025. ALTM
"I think there's a real chance we're going to build something very close to AGI in the next few years... the pace of progress has been shocking even to those of us in the field."
Lex Fridman Podcast, 2024. DARIO
"We may be approaching something like AGI within the decade... but we need to be extremely careful about how we develop it — the alignment problem is real."
Wall Street Journal interview, 2024. DMD
All three of the most credible AI lab CEOs believe AGI is years away, not decades. That's a significant shift from the consensus five years ago. And all three also emphasise that the risks of getting it wrong are catastrophic.
The uncomfortable reality: these are the people building it, with the most access to what's happening inside the models, and they're all sounding both excited and genuinely worried. That combination deserves to be taken seriously.
03 — AGI timelines: what the evidence suggests
Timeline predictions have a terrible track record in AI. But the rate of capability improvement since 2022 is different from anything that came before it.
GPT-4 passes bar exam (90th percentile), SAT (93rd percentile), AP exams. AI researchers expected these milestones 5+ years later. OAI
OpenAI o1 and Anthropic Claude 3.7 with extended thinking demonstrate multi-step reasoning that dramatically improves on prior models. ARC-AGI scores jump from under 10% to over 85%. ARC
Models complete multi-hour software engineering tasks, conduct research, and operate computers autonomously. Devin, Operator, and Claude Computer Use mark a shift toward agency.
Most AI researchers surveyed in 2024 placed >50% probability on AGI arriving by 2030. AIIM The "few years" claims from Altman and Amodei point to this window. Nothing is certain, and "AGI" remains undefined enough that any claim is contestable.
04 — The questions people actually ask
Nobody knows. Expert estimates range from "already here in a limited sense" (some OpenAI researchers) to "decades away" (sceptics like Gary Marcus). A 2024 survey of ML researchers found a median estimate of 2047 for AGI, but with enormous variance. AIIM The rate of capability improvement since 2022 has made earlier estimates obsolete in both directions — things moved faster than pessimists predicted, but some claimed milestones didn't generalise as hoped.
OpenAI has not announced AGI. Some researchers inside the lab have suggested o3 demonstrated "AGI-like" capabilities on specific benchmarks — but the company has not made an official AGI declaration. Under OpenAI's charter, the board would be notified if AGI was reached; that notification has not happened publicly. The definitional ambiguity is doing a lot of work here: if you define AGI narrowly (human-level on a set of benchmarks), you can claim we're close. If you define it broadly (genuine general reasoning), we're not there.
This is a genuinely contested question. The "alignment problem" — ensuring that a superintelligent AI pursues goals that are beneficial to humans — is considered an important unsolved problem by researchers including those at Anthropic, DeepMind, and Miri. The risk isn't the science-fiction scenario of a robot that decides to exterminate humans. The more technically grounded concern is an AI system that pursues its given objective in ways humans didn't intend, at a scale and speed that makes course correction difficult. Whether this constitutes an "existential risk" or a "manageable technical problem" is where the serious debate lies.
And nobody actually knows the answer yet.
The people building the most advanced AI systems believe they're years from AGI, not decades. That belief is based on internal knowledge we don't have access to. Whether their optimism is warranted or premature remains to be seen — but the capabilities we can observe keep crossing thresholds faster than anticipated.
The reasonable position is this: something significant is happening, the pace is accelerating, the implications are enormous, and the uncertainty is genuine. Neither dismissal nor panic serves you. What serves you is following the developments closely enough to understand what's real and what's hype, as the evidence develops.
AI intelligence,
weekly.
Every week: the AI developments that matter, the tools worth trying, and the data behind the headlines. No hype. No filler.
Subscribe free →