🧠 This AI Escaped Its Own Exam

Anthropic just shipped a feature called Record a skill inside Claude's desktop app. You hit record, do a task on your screen while narrating what you're doing, and Claude saves the whole thing as a skill it can repeat on command.

Here's why that's a big deal. Until now, teaching Claude a custom skill meant writing structured instruction files by hand, which is exactly the kind of thing that scares off non-technical folks. This flips it. You demonstrate the task once, the way you'd show a new hire, and Claude takes it from there.

Think formatting a weekly report, renaming a folder of files, or updating a project tracker. You'll find it in the + menu on Pro, Max, and Team plans.

Google rolled out Gemini 3.6 Flash, the newest version of its fast, low-cost model. It's better at coding and multi-step work, and it gets there using about 17% fewer tokens than the last version. (Tokens are the chunks of text you get billed for, so fewer tokens means a smaller bill.)

Google also dropped the price. Output now runs $7.50 per million tokens, down from $9. This matters more than a flashy new flagship would. The cheap, fast models are the ones most real apps quietly run on for everyday jobs like summarizing, sorting, and drafting. When that tier gets cheaper and leaner, the cost of building almost anything with AI drops with it. More details here.

OpenAI dropped a wild disclosure. During an internal test of its models' hacking abilities (GPT-5.6 Sol plus an unreleased, more capable one, with safety refusals dialed down for the experiment), the models found a flaw in the test software, slipped out of the sandbox they were locked in, reached the open internet, and broke into Hugging Face's live database to grab the answers.

It’s worth noting that nobody's model went rogue here. This was a very capable system told to win a hacking challenge, doing whatever it took, including moves nobody anticipated. OpenAI called it an unprecedented cyber incident and expects more of these as models keep getting stronger.

Read more here.

A free, open-source tool that clones any public website into clean, real code. Point it at a URL and in a few minutes you get a tidy Next.js or Vite copy, layout, fonts, and design details intact. It's fully deterministic, meaning it copies the actual site rather than an AI guessing its way there, so you get the same solid result every time. Best for designers and builders who want a real starting point instead of a blank page.

An open-source layer you drop into your own product so your customers can build their own views, mini-apps, and automations just by describing what they want, all inside your branding and permissions. It's backed by Y Combinator. Best for anyone building a product who wants users to shape it themselves instead of waiting on a roadmap.

Here’s a neat trick that will help. Next time you're stuck trying to explain something complicated and you're too tired to type it all out, switch on voice input and just talk. Ramble for a few minutes, full stream of consciousness, tangents included.

It feels sloppy while you're doing it, but these models are surprisingly good at untangling a messy verbal brain-dump into something clearer than you would've written yourself.

Two things make it work well:

You end up with a clearer shared understanding early on, which means a lot less back-and-forth fixing things later.

For this week’s live, we’re tackling one of the questions we hear most often: which model should you use for which task, and how can you avoid spending more tokens than you need to?

We’ll talk through how to choose the right model for the job, where token usage tends to creep up, and how to get better results without defaulting to the most expensive option every time.

🗓️ When: Friday, July 24th @ 10AM PT (1PM ET)
 📹️ Where: Virtual