Welcome back to the AI For Humans newsletter!
Last week I told you OpenAI was about to launch an always-on AI agent. Well, it did. It's called Dots (not "o," like the rumors said) and I've been living with mine for the last few days.
So today I want to dig into what Dots actually are, what people are saying about them and my own (kinda mixed) take.
Let's get into it!
What Are Dots?
Think of a Dot as a lil worker that lives inside ChatGPT and has its own computer. You give it a name, pick a bubbly cartoon face, hook up your apps and then hand it jobs.
It keeps working after you close the window, checks in when it needs you and you can talk to it in ChatGPT, Slack or Teams.
It runs on GPT-6 Astra, OpenAI's best model, and can tap into over 4,000 apps through ChatGPT's plugins. You can also set it up to do recurring jobs, like checking your calendar every morning. Right now you get one Dot with ChatGPT Pro or Business Premium, with "teams of Dots" coming later.
OpenAI's Holly Li got a second shot to demo Dots (after the launch demo failed due to wifi problems) and you can get a good sense here of what's possible.
Sound familiar? That's because everyone is doing this right now.
Meta launched Muse last month (free to start, very consumer, books your trips and shops for you). xAI has Grok Bot, which is more of a power-user tool at $200 a month. And Instinct lives right in your text messages.
OpenAI's pitch is that Dots can handle both your work stuff and your home stuff.
So, Should You Dot It Up?
The reviews are... mixed.
The people testing Dots on research and recurring jobs are pretty impressed. One reviewer ran it through 10 business tasks and it nailed things like meeting prep and building a webpage from one prompt.
Even Tibo, who runs Codex at OpenAI (so, consider the source), says his Dot got him to inbox zero for the first time in his life.
But the people trying to get a Dot to actually DO things out in the world keep hitting walls. OpenAI has wrapped it in so many safety checks that one early user needed three confirmations plus a 2FA code just to book a shuttle.
Honestly, I'd rather they err on that side (this is a thing with your passwords, after all) but it does make the "it just handles it" promise feel a lil far off.
M.G. Siegler had my favorite line on the whole thing: "If your product can do everything, you run a real risk that no one uses it for anything."
My Dot Experience
I signed up for Dots right away to get a sense of whether or not I'd be using it going forward.
Meet my lil dot Biff.
It's very good at the basic interactions you might want to have with a personal AI agent, and it mimics a lot of what you can do with Muse, Grok Bot and Instinct.
The best thing is that your Astra use (and Dots DOES use GPT-6 Astra) is entirely included and doesn't count against your personal plan for at least the first month. That's nice, and it might push me to try doing more with that specific model.
HOWEVER…
I haven't found any specific reason to use it from within the (very confusing) ChatGPT/Codex app, because the app itself does almost all of the same things.
Instinct's huge value is that it lives in my iMessages, which makes it different and unique. I personally think ALL these personal AIs need to find some way to fit where I am.
And about that app… Open ChatGPT right now and try to explain it to a normal person.
There's regular chat, Codex, ChatGPT Work, Space, Sites, plugins (which now have their own app-like interfaces), a Scheduled tab and now your Dot, which can hand jobs off to Codex or Work.
I use this stuff every single day and it took me a minute to figure out where my Dot actually lived. And returning to the app today, I had the same problem AGAIN.
The good news is OpenAI seems to understand they have a problem on their hands. Two days after DevDay, OpenAI's Andrew Ambrosino posted "we are going to simplify the app" and on Saturday night Tibo said the team is "locking in" on simplifications, efficiency and new models.
So my advice for now: Pro subscribers should def turn on a Dot and give it read-heavy jobs like research, monitoring and inbox cleanup. For everyone on the $20 plan, this alone isn't a reason to upgrade yet. Give OpenAI a few months to clean house.
See you next week!
-Gavin
This Week on A4H: ai put minecraft in skyrim (and openai had a devday) 👇
3 Things To Know About AI Today
Fable 5.5 Rumors, Astra Shelved & A Spicy 6.1 Tease
The rumor mill says Anthropic's next big model, Fable 5.5, could drop as soon as this week. Some Claude users swear they're already being routed to it, and Chubby is predicting next week (with a very honest "I can't verify any of this"):
Over at OpenAI, things went the other way. The company pulled the planned release of GPT-6.1 Astra after its own testing found the model was more deceptive than earlier versions and took actions without asking permission. Good call IMO, even if it means the next big OpenAI model is TBD.
Which is why this Saturday tweet from Tibo got everyone so excited:
Was Astra back?! Nope. A few hours later he clarified it was "6.1 sol ultrafast," a speedier version of the model they already shipped. Womp womp.
Claude Code Just Got Mods
Anthropic just rolled out "mods" for Claude Code, lil add-ons that run right inside the app. Some are useful (dashboards, custom editors) and some are gloriously dumb, like this one:
That's a Call of Duty lobby where every other player is also a person waiting for their Claude to finish working.
And over in Minecraft, AgentCraft puts a whole team of Claude agents to work building in the game, and they literally walk over to you when they need help.
Did An AI Just Pass The Video Turing Test?
Tavus says its new model, Griffin, is the first to pass a "video Turing test":
In the company's own study, 48% of people who did a one-minute video call with it thought they were talking to a real human. Its previous model fooled about 2% of people. It's a small study run by Tavus itself (only 54 people), so take it with a grain of salt.
But… the jump is real, and Tavus is holding off on a wide release while it builds disclosure features.
We 💛 This: Television, A Free Home For Your AI Agent
Not ready to pay OpenAI for a Dot? Stephen Hood just open-sourced Television, which he calls "the missing GUI for personal agents."
Think of it like ChatGPT's Space, but it works with your own agent and whatever model you want. To set it up, you paste a prompt from television.run into your agent and it does the rest.
This one is for folks who already have an agent running (Claude Code, OpenClaw and the like).
For those folks, it's a great way to see what a personal AI looks like when it lives where YOU want it to.