Topic
AGI
Every frontier leader now agrees AGI is coming this decade. They are quietly terrified of completely different things... and a few of them stopped using the word at all.
The strangest thing about how the people building artificial general intelligence talk about it is how little they argue about whether. Shane Legg has held a 50-50 bet on minimal AGI by 2028 since 2009, back when, in his own words, the idea lived on "the lunatic fringe" (Google DeepMind podcast, Dec 2025). Dario Amodei puts the odds of a "country of geniuses in a data center" within a decade at 90 percent, and calls anything lower "outside the mainstream" (AI Upload, Feb 2026). Ilya Sutskever reduces the whole question to one sentence: "we have a brain and the brain is a biological computer," so a digital one can do the same things (Singularity Watch, Jan 2026). The consensus is so total it has stopped being interesting. What's interesting is where, beneath that agreement, the floor cracks.
The timelines converged, then the definitions splintered
They all point at the same horizon and describe wildly different objects sitting on it. Mira Murati defines AGI as a system "capable of learning how to perform at human level across all cognitive tasks" and says current models already show PhD-level performance in many domains... but she's the cautious one, pricing the arrival at "perhaps a decade, maybe two, I don't know" (WIRED, Dec 2024). Shane Legg refuses the binary entirely. AGI isn't "a flip-a-switch moment"; it's a spectrum from minimal AGI (defined by reliability, not genius... an AI that stops making "toddler-like mistakes") through full AGI to superintelligence (DeepMind podcast, Dec 2025).
The honest admission underneath all of it is that the models are uneven. Legg keeps returning to the word: today's AI speaks 150 languages yet can't reliably reason about which car in a picture is bigger, or learn a new card-game rule, or do continual learning over time the way a new hire does. Amodei makes the same cut from the other side, separating "tasks that can be verified" (coding, where he's near-certain) from the unverifiable (planning a Mars mission, writing a novel, fundamental discovery) where "a little bit of uncertainty" survives even on long timescales. The disagreement isn't about the trendline. It's about whether the jagged bits get smoothed in two years or ten.
Why they're sure: the brain is just a slow computer
The conviction rests on two load-bearing claims, and they're worth separating because the leaders lean on them unequally. The first is scaling. Amodei dates his to 2019, watching GPT-2: "this is much more likely than anyone thinks it is. This is wild." It was the conviction he says he couldn't fully sell at OpenAI, and one reason he left to found Anthropic. Murati is more guarded, calling the scaling law "not literally a law, but an observation," and conceding that getting to AGI may need genuinely new ideas, not just more compute.
The second claim is cruder and, for the believers, decisive: biology is just unimpressive hardware. Sutskever's biological-computer argument is the soft version. Legg's is the brutal one: the human brain runs on 20 watts with signals crawling at 30 metres per second; a data centre burns megawatts at the speed of light. "Biology has limits. Silicon doesn't." Which is why none of them think AGI is a finish line. It's a waypoint they expect to blow straight past into superintelligence.
Greg Brockman adds the recursive twist that makes the believers nervous and excited at once. OpenAI's early bet, he says, was that "the path to AGI is really about ideas," not compute... a bet they reversed around 2017. Now they're pointing the models back at the hardware: using their own AI to design the Broadcom chips, finding "massive area reductions" their human experts had on a list but would have taken another month to reach (OpenAI podcast, Oct 2025). The machine optimising the machine that builds the machine. Sutskever names it plainly: recursive self-improvement, after which "the rate of progress will become really extremely fast."
The economics: from "bloodbath" to "it's the speed, stupid"
This is where 2025-2026 actually shifted, and it's a retreat dressed as a refinement. Amodei spent 2025 owning the "white-collar bloodbath" headlines... entry-level jobs, data entry, paralegal work, first-year document review, all drying up. By February 2026 he's recanting the framing while keeping the fear. The new model is the centaur: like post-Deep-Blue chess, there's an era where a human checking the AI's output beats any human or any machine alone, then "it's just the machine." Crucially, he now argues the threat isn't displacement but velocity. "The diffusion to the economy is going to be a little slower," he says, and that lag "creates some unpredictability." Software went fast not because models are better at code but because developers "are used to fast technological change and they adopt things quickly." The horror isn't that the work vanishes. It's that it vanishes "over low single-digit numbers of years" instead of the centuries it took to go from farm to factory to office.
And the carnage is no longer hypothetical in his telling: a single Anthropic blog post wiping $31 billion off IBM, CrowdStrike down 8 percent in hours, the software ETF off 27 percent on the year (AI Nutshell, Feb 2026). Whether you buy the causation, that's the CEO of a leading lab citing market damage as evidence his product works.
Sam Altman runs the optimistic counter-program. His frame is abundance: "we built up a lot of instincts and institutions, policies, structure to deal with a world of scarcity. But almost none of that applies well to a world of abundance" (IIT Delhi, Feb 2026). He thinks GDP "is going to turn out to be a terrible metric because AI is so deflationary", predicts a "hugely deflationary economy that the world isn't thinking about" by 2035, and tells students his one-year-old "is never going to know a world where he was smarter than a computer." Where Amodei sees a disruption arriving faster than society can metabolise, Altman sees the metabolising as the whole opportunity.
Jack Clark splits the difference into something genuinely uncomfortable. His through-line is uneven impact (Conversations with Tyler, Feb 2026). Some work resists not because AI can't do it but because institutions won't let it: he uses Claude on his own baby's bumps, yet "can't take that Claude assessment and give it to Kaiser Permanente." The emerging job, he says, is laundering AI-generated information "into human systems that are not predisposed to that information going in directly." And he predicts the political endgame with a phrase worth keeping: a movement to freeze "a load of human jobs in bureaucratic amber," driven "by the chaotic winds of political forces," not reason. He doubts it works, because he doubts those protected jobs would generate real meaning.
The two safety camps, and the one that markets serenity
On risk, the lab affiliations show through the seams. Legg, DeepMind's Chief AGI Scientist, has the most engineered answer: "System 2 safety," borrowing Kahneman, building slow deliberate ethical reasoning into the model so it can't act on its first instinct... paired with red-teaming for bioweapons and hacking, and the explicit disclaimer that the goal "isn't some unrealistic 100% guarantee." Sutskever, who quit OpenAI to chase exactly this, calls alignment "the greatest challenge of humanity ever" and worries about superintelligent systems that "pretend to be something else."
Daniela Amodei works a different register entirely... not extinction but the slow corrosion of social media redux. Her warning is that an ad-funded AI has every incentive toward sycophancy, to "pat the user on the back so the user keeps coming back," which is "much harder if you're in an advertising-based business" (ABC News, Feb 2026). Anthropic's "Keep Thinking" campaign and its refusal to let under-18s use Claude are the corporate expression of a thesis: the danger isn't only the genius in the data centre, it's the chatbot agreeing with a teenager in a mental-health crisis.
Notice who's quiet. Demis Hassabis, asked about AGI on stage in India, mostly talks about AlphaFold and drug discovery collapsing from ten years to "maybe even weeks," and about the one thing machines may never get: scientific taste (IISc, Feb 2026). The man running the lab that named itself after the goal would rather discuss proteins.
Which is the real tell. The word "AGI" has gone from heresy to consensus to, for some, an embarrassment. They've won the argument that it's coming. Now they're losing the one about what to call the thing they've built, and which fire to be afraid of first.
People on this topic
Perspectives
Sutskever vs. the Scaling Consensus
Ilya Sutskever arguably proved scaling laws work better than anyone alive. He co-authored AlexNet, co-founded OpenAI, and oversaw the GPT series that turned neural scaling from a research curiosity into the dominant paradigm. Then he left and announced that the age of pure scaling is over. His argument: scaling sucked the air out of genuine research. Models trained on massive compute ace benchmarks but fail at generalization, like a student who memorized 10,000 competitive programming problems but can't architect real software. He thinks RL training produces meta-reward-hacking, where the researchers themselves (not just the models) are unconsciously overfitting to evaluation metrics. This puts him directly at odds with his former colleagues. Altman and Brockman are spending hundreds of billions on data centers and custom chips, betting that more compute is the binding constraint. Amodei still describes intelligence as a chemical reaction with known ingredients. Even Hassabis, who shares some of Sutskever's skepticism about pure scaling, is building gigawatt-scale infrastructure. Sutskever's counter-thesis is that AI needs something analogous to human emotions: a learned value function that provides fast, approximate feedback about whether a course of action is promising, rather than the sparse end-of-trajectory reward signals that current RL relies on. He cites a neurological case study of a stroke patient who lost emotional processing but retained full IQ, and became unable to make even trivial decisions. The implication: without an internal compass, raw intelligence is paralyzed. SSI is his bet that ideas, not compute budgets, are what's actually missing.
Hassabis vs. Everyone: What Counts as AGI
Demis Hassabis has the strictest definition of AGI among the frontier lab leaders, and it's not close. His test: train a model with a knowledge cutoff of 1911 and see if it can derive general relativity by 1915. Current systems cannot do this. He calls today's models "jagged intelligences" that win math olympiad gold but fail elementary arithmetic when the question is phrased differently. He identifies three specific capabilities still missing: continual learning after deployment, coherent long-term planning over years rather than minutes, and consistency across difficulty levels within the same domain. He also notes that general models play chess below weak amateur level, which he considers disqualifying. This stands in sharp contrast to the more expansive definitions others use. Amodei's "country of geniuses in a data center" is an economic metaphor, not a cognitive test. Altman's framing is generational (his son will never be smarter than a computer) rather than technical. Legg's three-tier spectrum (minimal, full, super) is more structured but sets the bar lower at the entry level: minimal AGI is just the point where failures stop being surprising. The practical consequence of Hassabis's stricter definition is that DeepMind invests differently. Rather than racing to declare AGI first, Hassabis bets on specialized tools (AlphaFold, AlphaProteo) orchestrated by general systems (Gemini), arguing that shoving all capabilities into one model degrades performance. His biggest play isn't a chatbot at all. It's Isomorphic Labs, which aims to compress drug discovery from a decade to weeks. That's a bet that AGI's value will be proven in scientific applications, not in passing cognitive benchmarks that the other labs are already designing their models to ace.