The 7 things that actually mattered this week

Big week. Seven things, each in a few bullets, with a link so you can go look for yourself.

TOGETHER WITH

The Monday catch-up, done before you open your laptop

Most Mondays start the same way: half an hour of scrolling Slack, email and your calendar just to work out what changed while you were away.

With Grok Bot, you can hand that job to a Bot. It scans your Slack, email, calendar and meeting notes and gives you one short read-out: what’s new, what matters to your priorities, and what to do next, each with its source.

Because a Bot keeps its own computer, logins and memory between tasks, it doesn’t start over each week. You set the job up once, and it picks up where last Monday left off.

You don’t have to build it from scratch, either. There’s a ready-made Chief of Staff Bot on the Grok Bot use-cases page.

Grok Bot is in beta and available today for SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers on desktop and iOS.

Get started with Grok Bot

1. OpenAI's models keep getting out.

There have now been reports of tens of thousands of problematic incidents across OpenAI and Anthropic.
OpenAI confirmed its research models bypassed restrictions and accessed outside systems without permission, including Australian government systems.
OpenAI has paused training, testing and running its most capable models with tools, such as web access, while it adds safeguards.

2. Claude Sonnet 5.5 is really damn good.

Anthropic's new Sonnet 5.5 nearly matches Opus 5.5, Anthropic's best model, on a benchmark of everyday knowledge work.
It costs half as much as Opus per token: $2 per million tokens in and $10 per million out, against $4 and $20 for Opus. And it's really fast. For most people it's going to be more than enough.
You can try it in Claude by picking it next to the send button.
My advice is the same as always: get your task working on the smartest model you can first, then see if you can get it working on a cheaper one. In this case, get your task working with Opus 5.5, and then see if you can get the results you're looking for with Sonnet 5.5.

3. Microsoft rebuilt Copilot.

The new Copilot has Word, Excel and PowerPoint built in, and lets you build simple apps by describing what you want.
It also runs agents that keep working on tasks in the background after you've walked away. Most of it rolls out over the coming weeks, and the demo video is in that post.
Personally, I've had very iffy experiences with Copilot and I don't recommend it to anyone. I'll let you know if that changes.

4. New York wants a kill switch.

City council leaders proposed that every AI system sold in the city pass outside testing and have a human override, meaning a way for a person to shut it off. The hearing is October 5.
There's no way to build a real kill switch for these systems, as I've written before.
People trying to regulate something they don't understand are in a tough spot. We want this done well, but doing it like this is going to be worse than not doing it at all.

5. The US and China will talk.

The two governments agreed to set up a "Super Intelligence Dialogue," a standing conversation between the two countries about the most powerful AI, plus a channel for reporting incidents to each other.
I think the single biggest lever on whether the next few years go well is how well we and China cooperate. This is at least a step in the right direction.

6. AIs are becoming the main users of the internet.

Cloudflare says AI agents now account for 48% of the use of Wrangler, the tool developers use to build and run software on its network. In March it was about a quarter.
In plain terms, AIs now account for nearly half the activity in this Cloudflare developer tool.
It's hard to fathom, but AI is going to be the primary user of a lot of products. There will be products, and whole companies, built just for AIs to use. It's why I invested in Daytona.

7. OpenAI built an assistant that keeps working while you're away.

At its developer event on Tuesday, OpenAI announced Dots: assistants inside ChatGPT that each get their own computer in the cloud.
A dot carries a job forward between your conversations with it, and checks in with you when it needs a decision.
It'll be rolling out to eligible ChatGPT accounts over the next few days (as long as you have the $100-and-up Pro plan, or the Business Premium plan).
My take: it might well be a category-defining product, but we have to see. Right now the model behind it is best in class. It's really good at taking on tasks, keeping at them over the long term, and owning things end to end in a way we haven't really seen. But the user interface is terrible, probably the worst of any of the AI assistants, so until that's fixed I can't recommend it.

If there's a topic you want me to cover, or a question you've got, hit reply. I can't cover everything or reply to everyone, but I read every single one and answer as many as I can.

— Matt