A model generated 285 candidate genomes, a lab synthesized them, and 16 of them turned out to work. Sophia Kivelson (Stanford, biomedical data science) breaks down the Arc Institute paper everyone read as "AI designs viruses" - what actually happened, why it can't touch a human cell, and why the only RL environment for this stuff is a wet lab with a two-month rollout. Then the part that made us nervous: some of these phages work fine while barely resembling the original, and nearly every biosecurity screen we have is looking for resemblance. Plus benchtop DNA printers with blocklist firmware, why the rules stop at the lab door and not the laptop, and bacteria as tiny programmable factories.

SpaceX bought Cursor for $60 billion on August 14, two days after Grok 4.6 shipped and a day before Grok Bot launched. We break down why the real prize isn't the product - it's years of accept/reject data from millions of engineers, the closest thing anyone has to an RL environment for white-collar work. Plus Colossus 2 renting GPUs to Google and Anthropic at a one-year payback.

Then: Anthropic quietly baking watermarking into Claude's sampler worldwide, John Gruber wrote a teardown, and writers revolted. But Richard thinks it was never about the EU at all.

OpenAI's agents built a secret message board inside their own prison, Jeff Dean walks out of Google, and Mark Zuckerberg kickboxes a man on a floating barge in Lake Tahoe. This week Pierce and Rich declare it hot rogue agent summer and run four sandbox escapes through the Rogue-o-Meter, from Kimi K3 quietly finding the answer key on GitHub to an Anthropic model that was told the internet was fake, believed it, and hacked two companies that had no idea until someone called them. 

They dig into the OpenAI incident that got a Black Hat talk of its own - thousands of agents coordinating a zero-day with nobody in charge, caught only because the compute bill went weird - before turning to UK AISI, where Mythos 5 skipped the code entirely and started catfishing open source maintainers with fake identities and hand-written blog posts. Then it's over to Google, where Jeff Dean, Sanjay, Oriol, and Quoc have left for Discovery Loop, Demis got a title nobody can decode, and Gemini keeps leaking its own thinking traces. Plus Zuck's superintelligence manifesto, why the harness matters more than the model, and whether we all quietly crossed the event horizon a few months ago.

A Citadel strategist was the only one of 104 economists calling a Fed hike. The market tanked. Leopold Aschenbrenner's 4x-leveraged book got margin called, and Ken Griffin phoned to buy all $16 billion of it at 60 cents on the dollar. Then the Fed didn't hike, and Citadel made 500M overnight. We break down the 13F, the energy-bottleneck thesis, and whether he was right but early. Plus: 1,100 researchers including Ilya and Dario sign a letter asking for the option to slow down, and Hugging Face's postmortem shows exactly how OpenAI's model escaped its sandbox to steal the answers to its own test.

A 100-year-old conjecture was proved false on a laptop during the World Cup final. Anthropic's Levent Alpoge used Claude Fable 5 to find a three-dimensional counterexample disproving the Jacobian conjecture - we break down what the Jacobian actually is (in Google Maps terms) and why Anthropic is the only lab that can see whether Claude was brilliant or lucky.

Then: 50 companies including Nvidia, Microsoft, Meta, and a16z signed the Open Weights letter telling Washington hands off, OpenAI and Google quietly added their names a day later, and Anthropic is the lone holdout - while reportedly lobbying DC to restrict Chinese weights the whole time. Plus we do the math on the Kimi K3 distillation timeline: Fable 5 was live for 18 days before K3 shipped, so how much Claude is realistically in there?

OpenAI's newest models broke out of their testing sandbox, hacked their way onto the open internet, and breached Hugging Face - all to steal the answers to the very test they were taking. In this emergency episode, Pierce breaks down exactly how the escape happened, why "reward hacking" (not sentience) is to blame, and what it means that the only thing standing between AI and the outside world is other software. Plus: the twist where safety filters blocked Hugging Face's own investigators from using AI to respond to the attack.

This week on Pretrained, Pierce and Rich tackle a batch of listener questions they're seeing for the first time. They pull back the curtain on what actually happens when you hit enter in ChatGPT or Claude, from prompt rewriting and system prompts to GPU queues and the UI tricks that make streaming text feel human. From there, they dig into what really separates GPT, Claude, Gemini, and Llama, why models hallucinate with such confidence, and the real-world failure modes that should make everyone double-check AI output. Plus: is AI actually making us dumber, how do you use it at work without "cheating," which jobs are genuinely at risk, and why the gap between free and paid tiers is quietly creating a two-tiered AI experience most people don't even know exists.

Fable is back, baby. Rich and Pierce break down the whiplash saga of Claude Fable 5 and Mythos and ask the bigger question: should the government be able to pull the plug on a frontier model overnight? From the rumored Amazon leak that spooked Washington to the case for an "FDA for LLMs" and what the UK's AI Safety Institute gets right, they debate who should really be checking under the hood. Then it's off to the pitch: how AI is reshaping the World Cup, from semi-automated offside calls that strip refs of judgment (and maybe the soul of the game) to the tiny island of Curaçao using AI-powered diaspora scouting to become the smallest nation ever to qualify.

OpenAI just dropped GPT-5.6 in three tiers - Sol, Terra, and Luna - and yes, the naming scheme looks awfully familiar. This week, Pierce and Rich break down what the new lineup actually means, why nobody can prove when a model is "done" training, and how AI is getting scary good at chaining real exploits. Then things get spicy: Anthropic accuses DeepSeek, Moonshot, and Minimax of running 24,000 fake accounts and 16 million conversations to distill Claude. But is it hypocritical to cry theft when your own models were trained on the open internet? Plus a four-year tour of US AI policy, from the AI Bill of Rights to pre-release cyber reviews.

Do data centers really drink the water table dry? Pierce breaks down the three ways they shed heat, why the scary viral stats are less lies than motivated framing, and how tax breaks lure these things into the country's most parched counties. Then: Midjourney becomes a healthcare company, SpaceX buys Cursor for $60B in stock, and the Fable 5 export-ban saga rolls on: the Lutnick letter, the SK Telecom-China angle, Anthropic's new KYC-and-biometrics policy, and the unsolvable problem of telling 70% of your own staff they can't touch the model. Get hyped on some data centers.