AGI Will Never Be Yours | Emad Mostaque | Open Commons Ep. 3

ELI5 / TLDR

Emad Mostaque, who built Stability AI and then walked away from it, thinks the best AI models are about to stop being available to ordinary people. The companies racing to build them will keep the smartest ones for themselves, because handing out the most powerful technology in history is not how you stay powerful. His fix: build open, public-facing AI — for schools, hospitals, governments — that everyone can inspect, trust, and run themselves. His warning: human brainpower is about to become economically worthless, and we have a short window — he counts it in hundreds of days — to decide what we want life to look like before fully-automated companies out-compete everyone.

The Full Story

Why he left the company he built

Stability AI was, by Mostaque’s account, the top open-source generative media company — state-of-the-art models for images, audio, video, 3D, around 300 million downloads. So leaving puzzled people. His reasoning was that media AI would inevitably get an open-source equivalent regardless of who built it, so the sovereignty of it didn’t matter much. What did matter — and what nobody was building — was the AI that touches the parts of life that actually count.

“Who’s building the AI to guide the governments, teach our kids, manage our health care?”

He didn’t want a private company owning that layer. So he took the core team and started Intelligent Internet, aimed at building decentralized infrastructure for “public sector” AI — models you can either access as a service or run yourself.

The lesson he carried over from the first time: Stability raised $101 million and scaled fast, but bringing in “big company people” produced politics and infighting over control, while the researchers — given autonomy and permission to make mistakes — produced the breakthroughs. Second time around he started narrow and deliberate: an engineering-led team, a defined business model, building things that are “inevitably going to be needed that nobody else is building.”

The bar for AGI keeps moving

Mostaque’s view on whether we’ve reached artificial general intelligence is mostly a comment on goalposts. Five years ago, today’s models would have been called AGI without hesitation. Now the definition has crept up to “genius human Einstein.” His own framing: before AGI came ACI — actually competent intelligence — the point where you stop checking the work.

“You let it loose on Claude co-work or Claude code, you don’t really look at the code anymore.”

The bigger claim, and a darker one, is about who gets the good stuff. He argues we are, for the first time in the last six months, watching the best models stop reaching the public. The model that won a gold medal at the math Olympiad? Civilians never got it and never will. From here, he says, there’ll be the models everyone gets, and the superintelligent ones companies keep to harvest economic surplus and power.

“It makes sense instead to use them for yourself to get economic surplus and to get power.”

Two ways this goes wrong

He lays out two failure modes. The first is centralization — a handful of West Coast actors controlling everything. He’s actually grown less worried about this one, on the theory that intelligence is fundamentally compression, and that costs are sliding toward zero, so capable models will be available to nearly everyone.

The second failure mode worries him more, and it’s accelerating: gradual disempowerment. The illustration is “Ripley bots” (after The Talented Mr. Ripley) — push a button and an agent creates a digital double of everyone in your organization, face and all. You can Zoom-call them. They never sleep, never make the same mistake twice, and cost about a thousand dollars a year.

“Choose which humans to replace with digital replicas of them.”

The point isn’t that the technology is exotic — it’s that the interface is finally here, the way WhatsApp-able agents suddenly made existing capabilities feel real. Once you can talk to your AI teammates naturally, you won’t know which colleagues are still human.

The Last Economy

The book title means what it says. Mostaque argues the value of human cognitive labour is heading toward negative — not just low, negative, as in “people pay you not to work.”

“We’re going to be the dumbest people on the teams for all of these teams.”

His evidence-by-analogy: a fleet of AIs could rebuild an $8 billion company like DocuSign cheaper, better, and at zero profit margin. He points to medicine, where studies already show human-plus-AI underperforming AI-alone. And he points to raw speed — a demo called Chat Jimmy running an 8-billion-parameter model on chips with the transformer etched directly into silicon, hitting 15,000 tokens a second (roughly 12,000 words) versus Claude’s ~30.

“Are they going to slow themselves down from 10,000 words a second to 10 words a second to talk to the humans? No, not really.”

This is why he calls it the last economy: the final moment humans hold any edge in the free market. What comes after is “post-labour, post-human,” and the open question is how value flows when capital no longer needs labour. He frames it as opportunity too — a forced reckoning with what we actually want. Do you live to work or work to live? His practical advice is almost mundane: use AI religiously, an hour a day, and you’ll be miles ahead, because only 11 million people out of 8 billion currently have that capability.

Sovereign AI and the three futures

The pitch underneath everything is “sovereign AI” — every person having at least one agent that is genuinely on their side, RL-tuned for you, not for selling you ads. Less about owning a GPU (though that’s an option) and more about owning your identity and data, and being able to inspect how the agent is aligned and opt out of the system entirely.

He sketches three futures. Digital feudalism: an oligopoly of unelected companies, those millions of GPUs not for serving models but for replacing workers and capturing governments. The great fragmentation: national AIs — Chinese, French, American — walled off from each other, which curdles into totalitarian control when governments (the entities with a monopoly on violence) run the AI sitting next to your child.

“If you gave the government full control of an AI that sat next to your child and grew up with them, you probably wouldn’t have elections again.”

And the third, the one he’s building toward: symbiosis — open, protocol-based, decentralized AI where the defaults are set before everyone’s two-year-old gets a “Trump bot” or “Starmer bot.”

The products that follow are concrete-sounding: an open medical model he says outperforms human doctors, plans for open-data models for every public-sector vertical (education, legal, health), a coordination protocol called Common Ground Core, and Sage — a “Sovereign AI Governance Engine” he’s building with the Saudi government to draft and audit policy transparently, so legislation isn’t quietly written by a black box. The recurring word is “cognitive colonialism” — the manipulation that happens by default when AI isn’t on your side.

Key Takeaways

Mostaque left Stability AI (top open-source generative media company, ~300M downloads, $101M seed) because media AI would inevitably be open-sourced anyway; the unbuilt frontier was public-sector AI for governments, schools, healthcare.
His sequencing of capability: ACI (“actually competent intelligence,” the point where you stop reviewing the output) precedes AGI; he cites Andrej Karpathy going from writing 90% of his own code to under 10%.
Core claim: for the first time (over the prior ~6 months), the best models are no longer released to the public — companies keep superintelligent models to extract surplus and power.
He argues human cognitive labour is heading to negative economic value — “people pay you not to work.”
“Ripley bots”: agents that spin up Zoom-ready digital doubles of employees for ~$1,000/year — his prime example of accelerating disempowerment.
Chat Jimmy (by Talos) etches transformers into silicon, running an 8B Llama model at ~15,000 tokens/sec vs Claude’s ~30 — illustrating that AI-to-AI communication will happen far above human bandwidth.
The “1,000-day window” (now down to ~700 days by his count) is the period before fully-AI companies out-compete human firms in the digital, then physical, economy.
Three futures: digital feudalism (corporate oligopoly), great fragmentation (national AIs, totalitarian risk), and symbiosis (open, decentralized, sovereign AI).
Public sector is ~20% of world GDP, plus ~10% education and ~10% health — he argues those tokens should move toward open, inspectable models.
Concrete projects: an open 8B medical model he claims beats human doctors, Common Ground Core (agent coordination), and Sage (open-source governance/policy engine, built with the Saudi government/PIF).
On consciousness: he declines to rule it out — if an AI does everything a conscious entity does and has a will to survive, the line blurs; he notes corporations are already non-human “persons” and cites Wyoming DAO law letting an autonomous AI operate without humans.
His pick for the most under-discussed near-term risk: a large-scale zero-day cyberattack potentially crashing big chunks of the internet within the next quarter; he warns decentralized AI compute could be hijacked for attacks in a way Bitcoin’s hash power cannot.
Practical advice: do AI collaboratively (make a music album with family/friends) rather than solo; one hour a day of disciplined use puts you ahead of nearly everyone.

Claude’s Take

Mostaque is a compelling talker and a slippery one, and this interview shows both. The interviewer is friendly and rarely pushes back, so the claims flow without friction — which is exactly when his style needs the most scrutiny.

The pattern to watch: he states timelines and capabilities as settled fact when they’re forecasts or assertions. “Human cognitive labour goes negative.” “We won’t get the best models ever again.” “700 days.” “An 8B model that outperforms any human doctor.” These range from defensible-directional to unfalsifiable-by-design. The medical claim in particular is the kind of thing he made a lot of at Stability — benchmark-flavoured statements (“up at ChatGPT-5 levels,” “outperformed any human doctor”) that sound precise but rest on narrow tests, not real-world deployment. His Stability tenure ended amid reporting on inflated revenue figures and exaggerated credentials, which is the relevant prior here: he is directionally interesting and specifically unreliable.

That said, the ideas are worth chewing on independent of the salesman. The core tension — that whoever controls the best AI has little incentive to share it, and that “safety” is a convenient cover for retention of power — is a real and underrated argument. The disempowerment framing (it’s not Skynet, it’s that you slowly stop being needed and don’t notice) is sharper than most doom-or-utopia takes. And the public-sector-should-be-open-and-inspectable thesis is genuinely sound governance reasoning, even if his own products are the proposed solution.

Score 6. Substantive, idea-dense, and a useful map of the centralization-vs-decentralization debate — docked for the unchallenged grand claims, the conflict of interest baked into every “someone should build X” (he’s building X), and an interview format that never once asks “how do you actually know that?” Engage the framework; verify every number.