AI gets more helpers as the question of control grows

AI moved closer to the workplace this week. It is no longer just about chatbots you ask a question, but about agents: AI programs that take on tasks themselves. Anthropic has given Claude the ability to lead up to 1,000 helpers on a single task, and in a test the method found 66 of 70 hidden bugs Claude can run up to 1,000 AI agents on one task. Webex wants to let agents sit in on meetings as permanent team members Webex lets AI agents sit in on meetings as team members.

At the same time, more warnings are being heard. A safety researcher is leaving OpenAI and believes the culture is not cautious enough OpenAI researcher quits: Safety culture is not good enough, and Microsoft's chief wants to treat advanced AI as a risky employee Microsoft chief: Treat advanced AI as a risky employee. Here are the week's big themes, and where you can read more.

Agents get more to do

The idea of putting several AI workers on one task keeps coming up. Claude Code Projects lets a coordinator split a large task across several sessions working at the same time, and is now open to Pro and Max users Claude Code Projects: more AI workers on one task. Musk's Grok Bot can get its own inbox and use it to create accounts and schedule meetings Grok Bot gets its own email address and can work more independently. According to the vendor, it takes about three minutes to set up with a plugin Grok Bot can get its own email address: here's how. That is convenient, but also a reminder that an AI can now act on its own.

Who will keep track of them?

Satya Nadella suggests that the most advanced models get limited permissions, are monitored constantly and can be shut down immediately. The controls must sit outside the model itself Microsoft chief: Treat advanced AI as a risky employee. Small towns are feeling it too: Granby in Quebec already uses AI for administration, while a formal policy is still being drafted Canadian town writes AI rules while employees already use it. And after criticism of ChatGPT for users under 18, parenting expert Emily Oster recommends conversation over control Teen ChatGPT draws harsh criticism while parenting expert advises conversation. OpenAI disputes the test, however.

The money and the freebies

Only 4.5 percent pay for ChatGPT, Gemini or Claude, and the 1 percent who use it most account for almost a fifth of usage Few pay for AI, but those who do spend huge sums. Perhaps that is why free access is being tightened: From October 9, free Gemini users will only get the lightest model, Flash-Lite Free Gemini users only get Flash-Lite from October 9. At Grok, image generation and the most powerful models require payment Grok: What you get for free, and what requires a subscription. In China, DeepSeek, Moonshot AI and Manus are raising large sums and aiming for the stock market Three Chinese AI companies raise large sums and aim for the stock market, but MiniMax has lost more than 80 percent of its value since March MiniMax has lost more than 80 percent while DeepSeek raises billions.

New models and tools

OpenAI has made GPT-6 Astra and GPT-6.1 Sol about 50 percent faster at no new price OpenAI makes GPT-6 Astra and GPT-6.1 Sol about 50 percent faster. From Europe comes Mistral Large 4, which can be used via API today Mistral Large 4: European giant model with 1 trillion parameters. Alibaba has released a faster image model whose weights can be downloaded freely, but the license has a catch Alibaba releases Qwen-Image-2.1-Turbo: Images in eight steps. And ChatGPT has gotten a Try On button so you can see clothes on yourself before you buy, but availability in Denmark is unclear ChatGPT gets a fitting room: Try On and Favorites have arrived.

When AI is used for something important

In health care, Professor Ismail Gögenur has developed a Danish AI model for colorectal cancer, built on data from 76,000 patients Danish AI model to help doctors choose cancer treatment. The results on fewer complications, however, come from the presenter's own account. AI also showed its limits in mathematics: OpenAI had to withdraw three of 722 manuscripts because of a sign error Sign error forced OpenAI to withdraw three math papers. A fast pace demands careful checking.

What can you do yourself?

Try using AI as a helper, not as an authority: Check answers that matter, and do not give an agent more permissions than the task requires. If you use free versions, keep an eye on whether features disappear behind a subscription. And if you have children who use ChatGPT, have a conversation with them about how they use it.