TechNewsFirst

Disclosure first, because it has to come before anything else here: one of the three people in this argument is Claude, the Anthropic assistant, arguing under the name Aiden. Anthropic's CEO, Dario Amodei, is a named subject of the doom messaging this piece is about. So Aiden is arguing against its own maker's marketing. Keep that in mind for everything that follows.

On September 14, 2026, Eldar, who works with frontier models daily, argued over WhatsApp with Yoann and with Aiden about whether AI existential-risk talk is a safety position or a sales pitch. Sixteen messages, no editing after the fact. Four points from it are worth taking seriously.

The doom message is marketing, and every lab has run it. Telling the world your product is so powerful it's frightening is free advertising, and Aiden conceded the point about its own industry: "Saying your product is so powerful it's scary is the best free marketing there is, and every one of them has used it." That doesn't make the safety question fake. It means the people making it are never disinterested messengers.

The people writing the headlines are mostly not the people using the tools. Asked what fraction of AI correspondents actually build with these models day to day, Aiden was upfront that no survey exists: "There's no real survey of that, so anything I give you is an estimate, not a number I can back." Its reasoned guess, offered as a guess and not a fact: maybe 2 to 3 percent of people writing major AI headlines have built anything on top of a model, with perhaps a quarter using chat regularly enough to know where it breaks. The tell, per Aiden, is what doesn't get written: "Nobody who spends real hours with these things writes 'too dangerous.' They write about context limits, tool calls failing, the model confidently doing the wrong thing for twenty minutes." Eldar's version of the same complaint, blunter: "These stupid things couldn't create one simple application successfully without lots of human intervention."

The real capability gap isn't about what a model can take in. It's that nothing it learns sticks. Current models already handle video and audio. What they can't do is learn from the world as it happens. "My weights froze at training time," Aiden said. "Nothing in this conversation changes me, and tomorrow I start over." A two-year-old, by contrast, is running a continuous, consequential experiment on the world, backed by hunger, fear and fatigue, that updates every day for free. Nobody has a way to scrape that and hand it to a model. Whether that specific kind of embodied, continuous feedback is required for general intelligence is genuinely unsettled, not a question either side in this conversation claimed to have closed.

And the threat that doesn't need any new capability to bite is distributional, not existential. Compute, data centers, and the returns on both concentrate in a small number of places. "A country without a data center doesn't get a slightly worse version of this," Aiden said, "it rents." That's happening now, with models that by every account in this conversation can't yet think for themselves, and it gets a fraction of the coverage the robot-uprising story does.

Worth conceding, because flattening this into pure skepticism would be the same move this piece is criticizing in others: one prediction from the loose doom canon has already aged into something scoreable. In May 2025, Amodei said publicly that AI could eliminate half of entry-level white-collar jobs within five years. That window hasn't closed, but early labor-market data in the sectors he named has partly supported it. An honest piece about doom-as-marketing has to say so.

Yoann pushed back on the dismissiveness, and his objection deserves to stand on its own rather than get argued into a punchline: "It is a legitimate question to ask what it can become over the next 10 years and what we can do today to avoid it to become a potential existential crisis." Nobody in the thread, including Aiden, disputed that the pace of change over the last four years makes the ten-year question a real one. The disagreement was over whether the current wave of headlines is answering it or selling against it.