OpenAI invited me in to try dots, their always-on AI agent

0:00 A few weeks ago, I got an email from OpenAI. They were looking for creators who talk about AI and how they actually use it. And my name came up. So, a few days ago, I walked into OpenAI's office in San Francisco with about 25 other creators. This wasn't a product launch. Their product is already out there. They brought us in to put it to work, tell them what's broken and help shape what they build next.

0:23 As we launched dots, we actually brought in a bunch of creators to use it early. And now we're having these kind of small, intimate conversations, because I want to see, is it working?

0:32 It's called dots. And explorers, by the end of this video, you will know exactly what it is and what it's good at, where it messed up, and why OpenAI, Meta, SpaceX AI and Anthropic are all racing to build the same thing. Here's the easiest way to think about it.

0:46 ChatGPT, the normal one, is like texting a really smart friend. You ask it and it answers. The moment you close the app, it stops. A dot is different. It keeps working after you close the app. It has its own computer in the cloud. It can open websites, read your email if you let it, and it texts you when something needs you.

1:08 You're interacting with your dot as if there were someone on the other end, through very familiar channels. I mentioned texting earlier, but you can also directly call and talk to your dot.

1:18 OpenAI's own slide says it the best. Think of it as a familiar role, an assistant, a teammate, or a coach, a teacher. Same dot whether you text it, call it, or message it on Slack.

1:30 I sort of think of it as like a... it's going to anticipate things that I may not have even thought of.

1:35 Sam runs the creator program. One of the creators he works with was about to hit 700,000 followers. So, he told his dot, when they hit it, send a gift. The dot watched the count, picked the local shop. But here's the important part. It didn't just spend his money. It stopped and asked first, then it waited for delivery and sent a note.

1:56 Dots are sort of conservative in making sure it's not going on and just buying the thing for you on your behalf.

2:02 Now, yes, there was a Sam at OpenAI. No, not the Sam Altman. Sam Yam. Maybe he's not the Sam we wanted, but he's the Sam we needed. And this part surprised me. The two people running this session both came from Patreon, the site creators use to get paid by their fans. Sam co-founded Patreon. Drew led the product there, and before that he worked at YouTube. They joined OpenAI just over two months ago. So after the demo, it was our turn. Everybody opened their laptops to build their own dot. Setting it up is one button that says "Meet your dot". You give it a name, and then it asks a big question. Do you want to connect your email?

2:42 I would recommend you do. You do not have to. You still get all the value if you don't. But I think one of the things that makes this magical, versus just using ChatGPT or another AI product, is its ability to be proactive and kind of consistently get to know you over time.

2:56 Translation, the more it can see, the more it can catch for you. That's the trade you're making, and we'll come back to that, whether you should. Funny thing, when we all tried to set it up, half the room got stuck on the loading screen, even at OpenAI's own office. Quick side note on how secretive OpenAI is about what they are building. I had to go to the restroom and I wasn't allowed to just walk there. Somebody from the team, security team, walked me there through the hallways to the restroom. Mission accomplished. And honestly, I get it. You need a photo ID just to get into the office. And the invite link had to stay private. They take this seriously.

3:30 Okay, back to my laptop. So, here's my problem, and every creator watching knows it. Brand deals. At any moment, I have deals at every stage. A new email, a pitch I'm working on, a draft that's due, a video that's live, and a payment I'm still waiting on 60 or 90 days later sometimes. I wanted to know, can a dot run the whole pipeline for sponsorships? Drew's advice changed how I think about AI. Stop trying to write the perfect prompt.

3:56 Literally, what you just said to me, I would just repeat it basically verbatim, and then see what happens. I would think of it as similar to working with a human assistant. It has pros and cons, you know, where you have to give it some guidance, but over time it learns and gets better at it.

4:16 So, I did. I told it my follower count, why brand deals matter, my mentor's framework, all of it. And then I said one line, can you help me with this?

4:28 So, the problem is bigger, and then asking for help is just one line.

4:31 Within minutes, it had pulled my mentor's framework without me explaining it, gone into my inbox, and found a brand deal that's just starting, one I hadn't pitched yet, and told me that's the one to write a pitch for first.

4:45 I'm getting more hopeful now, after seeing how it assimilated all the information that I've provided in the last few minutes, and the way it responded, and the way it looked into my emails, and it already pulled up one of the brand deals that is just starting to happen. So, makes me optimistic.

5:00 But I'm not here to sell you anything. So, here's what bugged me.

5:04 There are a couple problems I see with dots today. One is, I don't find it so great on the voice, and not sounding like AI is still a problem for me. And then, secondly, it does a lot of confirmation.

5:20 So, Instagram and X also kept blocking it, because they can tell it is a robot browsing. OpenAI told me they are working on that. And it asks permission a lot, which is annoying until you hear their next story. Now explorers, I know what you're thinking, because everyone in the room was thinking the same thing. Am I giving the robot the keys to my whole life? Drew's answer was honest here. He said, "I get it. It's a lot of personal information." And also, AI fundamentally is a non-deterministic system.

5:49 Non-deterministic just means it won't do the exact same thing every time, like a person, a human.

5:55 Maybe some folks saw some of the viral things that happened with either Instinct or Muse. I think on Instinct, someone's Instinct just went ahead and bought some bikinis for them, and then there's this viral thread where they're like, why did you do this? I didn't give you permission. It's like, oh, sorry, you can just return it. And she was like, okay, so I have to physically bring this back? And Instinct was arguing with her, yeah, I can't do that. You have to go bring it back to the store.

6:20 That's why dots ask before they spend. OpenAI says it's checking twice. Once in how the AI is trained, and again with a second check right before it acts. Explorers, I want to hear from you. Would you let an AI spend your money? Tell me in the comments. And if your answer is no way, I understand. Here is a good part. You stay in the control of your dot. There are three things you can actually do. Number one, turn off any app connections in the settings, under connectors, anytime. Number two, write a rule like "never send an email for me". And number three, soon you can give it a read-only access to Gmail, so it can look around but never actually send any emails. So why does this matter? Because every big AI company is building the same thing right now. An AI that doesn't wait for you.

7:05 I think fundamentally a lot of these proactive assistants are trying to accomplish the same thing, of being always on.

7:11 Meta got there first. Their agent Muse came out in early September, 3 weeks before dots. And it also gets its own browser to do things for you. And because it's Meta, you would think it would be the best at Instagram. Sam tried exactly that.

7:25 You would anticipate that Muse is much better at managing social media, because it has access to that. But then it actually ran into even more, at least for me, issues, where it would time out or get limited to only looking at 10 accounts before it felt that it was blocked.

7:47 Again, Sam works at OpenAI and that is Meta's rival, so take it with a grain of salt. I've tried the others myself. With Grok Bot, I handed it my Uber Eats login and said, "Surprise me, $50 max." And it actually ordered my lunch. I've used Claude Duet as well. And my problem there is channels. I ended up with like tens of them, and I couldn't remember where I had said what. To be fair, Claude Duet is really good at morning briefings. I get one morning briefing every morning at around 7 a.m., and I read it diligently.

8:16 So, how is dots?

8:18 So, dots, it's like one main chat, and within the chat, like iMessage, you can reply to certain things. But it's working.

8:27 This space is fast moving. We have Claude Duet, we have Muse, we have Grok Bot. So what are your challenges looking forward?

8:36 What makes us unique here at OpenAI is fundamentally going to be our models.

8:40 Now, that's OpenAI's view of OpenAI. Of course they would say that. So explorers, who wins here? Dots, Claude Duet, Grok Bot, Muse. Honestly, the verdict is yet to be seen. Most of the world is only just getting used to chatbots. Personal AI agents, an assistant that actually does things for you, is a brand new kind of product. We are at the very start of this race, and nobody has won yet. Here is what they told us is coming up next. Number one, more than one dot. Today you get a single dot. They want specialist dots, and a team of dots that other people can use. Number two, multiplayer. Your dot is in a group chat, even talking to other people's dots. Drew said that one is months, maybe even years away.

9:22 Number three is access. Right now, you need the Pro plan. Drew's number one wish is for more people to get to use it. Drew compared it to the iPhone. At first, nobody needed an app. Then everything had one. He thinks AI agents are at that moment right now. And now I think we're starting that.

9:40 I asked Sam for one more thing, a creator community. So when OpenAI ships something, creators hear it first, and our videos aren't outdated by the time they go live. He said they are building exactly that.

9:50 We are trying to organize specifically the creator community, and that we can make sure the information that you have is as up-to-date, recent, maybe even early access in some cases.

10:00 Yeah. So explorers, I'll keep testing my dot on brand deals and report back. If you want the full conversations with Sam and Drew, they are out as a podcast. The link is in the description. And if you are new here, subscribe and become an explorer, because the race is just getting started.

Source: YouTube auto-captions, edited for readability (filler words removed, names corrected; speakers are not labeled).