Transcript 0:00 Did you miss Clippy? Say no more. OpenAI has announced Dots, and my nostalgia immediately brought me back to the animated, but so before its time, Clippy. But now it is time, time to make AI truly helpful. 0:14 And I'm here at Dev Day, where OpenAI has announced more than 20 things, and underneath the long list, I think there is a fairly simple question to tackle. 0:25 Can we finally leave some work with AI and get on with our lives? 0:30 When we started this YouTube channel, we talked about Kevin Scott's, from Microsoft, his idea about capability overhang, the distance between what AI can do and what we have learned to do with it, and we keep getting more capable models. 0:45 Meanwhile, I still have two inboxes with over 4,000 emails in them. Some of you have waited days for me to reply. 0:53 I apologize, and I think from now on I will be quick, and I will be on time, or maybe it will not be me, but it will be my dot, Turi. [upbeat music] Attention Span. Welcome to Attention Span by Turing Post. 1:07 I'm Ksenia, and today let's look at Dots, the cloud around them, and why I think the proactive research is such an important thing. Let's get into some technical details for a second. 1:19 A dot runs on GPT-6 Astra, which makes it quite expensive. It also has its own cloud computer, browser, and the apps you choose to connect. OpenAI says its plugin ecosystem reaches more than 4,000 apps. 1:33 The promise is that a project can continue between conversations with enough context to pick up the thread. 1:41 They also say that they don't want to compromise on quality, so for now there is no rerouting, just top-notch Astra. In that sense, Muse can be compared to it or even Grok. The model here is simply much better. 1:56 But how can you tell without trying? So I set my own dot up. The dot has a computer of its own in the cloud. Connecting my laptop is a separate optional step. 2:09 As I said, I can open its computer to see what it's doing, and I can set boundaries around the action it takes. It actually feels pretty intimate, almost inappropriate to look what the dot does in its own computer. 2:23 Then you, I, we can give it a name, a character, and something real to work on. 2:28 I was worried that my animated pet will be removed, and instead I would need to use a bit minion looking template dots, but you can also ask your dot to create an animated avatar based on what it knows about you. 2:42 I think it's fascinating because, for example, ChatGPT knows tons about me, so it decided my dot will be an owl, a lavender owl, because that's Turing Post new palette. And the funny thing is, my father collects owls. 2:56 He has more than 500. I've never told ChatGPT about that. But when I saw my owl, I thought, "Well, apparently this is becoming a family tradition." Mine is called Turi. Guess why. 3:06 And the first thing I did is asked it to help me deal with my two inboxes, starting with cleanup. That is a reasonably clear first assignment because there is plenty of material. 3:18 The first inbox result was rather ordinary, 14 work emails archived after my approval, a small beginning. My two inboxes were still very much there with thousands of emails. I'm, I'm not joking. So I, um... 3:31 it started small, trying the waters, and then it got better and started analyzing bigger batches. We are moving slowly, so from 3,500 emails, I now have about 1,000 to deal with. 3:43 Again, not me, the dot is dealing with it, and it is doable. We went through the first 2,000 in just a car ride. 3:51 It also gathered loose ends into a private scratch pad while we were doing the cleanup, and helped me prepare for an interview, and it also did a few work follow-ups that I forgot about, and that's exactly how it feels. 4:03 It acts proactively, and then you remember something, "Oh, I need to remind, you know, John to do this, this, and that," and you just unload your brain, not the curious part of it, though, but the very mundane part into the dot or onto the dot, and the dot connects everything together. 4:21 Yesterday, I was, uh, talking with one of the OpenAI guys who is dealing with, uh, developing documentation, and he has been using Dots for a couple of weeks. 4:31 He told me his dot noticed a change elsewhere in the system that the documentation no longer reflected. It suggested edits. He hadn't asked to check that particular change. He didn't even know about it. 4:44 It connected something happening in the project to something he was responsible for. And I just keep thinking about it. So much work lives between the things we write down. 4:55 Remember to check the document after the product changes. Remember to tell the person who is waiting. Remember that last week's answer may no longer be true. Remember, remember, remember. 5:06 We carry those connections around all day. 5:08 On a large project with people working independently and in parallel, some of these connections, some of these memories simply disappear, and the systems like Dot help with this very naturally. 5:19 For my work, I have another example. My dot already analyzed my Twitter and made a few suggestions how to improve it, and it can also notice that a new paper changes a point in script I'm about to record and fix that. 5:32 But with all that communication, and also remembering Clippy, is the dot annoying in its omnipresence? Surprisingly, no, at least I haven't noticed that. It is everywhere. It's in your phone. It texts you. 5:47 You can call it. It's in your browser. It's in the desktop app. But it feels more like a supporting system than to a drowning in information me. So how does this proactivity that is, seems to be so valuable work? 6:01 There is an important boundary here. OpenAI says the background research that looks for useful things is read only. It can inspect permitted connected information. 6:11 That research cannot itself send a message, change an app, or operate a browser or computer. Any action that follows goes through the normal rules. 6:20 The system checks proposed action against your instructions, against custom rules, and against built-in safeguards. Some steps can proceed with permission you have already given. 6:31 Others need another confirmation or a handoff. My inbox is a simple example. That's why I keep coming back to it. Finding a group of old messages and archiving them are two separate steps. 6:43 I want to understand the proposed group before it disappears from the inbox. That's what the dot does. Clippy tried to recognize what we were doing and offer help. 6:51 The difficulty was partly knowing where and when to stop offering. More capable assistants, especially the voice assist- assistants that came later, inherited the same problem. 7:02 They also had access to much more of our work, and that was a bit overwhelming. 7:06 And if Turi, my owl, brings me useful information every few minutes, I could end up with a third inbox inside my head, but so far it's been really helpful. [phone ringing] Hey, back so soon? Hi. Hi, Turi. 7:23 Can you, uh, tell me again how much progress have we done with the inboxes? Checking. We've archived 4,567 emails across both inboxes. 7:36 I asked at DevDay how OpenAI plans to route dots between models. The answer was that dots currently uses Astra and only Astra, and they are working on routing. 7:49 They had not announced it as a shipping feature, but they are working on it, and I think that is a very important issue for dots. 7:55 Currently, conversation with a dot does not count as tokens, and by the way, Thibault and Sam actually highly dislike measuring everything in tokens, but they say that is how most people understand it better. 8:07 So if we use Astra for everything, it gets really expensive, and in that sense, it is almost good that they had to held back GPT 6.1 Astra over safety concerns. 8:18 Model matters for sure, but we can also ask how much more useful the intelligence we already have could become with somewhere to work and a better way to leave work with it. 8:29 Of course, cloud computers have existed for years. What changes with an agent is that I don't have to operate the remote machine throughout the task. 8:37 A research project needs somewhere to keep its documents, run an analysis, and save the results and return when new information arrives. A conversation can describe all of that. 8:47 A working environment lets the steps actually happen. The model can run on separate infrastructure. 8:53 The surrounding software supplies context, tools, and permissions, and if you watched our harness episode, you'll recognize the idea. Much of an assistant's usefulness depends on the system around the model. 9:06 Today's Codex cloud environments, Agents API with computer use, and OpenAI-powered managed agents inside AWS, all this announcements that was made today at DevDay all belong in that picture, and that helps to understand how we moving away from a model and coming closer to the systems that actually make AI work and make it helpful. 9:27 And here it is interesting to look at what OpenAI has been buying. Some of the acquisitions make more sense through the lens if we look at what has been announced and what is happening with the system providing value. 9:40 OpenAI announced the Ona deal in June and closed in August. Ona brought persistent cloud environments. This is why OpenAI is betting on cloud so much now and is able to do a lot of cloud work. 9:51 So Ona brings persistent cloud environments, including environments under enterprise control, and OpenAI explicitly connected the acquisition to Codex. 9:59 Another company they acquired, Sky, brought a Mac assistant that could understand what was on the screen and act through applications. 10:07 Rockset now supports parts of ChatGPT's data infrastructure, including data plugins, conversation search, and keeping workspace knowledge up to date. 10:16 We don't know exactly which acquired technology powers each dot's feature. 10:21 What we can see, though, is the direction of investment, places to run the work, access to applications, and information that remains useful as things change. 10:31 For a company, all this comes with practical questions about credentials, access, and who can review an agent's actions. Giving an agent a computer means taking responsibility with the computer, too. 10:43 How do other companies deal with all of it? Are they behind? Are they on the same level? For sure, it's a crowded direction, and OpenAI has company here. 10:52 GroqBot already supports several bots coordinating, sharing context, and passing work between them. My husband swears by Groq. It also describes learning a workflow from a demonstration. 11:03 OpenAI is different in that sense. It starts with one primary personal dot. Teams of personal dots remain a future direction, maybe, because Sam and Thibault, both of them are very much pro one dot. 11:17 Meta's Muse has a dedicated cloud computer and works through WhatsApp, with AI glasses on its roadmap. That gives Meta a route into everyday life through communication and devices people already use. 11:28 Anthropic is bringing co-work and chat together with documents, slides, and longer-running work inside conversations. 11:36 It is becoming quite reasonable to expect an agent to keep working after the laptop closes, but everyone who I have asked say approximately the same. OpenAI by now has by far the best model. 11:49 It also has the most connections through Codex, through ChatGPT. A lot of people who are already using ChatGPT, they have this sort of environment created for them, and with dots, it just really connecting the dots. 12:02 So we will see how this competition will be going and what Anthropic will reply with. OpenAI's particular opportunity may be in how the pieces connect. What they announced is interesting. 12:14 ChatGPT Space puts people, ChatGPT, and dots around shared knowledge. If the user experience will be good, that will be a super useful thing. Pages give them somewhere to develop the work, also very convenient. 12:27 Collaborative slides are coming later. We've seen a smaller version of this change before. Shared documents replace the ritual of emailing attachments back and forth. Remember those times? 12:38 People could finally work on the same version. An agent adds another possibility. The document can stay connected to what happens around it. The product changes, so the launch materials need revision. 12:49 New evidence arrives, so... Or a figure and its explanation need another look. We're back to that documentation story. And plugin extensions let developers build interfaces inside ChatGPT. This is super exciting. 13:01 Connected app events can start work. Shareable profiles help people discover and reuse what others have built. That last part gives me an early limited sense of a GitHub-like ecosystem. 13:13 Useful work has somewhere to travel. 13:15 My read is that OpenAI wants our projects to accumulate here, along with the context and routines that make them work, so building all this environment that is connected and is working for you. 13:27 Microsoft's history is a useful reminder of how sticky that can become. Once colleagues depend on shared applications and formats, a software choice spreads through an organization. 13:39 Moving a model may be an easier thing to do, and may be easier than moving a team, and it's certainly easier than moving a team's working habits, permissions, and context. 13:50 So convenience deserves this question: How easily could I take the work somewhere else? As of now, Dot is not accessible to everyone. 13:58 It's rolling out to eligible Pro and business premium users with an admin-enabled enterprise beta. It is actually understandable. They need some control over this new release. They still don't know how it will work out. 14:10 Plus, it's very expensive. But during Q&A day, Sam said they will be rolling out things for mass market as well soon. Now, back to capability overhang. 14:20 When we talk about capability overhang, we tend to picture all the extraordinary things AI might do, but sometimes the gap looks like talking hectically to your phone and the phone sorting out things for you without asking too many questions, but just enough. 14:35 My theory is a lavender owl, and it's three hours old, but I already replied to 15 emails that were waiting for me the last three weeks, and archived over 2,000 emails I'd dreaded to look through. 14:47 And to the argument, if you need one dot or many, like OpenClaw, Sam and Thibault are adamant about one dot for everything. 14:55 Some of the people told me it's an ongoing discussion in OpenAI, but what I also know is that Peter Steinberger helped building those dots, because he works in OpenAI, and he did his really best to make them serve you, as OpenClaw does. 15:09 Through one dot, I think of it as chief of staff to make it less chaotic. Now, I would love to end up with a couple of questions. How do you feel? 15:16 Do the systems like Dots are helping deal with capability overhang, or is it still not that? And what assistant would you prefer, one for all, or many for different tasks? I'm curious about your thoughts. 15:29 Thank you for watching Attention Span. Subscribe to the newsletter and our channel, and I'll see you soon. [outro music]