Diogo Almeida, who at OpenAI helped build the research that taught ChatGPT to follow instructions, launched TypeSafe AI this week with a new kind of model called Jev.

Instead of writing sentences, Jev only answers in strict, pre-agreed formats: a category, a yes/no, a number, a score. Because it physically can't answer outside those formats, it can't hallucinate, which is the AI word for confidently making things up. Every answer comes with a confidence score.

The numbers are the story: responses in 70–500 milliseconds (roughly 40–200x faster than the big chatbots) at a fraction of the cost.

What this could mean for you, eventually:

You won't use Jev directly. It's in early access for developers. The rule for the rest of us: chat models are for thinking with you, and this new breed is for deciding inside your systems. If you've got a developer on tap, the waitlist is open.

On Wednesday, Anthropic announced that Cowork and the Claude chat app are merging into one Claude. You no longer pick the right tool for the job. You ask, and Claude works out whether it needs to chat, research, or roll up its sleeves.

The useful bits that arrived with the merge:

Rollout: Pro and Max plans first over the coming weeks, Team and Free to follow. Enterprise admins get at least 30 days' notice.

My honest read: this is less a new feature and more the end of a confusing question. If you sat out Cowork because juggling two Claudes felt like homework, that excuse just expired.

Try this once it lands in your account:

OpenAI is testing Sponsored Agents: click an ad in ChatGPT and instead of landing on a website, you start a conversation with that brand's AI agent. Ask questions, get personalised answers, then click through if you're still interested. It's live with select US advertisers, kept separate from your normal chats.

The part that matters for small businesses is who gets to buy in:

Worth remembering: ads inside AI chat are brand new territory, and nobody, including OpenAI, knows the etiquette yet. Probably don't move your whole budget there. But your customers are already asking ChatGPT things like "who's good for [what you do] near me?"

Try this: ask ChatGPT three questions a customer would ask before buying from someone like you. What it says today is the conversation your ads would be joining. Sign-ups are at ads.openai.com.

Google's new experiment, CC, is an AI agent for the household: it runs on its own cloud computer with its own Google account, reads whatever each family member chooses to share, and turns the chaos into a shared daily brief.

It tracks the dates everyone forgets, fills out permission slips, builds shopping lists, and plans the week's meals, for up to six people. Each person controls what CC sees, and it asks permission before doing anything beyond the group.

It's US-only for now (18+, with a waitlist), so Australia watches from the sidelines again. But the pattern is the one to steal: the same "give AI one job with clear rules" playbook that works in business, pointed at the school-run.

Try this while we wait: forward the school newsletter and sports schedule to your AI of choice each week and ask for one combined list of dates, actions and deadlines. That's 80% of CC with tools you already have.

Speaking of giving AI real jobs: on Friday's live call we walked through the five AI agents worth building in any business: Chief of Staff, Sales, Marketing, Operations, and Research.

The shift that matters: stop asking AI questions and start giving it jobs. A question starts from zero every time. A job owns one recurring responsibility, already knows your business, and does the same useful work every week.

The rule that stuck: don't build five agents on day one. Pick the one job you most want off your plate (for most people, that's your inbox), brief it like a new hire, and nothing gets published, spent or sent without your yes.

The full recording, the 7-part agent brief, and this week's job are in the Lab.