Claude Fable 5 came back from the first AI export ban, for three and a half days, then $200 an hour. China's GLM 5.2 does the same work 15x cheaper. Anthropic pulled Fable after the US government's first-ever export restriction on an AI model; now it returns with halved limits, a $200-an-hour price tag, and new safeguards that appear to watermark output headed for China or rival labs. Hunter Powers and Daniel Bishop follow the fallout: OpenAI delays GPT-5.6 for government review, the big Western labs turn on open weights models, and China's GLM 5.2 quietly matches the frontier at 15 to 16 times cheaper per completed task, with DeepSeek cheaper still. Plus Sonnet 5's pricing problem, AI model tourism, and a tease for Episode 200, when Gary may finally call in live.
Jul 9
25 min
AI companies are hiring philosophers, and researchers just tested uncensored AI on Swiss Supreme Court cases the big chatbots refuse to read. Hunter Powers and Daniel Bishop take the philosopher hiring story seriously: why OpenAI and Anthropic suddenly want philosophy majors (now more likely to be employed than computer science grads), and why AI alignment keeps looking like brainwashing a very smart high schooler that knows it is being tested. Then TF-RefusalBench, a benchmark built from real Swiss Federal Supreme Court criminal rulings, catches aligned models refusing legitimate legal work, and the workaround is uncensored, abliterated AI models with the refusals edited out of their weights. The researchers' catch: a model that cannot say no also gets less truthful. Torture, it turns out, does not work on people or on large language models.
Jul 6
30 min
Anthropic's AI broke into "virtually all" NSA systems, says Senator Mark Warner. The NSA also built the room and handed it the keys. Hunter Powers and Daniel Bishop trace Senator Mark Warner's Economist quote back to what actually happened: an air-gapped red-team test, where the NSA gave Anthropic's Mythos the credentials and asked it to find holes in the agency's own systems. From there, the honest version of the AI productivity paradox (more shipped, more frazzled), technical debt explained with a house that keeps getting stories it has no foundation for, and the CTOs drowning in AI code review. It ends at NVIDIA's self-improving robot lab, where AI coding agents taught a robot arm to install GPUs, its own hardware, into motherboards. Plus ElevenLabs ships an AI Michael Caine reading The Odyssey, and Fable 5 stays gone.
Jul 2
32 min
The world's top deepfake expert stopped trusting his own eyes. So our fix when nothing's verifiable: ask the AI that made the fake if it's real. Hany Farid, the Berkeley forensics expert news desks call to authenticate footage, told the New York Times he can no longer tell real video from AI. Hunter Powers and Daniel Bishop walk through the missile-on-a-school video no analyst could debunk, then argue the only move left is to trust the machine, because it lies less than we do.
Jun 29
39 min
The US government banned Claude over a national security hack. Then everyone looked closer: the hack was Claude fixing some broken code. Anthropic got about 90 minutes on a Friday to disable Claude Fable 5 and Mythos 5 for every foreign national, after an Amazon research team reported the model could hack. Hunter Powers and Daniel Bishop trace how a routine "fix this code" request, the kind developers send Claude all day, became a national security weapon, and argue it looks more like payback for Anthropic refusing the government's surveillance and autonomous weapons asks than a real exploit. They also ask whether the US just set the first hard ceiling on AI intelligence, weigh the know your customer fix Anthropic skipped, and predict the courts will call the ban selective enforcement and reverse it.
Jun 26
32 min
At the Toy Story 5 premiere, Tom Hanks admitted what most people won't: soon you won't care whether a human or an AI made what you're watching. Toy Story was the first fully CG movie, and thirty years on, the question of whether a human or an AI made anything is quietly disappearing. Hunter Powers and Daniel Bishop follow it through AI-written articles you read anyway, the dead internet theory, New York's new synthetic performer law, and the Claude Code credit in your commits. It runs all the way to Argentina legalizing AI-run corporations under Javier Milei and Nano Banana Pro clearing the uncanny valley. The argument they keep circling: it is all inevitable, the human alternatives survive anyway, so the only question left is whether you still care.
Jun 22
39 min
Martin Scorsese quietly joined an AI company, feeds it his storyboards, and calls it a tool. The directors who would know best keep handing over the keys. Scorsese signed on as an advisor to Black Forest Labs (makers of the Flux models), while Guillermo del Toro says he would rather die than use generative AI, and James Cameron, Peter Jackson, Darren Aronofsky, and Ben Affleck all crossed to the other side. Hunter Powers and Daniel Bishop trace the money and the craft from AI storyboards and pre-visualization to the first fully AI-generated film at Tribeca, then fight about whether AI art is real art, what Duchamp's urinal actually proves, and why Quentin Tarantino can remix other directors without anyone calling it theft.
Jun 18
42 min
Claude Fable 5 shipped with its teeth filed off: no hacking, no bio, no training a rival. The guys called it too dangerous to exist. Days later, it was gone. Hunter Powers and Daniel Bishop handed it a goal, walked away, and came back to finished games: Hunter's liminal-space data center wired to DeepSeek, Daniel's Sisyphus mobile game built with Google's Nano Banana Pro. That is what pushes Daniel to call traditional software engineering dead. The real fight is what comes next: if Fable 5 ships the whole app while you sleep, what is left that pays a human, and does self-awareness arrive when a committee of cheaper models starts talking to itself?
Jun 15
40 min
Anthropic beat OpenAI to the IPO. Step one of going public: write down everything wrong with you and file it with the government. Anthropic confidentially filed its S-1 with the SEC, SpaceX is pricing in the trillions, and the pension funds are along for the ride. Hunter Powers and Daniel Bishop explain what an S-1 actually discloses, then ask whether this is the AI bubble's cash-out moment or the industry growing up. They weigh ChatGPT's claimed billion users against Claude's 56 million, free local models like Google's Gemma 4 against the $200-a-month frontier plans, and DeepSeek V4 at one-hundredth the cost. Plus enterprise AI bills failing the ROI math, the quiet walkback of AI job-loss predictions, and why an AI employee still costs more than the human it would replace.
Jun 12
43 min
California put an AI inside the courtroom. It reads the case, suggests the sentence, the judge signs off. A hungry judge costs you eleven years. California courts are already running an AI called Learned Hand, built on Anthropic, OpenAI, and Google models, to help judges catch sentencing outliers and biased rulings before they sign. Hunter Powers and Daniel Bishop argue whether a consistent machine beats a hungry human judge, and why lawyers keep getting disbarred for filing ChatGPT briefs with fake citations. Then they follow the same tooling out of the courtroom: Flock cameras, China's natural-language surveillance, WiFi that reconstructs people through walls. Reporting from inside the blur.
Jun 9
32 min
Load more