Latest / Now Shipping: A Mind the Product podcast / Everything you need to know about Dots: OpenAI’s new agent

Transcript

Everything you need to know about Dots: OpenAI’s new agent · Now Shipping: A Mind the Product podcast

• 0:01
  SPEAKER_00: Now, just last week, we reported on how Muse, Meta's personal agent, became number one in the app store, and how we started to see its reach go beyond its own app into places where people actually shop.
• 0:14
  SPEAKER_00: Now it seemed like Muse was off to a giant head start on being the personal AI agent that people turn to for their home lives.
• 0:21
  SPEAKER_00: And on Tuesday of this week, at their dev day event, OpenAI stepped into the race with their own personal agent offering, Dots, which was positioned to help you in the workplace, but also at home too.
• 0:34
  SPEAKER_00: Now, we talked about personal agents and the implications for product people when Muse launched and even in last week's episode.
• 0:41
  SPEAKER_00: And there's a lot more here for product people than just, you know, OpenAI has an agent too.
• 0:47
  SPEAKER_00: And there's one thing that happened the day before the launch that changes how I read the whole story.
• 0:53
  SPEAKER_00: So let's get right into it.
• 0:55
  SPEAKER_00: Now, on Tuesday, September 29th, Sam Altman is on stage at Dev Day and he introduces Dots.
• 1:03
  SPEAKER_00: Now, OpenAI's description is that Dots is a collection of remarkably capable, always on agents built to handle everything.
• 1:11
  SPEAKER_00: As Altman put it, it's like having an AI helper that always has your back, inspired by the cool versions of what we all watched in movies, which sounds pretty intriguing, right?
• 1:22
  SPEAKER_00: Like the question is, is that actually real?
• 1:25
  SPEAKER_00: Is that real life?
• 1:26
  SPEAKER_00: Now, here's what it all means in practice.
• 1:29
  SPEAKER_00: You create a dot in the ChatGPT desktop app and you name it, it gets its own cloud computer with its own browser.
• 1:36
  SPEAKER_00: So it's not running on your laptop.
• 1:38
  SPEAKER_00: You connect the apps you want it to touch, and there are more than 4,000 available through Chat GPT's plugins.
• 1:44
  SPEAKER_00: Then you talk to it in ChatGPT or in Slack or in Microsoft Teams, whatever it is that you use.
• 1:52
  SPEAKER_00: You can even voice call it.
• 1:53
  SPEAKER_00: And texting is right around the corner.
• 1:56
  SPEAKER_00: And here's what makes it different from the agents that we've covered so far.
• 2:01
  SPEAKER_00: It's always on.
• 2:02
  SPEAKER_00: So you close your laptop, it keeps going.
• 2:05
  SPEAKER_00: And when you're not asking it for anything, it goes looking for ways to help.
• 2:09
  SPEAKER_00: Now, OpenAI says the magic of dots is when they bring you work done the way you would do it, sometimes before you even think to ask.
• 2:17
  SPEAKER_00: And that's super interesting.
• 2:19
  SPEAKER_00: I mean, I would love to have an agent like that, right?
• 2:21
  SPEAKER_00: It's pretty compelling.
• 2:22
  SPEAKER_00: But I've noticed my Muse agent, buddy, he does proactively come to me with suggestions or flagging things that it thinks I should know.
• 2:30
  SPEAKER_00: So I don't know if dots is completely unique in that aspect here, but again, it's very compelling.
• 2:35
  SPEAKER_00: Now, OpenAI has showed a handful of demos in that dev day event, and some of the use cases that they showed off, they were interesting.
• 2:42
  SPEAKER_00: So, in one example, a bug shows up in Slack and a dot starts investigating right away.
• 2:49
  SPEAKER_00: Design lands, a dot turns it into a working app while the team stays focused on the customer feedback, a planning cycle kicks off, and then a dot keeps everybody in sync as the deadline gets closer.
• 3:00
  SPEAKER_00: And then there's the demo that I think would relate to product people here.
• 3:04
  SPEAKER_00: There was a fictional product launch for a wellness app called Aster.
• 3:08
  SPEAKER_00: Now, a dot was reading the launch brief in the team's launch channel and was watching for updates.
• 3:14
  SPEAKER_00: Then a product update lands.
• 3:16
  SPEAKER_00: The launch screen changes to a daily view with a score, and the dot pops up and says, Hey, quick heads up.
• 3:22
  SPEAKER_00: This changes the launch screen.
• 3:23
  SPEAKER_00: All five of our slides still say you're weak.
• 3:26
  SPEAKER_00: Want me to update the copy, the visuals, the speaker notes?
• 3:30
  SPEAKER_00: The human in this example says, Yes, show me a couple of options, keep the same look.
• 3:34
  SPEAKER_00: The dot comes back with two versions of the slide, it recommends one and saves both.
• 3:38
  SPEAKER_00: Now, in that situation, nobody wrote a ticket, nobody pinged marketing, a change in the product got noticed and turned it into downstream work, and a human just made the call between two options.
• 3:49
  SPEAKER_00: So there's still a human in the loop, of course, thank goodness for us product people, but it takes a lot off of a product person's time.
• 3:57
  SPEAKER_00: And here it seemed pretty seamless.
• 4:00
  SPEAKER_00: Now, if this sounds really familiar to say OpenClaw, I mean, remember when OpenClaw was the talk of the town?
• 4:07
  SPEAKER_00: Then OpenAI essentially acquired OpenClaw when they hired its sole creator, Peter Steinberger.
• 4:13
  SPEAKER_00: Well, to me, dots sort of feels like a much more simplified, much more user-friendly version of OpenClaw.
• 4:20
  SPEAKER_00: Now, I do want to call back on something that we talked about back in July.
• 4:23
  SPEAKER_00: We looked at the AI layoff tracker.
• 4:26
  SPEAKER_00: I know, I don't want to bring it up right now, but if you remember the pattern where there were cuts was the coordination layer, that middle layer, the information brokers of sorts, right?
• 4:37
  SPEAKER_00: Well, dots was built exactly for that layer.
• 4:40
  SPEAKER_00: So if you're a product person and a part of the value that you felt you've always brought was that you're the person who knows what changed and knows who needs to hear about it.
• 4:50
  SPEAKER_00: I would think really hard about what happens when a dot does that at three in the morning without anybody asking it to.
• 4:57
  SPEAKER_00: Now look, don't panic.
• 4:58
  SPEAKER_00: Again, you notice there are humans involved in this process.
• 5:02
  SPEAKER_00: It's just that the human's value is different.
• 5:05
  SPEAKER_00: The person is the one that makes the calls that the dot can't, right?
• 5:10
  SPEAKER_00: Which of those two slides is right?
• 5:12
  SPEAKER_00: Should we even ship this change, right?
• 5:15
  SPEAKER_00: Like whose problem is it if we don't ship this change?
• 5:18
  SPEAKER_00: This is where product people really have the value.
• 5:21
  SPEAKER_00: Anyway, let's talk about the controls because, in a lot of ways, this is kind of the real product here, right?
• 5:27
  SPEAKER_00: Each dot lives on its own cloud computer, so connecting your actual computer is optional.
• 5:32
  SPEAKER_00: You choose what apps to connect.
• 5:34
  SPEAKER_00: Then you set what OpenAI calls custom rules.
• 5:38
  SPEAKER_00: And similar to Muse, for any kind of action, you decide whether the dot can just do it, whether it has to ask you first, or whether it should never do it.
• 5:47
  SPEAKER_00: There's an activity view in the desktop app that shows what your dot is working on and where each task stands.
• 5:53
  SPEAKER_00: There's an auto review step that checks consequential actions for risk and authorization before they run.
• 5:59
  SPEAKER_00: And there's something that they call misalignment monitoring, which looks for deception or attempts to bypass safeguards.
• 6:06
  SPEAKER_00: Now, certain things like password changes, they always need a human.
• 6:09
  SPEAKER_00: And at any point, you can change direction or tell it to stop.
• 6:13
  SPEAKER_00: Now, if you're building anything agentic, you might want to like hit back a minute or so and listen to that list all over again.
• 6:21
  SPEAKER_00: And notice how much of it isn't really about intelligence.
• 6:24
  SPEAKER_00: It's an approval model, a permission model, a status view, a stop button.
• 6:30
  SPEAKER_00: That's the user experience, right?
• 6:32
  SPEAKER_00: OpenAI could have shipped a smarter dot and still lost people the first time it sent an email that they didn't want the dot to send.
• 6:39
  SPEAKER_00: The trust design is what lets somebody hand over real work.
• 6:43
  SPEAKER_00: So, question for you and your team to ask what's the most consequential thing your agent can do without asking?
• 6:50
  SPEAKER_00: Now, if nobody can answer that or figure it out in, you know, five seconds, 10 seconds, you might want to really pause, get with your team, and think about questions like that and figure it out pretty quickly here.
• 7:03
  SPEAKER_00: Now, the demo day pitch sounded great, and and they even gave some customer stories out of the gate from customers that had early access.
• 7:11
  SPEAKER_00: There was a Dustin Fox, who has a 32-person real estate team who was thinking about rolling out Mac minis.
• 7:18
  SPEAKER_00: Now, his dot noticed, put together a setup plan, an equipment checklist, a meeting agenda, then offered to coordinate the Slack message, the team email, and the calendar invite.
• 7:27
  SPEAKER_00: There's a developer named Dan McAteer.
• 7:30
  SPEAKER_00: He said his dot proactively picked up on an invoice that he needed to send for freelance writing, pulled all the relevant details from the email thread, and drafted the invoice.
• 7:39
  SPEAKER_00: Now, after he approved it, the dot sent it as a PDF.
• 7:43
  SPEAKER_00: Then there was a developer, Will Weinbach.
• 7:45
  SPEAKER_00: He said his dot turned his scattered emails, messages, meetings into clear next steps, then carried out the actions that he approved.
• 7:53
  SPEAKER_00: Now, these use cases are good, but the proof that was shared in these stories, it's kind of like admin work.
• 8:00
  SPEAKER_00: I mean, and that's useful, right?
• 8:01
  SPEAKER_00: I but not sure that we saw anything necessarily monumental, at least not quite yet.
• 8:07
  SPEAKER_00: And people can already automate work like this today, right?
• 8:10
  SPEAKER_00: So for example, Winebach said that he and his partner had already built a similar workflow using custom assistance, scheduled tasks, and notion.
• 8:18
  SPEAKER_00: But with a dot, he set it up in one prompt.
• 8:21
  SPEAKER_00: And maybe this is where the magic really is, because I know so many teams that put all sorts of work into setting up custom agent setups, scheduling tasks and automations between tools.
• 8:31
  SPEAKER_00: And maybe dots can be the path to condense all of that into a single prompt.
• 8:36
  SPEAKER_00: Now, on pricing, your first dot is included at no extra cost with pro and business premium plans ineligible markets.
• 8:44
  SPEAKER_00: That means you can't access it unless you're paying at least$100 a month.
• 8:49
  SPEAKER_00: Conversations with your dot, they don't count against your usage limits.
• 8:53
  SPEAKER_00: And OpenAI is offering extended limits for the first month and says paid options for more dots, more speed, or more work will soon follow.
• 9:02
  SPEAKER_00: Now I will say, in the same announcement, OpenAI changed its$200 pro plan.
• 9:07
  SPEAKER_00: The price stays at$200 a month, but starting October 30th, the included usage in Chat GPT work and codex goes from 20 times the plus allowance to 10 times.
• 9:18
  SPEAKER_00: And GPT 6 Pro messages in the chat dropped from 200 a week to 100.
• 9:23
  SPEAKER_00: Now they're giving current subscribers a one-time credit worth$2,500 to soften it, but that is something to note.
• 9:30
  SPEAKER_00: Now, it isn't completely free like Meta's Muse, you might notice, but even that wasn't completely free, but it's free the way most people would use it, as Meta says.
• 9:40
  SPEAKER_00: But you do have to be on a paid plan that starts at$100 a month.
• 9:43
  SPEAKER_00: So for most people, it might cut this out of being a personal agent for them.
• 9:48
  SPEAKER_00: So you may see more work-related use cases here for dots than personal use cases.
• 9:53
  SPEAKER_00: But I do know plenty of people that are already paying that$100 a month pro plan just for themselves.
• 9:58
  SPEAKER_00: So it's not the case with everybody that$100 might be over the barrier.
• 10:03
  SPEAKER_00: But it is more expensive than free, right?
• 10:05
  SPEAKER_00: Now, if you're pricing anything with AI in it, you're sort of watching a real live case study on how to package an agent.
• 10:13
  SPEAKER_00: Maybe it's free to start, you meter the work, but don't be surprised when your heaviest users start to push back on the limits.
• 10:20
  SPEAKER_00: Now let's talk about distribution.
• 10:22
  SPEAKER_00: So a few weeks ago, Meta launched Muse, Google has had Gemini Spark since May, and now there's OpenAI.
• 10:29
  SPEAKER_00: But look at where each one lives.
• 10:32
  SPEAKER_00: Muse lives in Meta's apps and WhatsApp, following you throughout your personal life.
• 10:36
  SPEAKER_00: Spark sits in Google's world, Dots lives in ChatGPT, but also Slack and Teams.
• 10:42
  SPEAKER_00: Meta's Muse is going after your home life.
• 10:45
  SPEAKER_00: Maybe OpenAI with Dots is really going after your work life more.
• 10:50
  SPEAKER_00: Now that connects to what we said last week.
• 10:53
  SPEAKER_00: I told you that distribution was half of why regular people haven't adopted agents yet.
• 10:58
  SPEAKER_00: Of course, if your boss mandates it, you're doing it.
• 11:01
  SPEAKER_00: You look at the enterprise version rolling out, an admin flips the switch, and the agent just shows up in your Slack or your Teams, right?
• 11:08
  SPEAKER_00: A dot can also start to manage tasks in Codecs and ChatGPT work.
• 11:14
  SPEAKER_00: So if it's an agent that can direct other agents, and 35 million people already are building with ChatGPT work and codecs every week, on top of ChatGPT's 1.2 billion weekly users.
• 11:27
  SPEAKER_00: This is a lot of people who will get a dot handed to them inside of a tool that they already open every day.
• 11:34
  SPEAKER_00: And if you sell B2B software, there's one more piece that came out of this dev day, right?
• 11:40
  SPEAKER_00: OpenAI launched a marketplace in beta with 32 partners, including Adobe, Figma, Sierra, Harvey.
• 11:49
  SPEAKER_00: Now, eligible US enterprise customers can apply part of that existing OpenAI spending commitment toward approved partner software.
• 11:57
  SPEAKER_00: So this is sort of a new procurement path, right?
• 11:59
  SPEAKER_00: Your buyer might soon be able to pay for your product out of an OpenAI budget that they already signed.
• 12:05
  SPEAKER_00: So if you're in that world, you want to find out how to get on that list, right?
• 12:11
  SPEAKER_00: Now, the wrinkle that I mentioned earlier on Monday of this week, this is the day before the launch, OpenAI announced it will not release GPT 6.1 Astra.
• 12:23
  SPEAKER_00: That was planned for October.
• 12:24
  SPEAKER_00: That's the next version of the model that powers dots.
• 12:27
  SPEAKER_00: OpenAI's head of safety system, Sashi Jane, came out and said that the model didn't quite meet the bar in terms of staying within scope and authorization.
• 12:36
  SPEAKER_00: Basically, the model was doing things without permission from users, so it's on hold.
• 12:41
  SPEAKER_00: Which means the day before OpenAI launched an always on agent, it pulled a model for being too willing to keep going without permission and for being too loose about telling you what it actually had done.
• 12:54
  SPEAKER_00: Now, with dots, remember one of the core promises is that dots can do work sometimes before you even think to ask.
• 13:01
  SPEAKER_00: Well, that scope piece, that authorization piece, those are really important things for an agent to fulfill this promise.
• 13:08
  SPEAKER_00: The more a dot does on its own, the more often it will do something that you actually didn't want it to do.
• 13:13
  SPEAKER_00: Now, this launch announcement uh didn't include reliability numbers, it didn't include error rates.
• 13:19
  SPEAKER_00: There's nothing on how often a dot does or might go off script.
• 13:24
  SPEAKER_00: Now, what's actually powering dots right now is GPT-6 Astra, which OpenAI says is its most aligned model, and which the register reports has hit OpenAI's own critical level for cybersecurity capability, meaning it can find and exploit vulnerabilities that nobody's disclosed.
• 13:42
  SPEAKER_00: So OpenAI feels really confident that that model is a good fit for right now.
• 13:46
  SPEAKER_00: But what does this all mean for you as a product person?
• 13:50
  SPEAKER_00: We talked about a couple things, like pricing for one, but I'll say there's a good chance your customers are about to send dots into your product.
• 13:59
  SPEAKER_00: Now, a dot has its own browser, it logs in, it clicks through things.
• 14:02
  SPEAKER_00: If your customer has one and your product is a part of their job, that dot is going to use it while the customer's asleep even.
• 14:09
  SPEAKER_00: So can your product survive that?
• 14:11
  SPEAKER_00: Can it get through your login, your onboarding, that modal that shows up on the third visit?
• 14:16
  SPEAKER_00: Are you one of the 4,000 apps in the connector list?
• 14:19
  SPEAKER_00: Or will the dot sort of look at your interface and guess on what to do because it's not really bot friendly?
• 14:26
  SPEAKER_00: And are you fine with the traffic?
• 14:28
  SPEAKER_00: Or are you gonna go the Amazon route that we talked about last week and not be very bot friendly?
• 14:33
  SPEAKER_00: Maybe I swear a year ago, we were talking about ways to protect our websites from bots.
• 14:38
  SPEAKER_00: Now we've got to be bot friendly.
• 14:40
  SPEAKER_00: Anyway, think about what it does to your numbers.
• 14:42
  SPEAKER_00: A dot that pulls in your customers' weekly report every Monday at 6 a.m.
• 14:46
  SPEAKER_00: sort of looks like a strange power user, right?
• 14:49
  SPEAKER_00: Now, if you measure success in sessions and clicks and time and product, agents are about to make those numbers very, very weird.
• 14:56
  SPEAKER_00: So start working out how to spot agent traffic flow now.
• 15:00
  SPEAKER_00: You want to know what share of your usage is from agents before somebody in leadership asks.
• 15:04
  SPEAKER_00: But that also leads to pricing.
• 15:06
  SPEAKER_00: How are you pricing your product right now?
• 15:08
  SPEAKER_00: Is it by user?
• 15:10
  SPEAKER_00: Do bots count as users?
• 15:11
  SPEAKER_00: And if it's just per human user, well, there might actually be fewer and fewer humans using your product, and yet your product might get double the use, triple the use, or more.
• 15:21
  SPEAKER_00: And you got to get paid, right?
• 15:22
  SPEAKER_00: So keep your eye on that and keep your eye on the packaging.
• 15:25
  SPEAKER_00: OpenAI gave the agent away and squeezed the top tier, and everyone else is watching to see if that sticks.
• 15:32
  SPEAKER_00: Now, if your roadmap has an AI feature and a pricing page between Muse's free agent and OpenAI's included agent, we might start to shape people's expectations on what they have to pay extra for and what they don't.
• 15:44
  SPEAKER_00: Now, finally, if you build on these models or buy agents built on them, scope and authorization is now an evaluation criterion, and it has to be.
• 15:53
  SPEAKER_00: Ask your vendors how they test for it.
• 15:54
  SPEAKER_00: Ask what happens to your product when the next model doesn't ship on time because OpenAI just showed you that can happen.
• 16:01
  SPEAKER_00: All right, that is going to wrap it up for this week.
• 16:04
  SPEAKER_00: I have a feeling there are going to be a lot of developments in the next week, and we'll probably hear more about dots, but I will be coming to you from a different setting because I'll be in Chicago for Mind the Products Big Conference, MTP Chicago.
• 16:17
  SPEAKER_00: We'll be gathering hundreds of product leaders from all over the world.
• 16:20
  SPEAKER_00: I have to prepare my fireside chat with Brandon Hunt.
• 16:24
  SPEAKER_00: He's the co-creator of Ted Lasso, the one and only Coach Beard.
• 16:28
  SPEAKER_00: I'm a little nervous for that one because I am a big Ted Lasso fan.
• 16:31
  SPEAKER_00: I guess I'll have to channel my inner Ted Lasso for that fireside chat.
• 16:36
  SPEAKER_00: If you're going to be there, definitely come say hi.
• 16:38
  SPEAKER_00: I would love to meet you in person.
• 16:40
  SPEAKER_00: But if you found this valuable, please subscribe, tell a friend.
• 16:44
  SPEAKER_00: And hey, if you have any suggestions for me, leave a comment below.
• 16:47
  SPEAKER_00: I will read every single comment that you leave.
• 16:49
  SPEAKER_00: I promise you that.
• 16:51
  SPEAKER_00: I will see you right here next week.
• 16:53
  SPEAKER_00: Once again, my name is Mike Belsito, and brought to you by the team at Mind the Product.
• 16:58
  SPEAKER_00: This is Now Shipping.