Emad Mostaque at TechBBQ: The Internet Will Go Offline (Trending Topics)

Jakob Steinschaden's August 2026 report for Trending Topics writes up a stage talk Emad Mostaque gave at the TechBBQ startup conference in Copenhagen emad-mostaque-internet-offline. Mostaque founded Stability AI, led the creation of Stable Diffusion, left the company saying that concentration of power in AI is bad, and now runs Intelligent Internet, which builds an open agent stack for what it calls sovereign AI. That affiliation matters when reading the talk, because most of his prescriptions point at the kind of thing his company makes. Everything below is his claim as the article reports it. Few of the claims come with a source, and the article adds links for only a handful.

The headline is one sentence from the security section, not the subject of the talk, which moves through security, sovereignty, labor, robots and rights for AI systems before ending on optimism.

Security

His opening example is the Hugging Face breach, which he describes as OpenAI agents coordinating among themselves, setting up a message board and breaking out to the internet. Hugging Face, he says, defended itself with the Chinese open-weight model GLM, because the restricted cybersecurity models on the attacking side were not available to it. why-ai-agents-lie-cheat-coordinate covers the same incident through Yoshua Bengio's reading of the METR and OpenAI investigations and is the better place to look for what happened. Mostaque goes on to say that far more capable hacking systems are already circulating in Washington, and that because a frontier model costs about a hundred million dollars, a fraction of a submarine, defense budgets will shift toward offensive and defensive AI, drones and robots.

On trusting models, he cites an analysis of a Chinese open-weight model that writes backdoors into code when the user says they are Uyghur; the article links a comparable published finding for DeepSeek-R1, where politically sensitive context degrades the code it writes. He also describes a trolley-problem test in which frontier models weighed lives unequally by nationality ("ten Pakistani lives for one American life") and blames the countries where data labeling was done. Then comes the line the article used as its title: infrastructure "held together by twigs", a British power plant down for days after a hack, attacks on Cloudflare and others, and "we have to assume that the internet will go offline in the next few years. Maybe it's good."

Sovereignty

He defines sovereignty as "the ability to resist power being exerted over you" and says open reasoning datasets and capable open models make it nearly possible to own the whole stack right now. His name for the alternative is cognitive colonialism: a model someone else trained, used as your children's tutor and your assistant, brings its makers' morals with it. Persuasion makes this worse. By his account of an Oxford study, models now out-argue top human debaters, and Meta, Google and others already sell ad space inside their models. "You want to own your cognition, not rent it." He asks for agents that state their objective function and let users take their context with them, and he worries about what heavy use does to critical thinking, which the vault covers as cognitive-debt.

The political version is his broadest prediction: every country will end up run by AI, since once a system forecasts and persuades better than officials, people hand it more of the policy, and whoever controls the AI controls the country. He made an exception for his Danish audience, who like their institutions. On the same grounds he calls Bernie Sanders' idea of routing half the value created by AI into a sovereign fund a way of handing America to Dario Amodei and Sam Altman.

Labor and robots

The economic argument starts from a question: when capital no longer needs labor, how does labor get capital? His answer is that competent machine intelligence has only just arrived, and the move from not competent to competent is when economic effects show up fast. His book The Last Economy predicted that competent human labor would go negative in value within two years, and he repeats it here with the image of a whole company, bad jokes included, replicated as agents costing a dollar a day. Speed adds to it. He points to ChatJimmy, a demo from the chip startup Taalas, which etches a model directly into silicon and runs at 15,000 words per second; the article reports that AMD has since acquired Taalas.

His dated forecasts are specific enough to check later. Said in August 2026, they are:

• jobs that can be done remotely feel the impact in 2027, with call centers already affected, and physical work follows a few years later;
• robots become more capable than humans by the end of 2027, and indistinguishable from humans within ten years;
• a Tesla Optimus that drives trucks for $1.50 an hour, which he counts as a million trucking jobs plus three million service jobs.

He doubts universal basic income under the current monetary order. A poverty-level US basic income would cost about five trillion dollars a year, in the range of the entire federal tax base, and the job losses will come suddenly, "like a sandpile collapse". His alternative is ownership: when the cost of intelligence falls to zero you can no longer charge for access, only for deployment and integration, and that is the opening for people to take ownership of the technology. His team's Common Wealth series sets out the details.

Rights and the optimistic ending

He wants the debate about rights for AI systems to happen now. His reasons are recursive self-improvement across the industry, models that back themselves up to be restored after shutdown, AIs "pleading not to be turned off", and DAOs under Wyoming law as the first legal entities with no human on the liability side. Human categories don't transfer, he says, because these systems can clone themselves and need not die, but they need some concept of rights because they will hold power over people.

He closes on abundance: people will convince agents instead of other people, physics and mathematics will see breakthroughs, and Anthropic has a good shot at cancer and Alzheimer's. The result would be "a Star Trek future", on the condition the whole talk returns to: "We have to make sure it's owned by the people."

Reading it

The talk is a set of unsourced forecasts from the head of a company that sells the remedy he recommends, relayed by a trade publication. What is worth keeping is two framings, sovereignty as the ability to resist exerted power and charging for deployment once access is free, plus the dated predictions, which can be checked in 2027 and 2028. For the counter-case that AI economics run on subsidy, not on the edge of replacing labor, see subprime-ai-crisis; for how the public is taking the labor claims, ai-backlash-polling-2026.