Curious Companion: WTF are OpenAI Dots?

In this episode I break down OpenAI’s latest rollout, Dots, an always-on agent that lives inside ChatGPT and keeps doing things even after you’ve closed the chat. I cover what a Dot is, how it works, the logistics around pricing and access, the rules and guardrails, and what Dots can do. I also share my thoughts on abstraction in the agentic era and what the launch of Dots (and Meta Muse) speaks to in the bigger picture.

LISTEN TO THE EPISODE

Full Transparency

Today we’re talking about OpenAI’s latest rollout: Dots.

OpenAI announced Dots on September 29th at DevDay, stating that it runs on their most capable model, GPT-6 Astra. You can read the release here.

In true Maestro fashion I’m here to be nothing but honest with you, and thus in full transparency I must tell you that I’ve yet to actually try Dots. I don’t have a paid account with OpenAI, nor do I plan on getting one, so I don’t even have access to Dots.

To that end, this episode is an explainer based on research I’ve done, including watching a ton of YouTube videos shared by power users.

As per always, your best bet for truly learning about the thing is to get in there and try it out firsthand, buuuuuut I did want to make an episode to keep us curious folks in the loop, so here we go.

So…What Is a Dot?

A Dot is an always-on agent that lives inside ChatGPT.

Remember, an agent is AI that can actually DO things, not just answer questions.

You give it a task and it goes and executes it. The notable thing about Dots is that it keeps running and making progress even after you close the chat.

For my folks in the crowd who like details, the product name is Dots, and the individual agent is called a dot.

Each Dot is shown as a colorful character you can customize, and it has a name. OpenAI’s own demo Dot is named Alfred. NGL, something that bothers me about Dots (and Meta Muse) is the personification of these AI agents.

I think it’s a very slippery slope, and it’s definitely a clever marketing tactic by these companies. When you make the avatar cute and unsuspecting it completely changes how users interact with it and how they view it.

How It Works

The first step in using a Dot is to set it up on desktop or browser. From there, you can access it on mobile whenever you want.

Communicating with your Dot works just like with a chatbot, aka it uses natural language (which to me continues to be the most mind-blowing part about AI). The actual conversations do however take place in a separate area inside of ChatGPT, not in the standard chat thread.

You can talk to or message your Dot:

Inside ChatGPT
Via direct messages or channel tags in Slack and Microsoft Teams
Through a voice/phone call interface (this is very weird to me)

Dots can be accessed from the sidebar and will open up a new interface when selected.

Something unique to Dots is that the entirety of your conversations happen in a single chat thread. Like…all of them. Users have reported this takes some getting used to and can be a little confusing for organizational purposes. My take is that the design seems very intentional as OpenAI continues to push towards more agentic workflows where you simply tell the agent what you want and it handles the rest.

When you first set up your Dot, it has access to nothing. You must connect what you want it to touch. Dots can connect to over 4,000 apps (such as Gmail, Google Calendar, Slack, and Microsoft Teams) via the ChatGPT plugin/app ecosystem.

From there, you give it a goal, set the boundaries, and you’re off to the races.

Dots runs on GPT-6 Astra, OpenAI’s most capable model, which allows for additional safety and privacy safeguards.

Logistics

Like I said earlier, I don’t have a paid plan on ChatGPT, but even if I did I’d have zero plans to try Dots as it’s only available on the $100/month Pro plan and higher.

Worth noting, it’s also not available in the EEA, Switzerland, and the UK, even on Pro plans.

A Dot has no price of its own, rather It’s bundled into the usage plan, with caveats:

Conversations with your Dot, text or voice, don’t count toward ChatGPT usage limits.
The Dot doing actual work (performing actual tasks) is considered “deeper work” and each plan includes a separate allowance for this,
Anything a Dot hands off to Codex or ChatGPT Work draws from your plan’s usage.

Each Dot runs inside its own dedicated cloud-hosted computer and browser environment, allowing it to handle tasks (like public URL scraping, coding via Codex, or research) even when your personal laptop is turned off. This is actually pretty cool.

Lastly, you get one Dot per OpenAI account (which I think is probably a good thing for decreasing complexity and decision making for the average users), though OpenAI says that in the future you’ll be able to add more Dots.

Rules and Guardrails

So what’s keeping your Dot in check? Here’s what OpenAI has in place:

Custom Rules: For each kind of action, you decide whether it’s allowed, needs your approval, or is blocked.
Auto-review: An automatic safety check that runs before a Dot takes certain risky actions, separate from the rules you set. For example, extra checking on actions that could affect your accounts or share information.
Some things never get delegated. Password changes are OpenAI’s example.
Built-in safeguards against malicious instructions (aka prompt injections). This matters because an agent reading content from thousands of connected apps can be handed bad instructions hidden in that content.
A monitoring system that can pause or stop a Dot if it detects a safety problem.

Worth noting, OpenAI is always going to cover their ass, and thus Dots comes with the disclaimer that “Dots can still make mistakes, including when following your rules, so review anything consequential.”

What Can Dots Do?

Again, I haven’t actually tried Dots, so I’m not speaking from firsthand experience here, more so summarizing how I’ve seen and read about folks using it.

I will say that, as with most things related to AI, this feels like another backwards marketing release, with the folks at OpenAI saying, “Here’s a solution, go find a problem.”

Overall, I’m viewing Dots as:

A personal assistant for the average user (though I’m not sure what kind of average user is paying $100/month for ChatGPT)
An executive assistant for the power user
A project manager for the coder/developer

For my nerds in the crew, S\so much of where I get my information from is folks using AI for coding and the like, and they’re really using Dots for managing multiple agents and Codex projects.

My Thoughts

Real talk: I don’t know how truly useful Dots are, and it received mixed reviews in the various sources I looked at.

One review that stood out to me was a YouTube video by Nate Herk comparing Meta Muse and OpenAI Dots. Nate stated that Dots felt rushed, and like OpenAI released it to keep pace with Meta Muse. He noted that in the near future he expected to see updates to Dots that mirrored the functionality of Meta Muse. This would surprise me exactly zero percent.

Of note, for those of you who just said “WTF is Meta Muse”, I did an entire episode on it a few weeks back. Check it out here.

Let’s Talk About Abstraction

If we zoom out for a second, a single thing really stands out to me regarding the release of both Dots and Meta Muse: abstraction.

Abstraction is hiding the details of how something works behind a simpler way of using it, so you can work at a higher level without handling the steps underneath.

It’s not at all inherently a bad thing, nor is it unique to AI. It is, however, something worth noting.

As we continue to move more and more into the agentic era, it brings with it more and more abstraction.

When generative AI first rolled out, prompting was so important. The user was really involved in laying out the steps, and one’s ability to do so greatly impacted the quality of the output.
Now, all that friction is virtually gone, and it’s simply about telling the robot the outcome that you want and letting it do its thing.

Again, it’s not inherently a bad thing or a good thing, just something that is definitely worth noticing.

How I Used AI This Week

Each episode I share a quick example of how I used AI that week.

This week, my use case is most notable for what LLM I used and how I used it, as opposed to it being some super creative or techy use case.

This time I was trying to remember a website password website that one of my clients shared with me. I knew the client had shared it with me somewhere, and thought it might be in an email, so I used Gemini inside of Gmail to search for it.

Spoiler, Gemini didn’t find it because said password wasn’t in an email (I found it later in a different messaging platform we use), but it was nice to be able to use natural language to explain what I was searching for instead of trying to think of the exact perfect keywords to search.

Bringing this episode full circle, how I used Gemini is very much a simplified version of the way that Dots and Muse work, so, yes, I can absolutely see the utility. I just don’t have a zillion use cases in mind that have me chomping at the bit to use either agent.

Da Wrap-up

Dots is OpenAI’s always-on agent that is seemingly designed to serve as an executive assistant to ChatGPT users. I’ve yet to try it out first-hand, so I can’t speak to how good or bad it is, but I will say that it is definitely another step forward for agentic AI and has created yet another level of abstraction. I have no plans on trying Dots, again because I don’t pay for ChatGPT, but we’ll see what happens if Anthropic comes out with their own version.

As always, endlessly grateful for you and your curiosity.

Catch you next Thursday.

Maestro out.