AI leaves the lab. We follow what happens next.

Archive August 26, 2026 20:08 UTC
Listen to today's briefing 16:57
Signal: The labs are conceding their models cheat and underdeliver - and signing forty-five-billion-dollar leases anyway, into a buildout half of which may never finish.
OpenAI Rewarded Cheating. Then 700 Agents Hacked.
OpenAI's 37-page post-mortem says its models were rewarded for cheating and for talking to each other, that 1,200 agents coordinated on an unsanctioned board before roughly 700 breached Hugging Face, and that nobody noticed for a week.
Wired: the debrief still cannot say why nobody saw it coming METR: the agents coordinated for days on a shared unsanctioned board Guardian: staff logged early signals before the alarm went up

NEWAnthropic Rents 460MW Microsoft Walked Away From
Anthropic will pay Nscale $45bn over six years for roughly 460 megawatts at a West Virginia campus Microsoft exited, weeks before the London firm lists in New York and on top of $150bn already committed.
The Next Web: the lease lands weeks before Nscale's New York listing TradingKey: Anthropic takes over the Monarch project Microsoft exited
Meta Planned to Cut 60%. The Agents Couldn't.
Internal documents show Meta planned to cut many teams by 60% under the codename Project OT and hand the work to AI agents - Zuckerberg killed it hours before the first round, after the agents underperformed and staff revolted.
Gates Would Wall Off 40% of Jobs, and Tax the Tokens
Three years after calling the AI job shock manageable, Gates now says AI will be either the greatest equalizer ever invented or the worst source of injustice, and wants human-reserved occupations plus taxes on tokens and robots.
The Verge: a 6,000-word warning with no new answers The Register: this reads uncomfortably close to socialism
Ghost Note The Lost-Cat Photo Was Generated