Adam here!
This was the week AI stopped waiting for instructions.
OpenAI gave ChatGPT "dots": agents that keep working after you close the chat. Shopify's AI now designs your whole store while you talk to it. And one startup's AI passed for human on live video calls.
The tools are starting to act, not just answer. Which makes the boring questions (what is it allowed to do, and who checks?) the interesting ones.
Let's get into it.
OPENAI
ChatGPT's New Dots Keep Working After You Close the Chat
OpenAI used its DevDay on 29 September to launch dots: always-on agents that live inside ChatGPT, run on its top model GPT-6 Astra, and keep working after you close the chat.
Unlike a chatbot, which answers and forgets, a dot holds a standing job. It gets its own computer in the cloud (so it can browse, build files and use apps without touching your machine) and connects to more than 4,000 apps, with Slack and Microsoft Teams built in alongside ChatGPT itself.
Your first dot is included on Pro and Business Premium plans in eligible markets (Australia made the list), with an Enterprise beta admins can switch on. Rollout can take a few days to reach your account
Chatting with your dot doesn't count against your usage limits, though the tasks it runs draw on your plan's allowance
Background work is read-only by default: a dot can research and prepare, but sending, editing or paying needs your explicit rules, and things like password changes always stay with you
My honest read: this is the first mainstream product built around giving AI a responsibility instead of a task. That's useful and slightly unnerving in equal parts, and OpenAI's own launch notes say dots can still make mistakes, so review anything consequential. The week one play: give it a job where a mistake is annoying rather than expensive.
Try this: set up your first dot with one read-and-report job and nothing else:
You are my client watch dot. Your one responsibility: watch my inbox and calendar for anything a client is waiting on.
Check each morning at 7:30am and send me a short report with three headings: needs a reply today, commitments coming up, nothing needed.
You may read only. Do not send, edit or delete anything. If something looks urgent, flag it and ask me.MORE AI NEWS
SHOPIFY
Shopify's Canvas Designs Your Whole Store by Chat
Shopify launched Canvas on 1 October: a visual workspace where you redesign your store by describing what you want to Sidekick, its built-in AI assistant.
What this could mean for you:
You describe the change ("make the homepage feel more premium and move the reviews up") and watch it happen to your real store code, live, instead of fighting a theme editor
You can click any element to tweak it directly, and preview how pages look across screen sizes before anything ships
It's desktop-only for now, and third-party themes and app blocks aren't supported yet, so heavily customised stores should wait a version or two
The practical rule: start with the one page you've been putting off (usually the about page or a tired product page), and check the mobile preview before you publish. Nothing goes live until you approve it.
Gemini 4 Argon Might Be the Best Model Yet. You Can't Use It
Google announced Gemini 4 Argon on 30 September, calling it its most capable model for long, complex work. The headline number: it can produce up to 1 million tokens of output in a single pass, up from 64,000. In plain English, that's a book-length piece of finished work from one instruction.
For now it's only available to vetted cybersecurity defenders through Google's Fairwind Program; everyone else waits for a phased rollout
When it opens up, access starts with paid API customers and Google AI Ultra subscribers, at an announced US$2 per million input tokens and US$10 per million output
It's the third frontier model this year to launch to a closed, trusted group first. The "everyone gets it on day one" era is quietly ending at the top end
The practical rule: build your workflows on the models you can use today. The leapfrogging at the top never stops, and a system that works on this month's models will simply work better on next month's.
AI VIDEO
An AI Just Passed for Human on Live Video Calls
AI startup Tavus previewed Griffin, a model that joins a live video call as a lifelike person. It watches, listens and reacts in real time, with a delay of about 0.43 seconds, and it nods along mid-sentence instead of waiting its turn.
In Tavus' own test, 26 of 54 people (48%) on a one-minute video call believed they were talking to a human, up from 2.4% for its previous model
It's limited to trusted testers while Tavus builds disclosure and safety features, and those numbers are the company's own, not independently verified
The upsides are real (a patient tutor, company for an ageing parent), but the same technology is a gift to scammers
Try this instead of a prompt this week: agree a code word with your family and your team. Any surprise video or voice call asking for money, passwords or gift cards gets verified by calling back on a number you already know. Seeing is no longer believing.
TWEETS OF THE WEEK
MORE AI IN ACTION
AI SYSTEMS LAB
If the lead story left you wondering what job to actually give a dot, Thursday's Lab call answered exactly that: your first dot, and the one job to give it.
We built a client watch dot live on the call: Gmail and Calendar only, a 7:30am report under three headings, no sending and no editing. We also covered the four lower-risk starter jobs, the seven-part responsibility brief, and how to correct the first report so the second one is better.
The rule from the call: name one responsibility, set the rules, watch three runs. Read-and-report can stay your permanent setup, and for most businesses it probably should (for now).
The full recording, the brief template, and this week's job are in the Lab.
