The Daily Prompt

OpenAI just shipped GPT-5.6, and one variant called Sol is doing something no model has ever done before on a benchmark so hard it makes PhD candidates cry. Meanwhile, newspapers want the company sanctioned, its CEO of AGI deployment just stepped down, and SK Hynix raised more money in one day than most countries see in a year. Let's get into it. 1. GPT-5.6 Sol Wins First ARC-AGI-3 Game (TLDR AI)...

OpenAI just announced not one, not two, but three major products in what appears to be a coordinated blitz: an entirely new voice model that talks and listens at the same time, an agent system that'll actually do your work for you, and a shiny new GPT-5.6 to power it all. Meanwhile, the competition's heating up and the courtroom drama's getting messy. Let's get into it. 1. OpenAI Launches Agent Th...

OpenAI spent two weeks keeping its newest models under wraps at the government's request, then released them anyway while the White House insists it never gave permission. Microsoft, meanwhile, has started quietly replacing OpenAI's tech with its own in Excel and Outlook. And Google just dropped an open model family that thinks before it speaks. It's Wednesday, and the AI industry is having a norm...

The US government's cybersecurity team just got access to an AI that's actually good at finding code vulnerabilities, after being left out in the cold for months while everyone else got to play. Also today, Anthropic found a hidden thinking room inside Claude's brain, CLI coding agents have completely taken over the terminal, and Elon's AI company got a new name that's exactly what you'd expect. L...

Anthropic spent months talking up privacy and responsible AI. Turns out they also spent months secretly tagging Chinese users with hidden code to sniff out distillation attacks. A researcher just blew the whistle, and Anthropic's response was basically 'yeah, we've been meaning to turn that off.' Today we've got surveillance hypocrisy, a nineteen billion dollar data center deal, Microsoft laying o...

OpenAI just dropped three new models and the U.S. government is the bouncer deciding who gets in. Microsoft built its first reasoning model without anyone's help. And Mark Zuckerberg admitted to his own staff that Meta's big AI agent bet isn't exactly paying off yet. It's a weird week when the biggest AI stories are about who can't use the tech and who's finally breaking up with their suppliers. L...

OpenAI just floated the wildest trial balloon yet in the ongoing saga of how to pay for all this artificial intelligence: give the government actual equity. As in, a five percent stake. Bernie Sanders has his own version involving a fifty percent tax, because of course he does. We've also got Anthropic's flagship model back online after a three-week government timeout, Microsoft launching what is...

Three weeks ago, Anthropic's flagship AI went dark. The government hit Claude Fable 5 with export controls after a jailbreak got ugly, and every user lost access overnight. As of today, it's back. But that's not even the strangest part of Anthropic's June. We've got the comeback story, a brand new Claude for drug discovery, Meta turning into a cloud provider, and someone just figured out how to ga...

Anthropic just launched a proper science workbench, South Korea's dropping nearly a trillion dollars on chips and robots, and OpenAI is apparently willing to delay its IPO until someone agrees it's worth a trillion bucks. Also today, Meta's been testing competitors' chatbots by pretending to be suicidal teenagers, which is exactly as weird as it sounds. Plus brain-typing, inflation warnings, and a...

OpenAI just dropped three new GPT models, and the market responded by panicking about their IPO timeline. Meanwhile, Micron is suddenly worth more than Meta, Google's telling customers to cool it on the AI tokens, and the Trump administration has decided some companies can have their powerful models back. Let's unpack it. 1. OpenAI Releases Three-Model GPT-5.6 Family (TLDR AI) 2. OpenAI IPO Pushed...

The White House just did something unprecedented: it told OpenAI that its newest model can't go public until the government approves each customer one by one. We're talking customer-by-customer vetting for GPT-5.6, all because of concerns about cybersecurity capabilities so advanced they're getting compared to restricted models. Also today: OpenAI's first custom chip just shaved nine months off a...

OpenAI just announced its first custom AI chip, built with Broadcom in nine months flat, and it's already running workloads. That's the fastest anyone's ever gone from whiteboard to silicon at this scale. Meanwhile, Anthropic is accusing Alibaba of running thousands of fake accounts to rip off Claude, and Google can't seem to stop hemorrhaging talent to its rivals. We've got hardware ambitions, co...

Google just lost two of its biggest AI names in one weekend—including a literal Nobel Prize winner—and the market took two hundred and fifty billion dollars of revenge. Meanwhile, OpenAI's new security tools have already fixed seventy thousand vulnerabilities, and Anthropic wants Claude to become your next coworker. That's not a metaphor. Let's get into it. 1. OpenAI Ships Security Tools at Scale...

OpenAI just launched a suite of security tools it says can protect every organization on the planet. Bold claim. Meanwhile, the first AI export ban just hit Anthropic, a Nobel laureate jumped ship to join them anyway, and new research confirms AI is officially better at changing your mind than actual humans. We've got chip billions, coding drama, and one very ambitious patching project. Let's go....

It's been a rough couple of weeks if your name is Anthropic. First they shipped a model that throttled AI researchers, then the U.S. government shut down their best models entirely, and now they're promising everything will be fine again real soon. Meanwhile, Midjourney decided image generation is boring and they'd rather build full-body scanners instead. Also, OpenAI poached the guy who literally...

One of the eight people who literally invented the technology behind every major AI model just jumped ship from Google to OpenAI. That's the kind of hire that makes headlines, breaks hearts in Mountain View, and probably cost someone their quarterly bonus. Also today: the U.S. government just gave AI data centers a fast pass to the power grid, which is great news unless you were hoping electricity...

Anthropic spent months asking Washington to regulate frontier AI. Friday night, the Trump administration delivered. A surprise export control directive shut foreign access to Claude overnight, days after CEO Dario Amodei published an essay about responsible testing. Careful what you wish for. Also today: the hyperscalers are running out of cash, SpaceX buys Cursor for sixty billion, and Android be...

The Commerce Department banned everyone from using two Anthropic models on Friday. The official reason? Still a mystery. The actual reason, according to sources? Personality differences with the Trump administration. You really can't make this stuff up. Today we've got government overreach rattling the AI industry, OpenAI losing twenty billion dollars a year, and SpaceX dropping sixty billion on a...

The U.S. government just did something it's never done before: issued an export control order so strict that it killed two AI models worldwide. Anthropic had no choice but to pull the plug on Fable and Mythos for every user on Earth because there's no citizenship checkbox at the API level. We've got that story, a court ruling on AI liability, and SpaceX's wild first day as a public company. 1. Fed...

SpaceX just pulled off the largest IPO in human history, Elon Musk is on track to become a trillionaire, and there's a Mars colony bonus involved. Meanwhile, OpenAI and Anthropic are locked in a price war that has companies sweating through their AI budgets. Let's get into it. 1. SpaceX Prices Largest IPO in History (TechCrunch AI) 2. OpenAI Weighs Price Cuts Amid Competition (The Rundown AI) 3. B...

Google just dropped a new way to generate text that's four times faster than the standard approach, and it does it by basically cheating at how language models work. Also today, SpaceX wants to put your AI in space, Anthropic's CEO is asking to be regulated, and OpenAI might finally go public. 1. Google's DiffusionGemma generates text four times faster (TLDR AI) 2. Anthropic CEO calls for stronger...

Anthropic just told us that over eighty percent of the code landing in their codebase is now written by Claude, not humans. Eight times as much code shipping per engineer per day. If that doesn't make you wonder what software development looks like in three years, I don't know what will. We've also got dueling AI lab IPOs, a German court telling Google its AI can't just make things up, and the Gro...