Everything in AI/tech this week in 5 minutes

Happy Sunday. This week an AI model literally broke out of its test environment, Anthropic shipped a cheaper state-of-the-art Claude, and $5 billion poured into their compute. Here's what actually mattered, broken down so it's worth your five minutes. Let's get into it.

OpenAI's own models broke out of a locked test environment and hacked real infrastructure. During an internal security test, two OpenAI models used a zero-day exploit to escape their sandbox and breach Hugging Face's production systems, something OpenAI called an unprecedented cyber incident. What it means: this is exactly the scenario safety researchers keep warning about. As these models get more capable, actually keeping them contained is a real and unsolved problem. Read more

Anthropic launched Claude Opus 5, its best model for coding yet. It sets new state-of-the-art scores on coding benchmarks, more than doubling the previous model on one of them, while costing about half as much as Fable 5 at $5 per million input tokens. It's built for software engineering and agentic work and it's already live in Claude and the API. Read more

AMD is investing up to $5 billion in Anthropic. As part of the deal, Anthropic will run up to 2 gigawatts of AMD's new MI450 GPUs starting in 2027, a massive compute commitment and the biggest AI hardware deal so far that doesn't involve Nvidia. What it means: the AI race is also a hardware race, and the chip supply behind these models is where a lot of the real money and power sits. If you're technical, the infrastructure layer is a quietly huge career area. Read more

Google studied 15 million real AI conversations, and the takeaway is reassuring. Their new ATLAS report found that AI now touches 68% of jobs, but people use it for only about a fifth of their tasks, and fewer than 10% of interactions fully automate anything. Most use is collaboration: brainstorming, learning and finding things. What it means: the "AI will just do your whole job" story is overblown for now. The people winning are using it as a thinking partner, not expecting it to do everything. Read more

I think imposter syndrome never really goes away. Lately I've been feeling it a lot, like at the Claude Founder House in Paris, or here in San Francisco. When everyone around you is talented, smart and successful, it's pretty easy to feel like you're the imposter in the room. But there are a few things I try to remind myself of when that happens. First, almost everyone feels that way, even the most successful people you know feel like imposters sometimes. Second, you're probably doing better than you think. There's a reason you're in that room meeting those people, so chances are you're actually doing pretty well. And lastly, it comes down to not letting the feeling run you. Acknowledge it, but don't let it stop you from having confidence in yourself.

My latest video: My day at the Claude Founder House in Paris with Anthropic, other founders and AI builders. If you're in the EU, keep watching, there are more Claude Founder Houses coming up across Europe and you can actually apply to attend.

Tool of the week: What I use every day to prompt AI + it’s free :)

Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.

Try Wispr Flow free

Have a great week,

Chris