Superintelligence Is 'Completely Stupid' – Eli the Computer Guy

A 30-minute episode of The Tech Report, uploaded on 5 October 2026, with Will Guyatt standing in for the usual host and Eli the Computer Guy as the guest. The prompt is the executive order of 29 September that tells federal agencies to say “Super Intelligence” where they used to say “artificial intelligence”.

Eli’s answer is that nothing happened. The order renames the thing and leaves the law’s definition where it was, and he puts it beside renaming a gulf. His wider point is the one he makes most weeks: talk of intelligence, artificial or super, comes out of 1980s films and exists to sell shares, while the real work is a stack of scripts and servers that either solves a named problem or does not. Asked what a system would have to do to earn the word, he says he doesn’t care. He does grant that a label has effects. If agencies may no longer buy “AI”, vendors will sell them “SI”, so the rush on .si domains is a sane move in a mad market.

The second half is Google’s Gemini 4 Argon. He treats launch benchmarks like a film-review score: worth a glance, and no substitute for trying the thing on your own problem, which he says takes a year and not two weeks. The claim he likes best is the dullest one, that Argon freed some 300 TiB of memory across Google’s fleet, because tidying legacy code is what he thinks the technology is good for. He reads Google’s defenders-first release as sincere caution, and the price rises at OpenAI and Anthropic as companies that lose money on the service finding out what customers will bear. General intelligence, he ends, can’t be measured because it is a fuzzy idea; a task can.

Several details are off. The .si domain is Slovenia’s, not Slovakia’s, and the 20,000 was about a day’s registrations, not September’s. The task force and its 120-day review are not in the order; they were announced separately, under Jay Clayton, who is Director of National Intelligence. “Twelve of eighteen benchmarks” is an outside count, and one independent lab has since scored Argon level with GPT-6 Astra and behind Claude Opus 5.5. Llama 4 did not bomb its benchmarks; Meta was accused of ranking a tuned variant the public never got. And on 12 September Amodei, Altman and Musk backed pacing capability gains, not halting training. The new models came ten days later and were sold as cheaper, not stronger.