OpenAI DevDay 深度复盘：常驻代理 Dots、协同空间 Spaces 与超快推理 Astra Ultrafast

Today I spent the day in San Francisco at OpenAI's dev day and it was a fun one. Not just because Tibo let me smash the codeex reset button for you all. I was there from the keynote throughout the day and I have good news for you and bad news for you. The good news is OpenAI released a lot of stuff. The bad news is OpenAI released a lot of stuff. There was a lot jam-packed into the Devday keynote. um everything from OpenAI's answer to Grockbot and Muse to platform components that developers are really going to use felt like every two or three minutes the team was announcing something new. I was able to test some of it early. I want to share from you the highlights from the releases so you know what to pay attention to depending on what's important to you and also maybe three or four things that I've spent a little bit more time with and have come to an early opinion on so you can see what I think about all these new releases.

Okay, so the big release of the morning, the big bang thing first thing up in the keynote is dots. Dots are an always on longunning agent that can work across chat GPT across codecs in its own computer on your computer on your phone whether through text or through call even in your Slack to get work done for you. And I was able to test dots pretty early. In fact, so early it's hard for me to give you all a full review because it's very clear that the OpenAI team is building and building and building on this every minute, every hour, making it better. I think what I want to get across to you is like where do bots sit in this ecosystem of like Grockbots's very popular, Muse is getting some traction, Instinct is still out there, you know, texting people. What are dots? What do I think of them and where do I think they're going? So again, dots are your agents and I asked the team, do I get one dot or many dots? And right now we are issued a single dot. So if you're on pro business enterprise, a couple of the key plans, you should have your dot in chat GPT in the codeex app right now. So this is my dot. Her name is Ba. She's a pink blob with glasses and googly eyes. And Bae has her own computer. So, as you can see here, there's a machine that your doc can use or you can take over and use it yourself. It's loaded with, of course, a browser, but also a bunch of apps, things that are probably pretty interesting to developers. In addition to their computer, you can give your dot access to your local computer. So again, sort of this like always on in the cloud anywhere bot that can work locally or it can work in its own machine. Some interesting things that they're pushing with dots is the kind of like multi-channel experience of a dot. So you can of course use your dot in the chatbt desktop or mobile app, but you can also call your dot. So they they've actually made this not voice mode with your dot. really calling it call your dot. It rings. It picks up. I'm curious. It's like an interesting unique interaction user experience pattern. I'm not sure that I quite get all the interactions between transcribing voice, live voice, and then calling, but you know, you do you bae. And then of course, because we're all enterprises, we're all teams, and we're all on Slack, the ability to install your dot in Slack. Now, one of the things that I find most confusing about dots is just like where does it sit in the ecosystem of all the other chatbt surfaces um and all the codec surfaces. So, you know, dot can do work. They position it as a sort of like replica or delegate of you. So, it's sort of like your main thread. Um, and so you have like kind of like one master dot that is responsible for all your work, can be proactive towards you, can remind you of things, can do work on its own, but it can also kick off codeex threads, um, and I think chat threads. They're kind of like one and the same. Um, and do work on your behalf. And so, you know, I think they still have to figure out the information architecture. I want to do a full dot review, but I want to do it when it's baked and I know that it's really going to work. But it is worth paying attention to. I will say that early takes are the work itself that these dots do is good. Like I had it go shopping for me. It found the right shoes. I had it code something for me. It coded it well. It's proactive in exactly the right way, which is I was trying to figure out my kids's weekend schedule and whether their swim classes conflicted with another thing we were trying to do with them. And then this morning it said, "Hey, do you want me to look for November ones? Here's your option." So I I I say the bones of your dot, which is a very weird thing to say, are good. There are a bunch of rough edges around the user experience. I think there's a bunch of rough edges around dots position in all the primitives that you have going on with chat and codeex. I think it's very confusing to understand what's a dot thread, what's a chat thread, what's a codeex thread. But I think the surface area is moving quite fast and I think people are going to have to decide what they think about using their dot. Where do I think I'm going to use it? I'm probably going to respond to its productivity. So, dot is going to be able to pull me in if it is proactive and suggests using my chatter codeex memories useful things it can do for me. I think that's how an openi agent gets my attention is it pushes interesting things to me and I start to engage with it a lot more. I will say right now early take and again I'm not doing a full dot review. I'm still sitting and using my Grock bots for work. I just like defining microaggents for work just like I would hire someone. I like to hire an agent for a job. So, I'm not really a one one agent to rule them all person. And then I think the user experience of Muse is just a little more consumer friendly, a little bit more approachable. Um, the avatar is super cute. So, I'm going to give it a whirl, but I'm not going to give you all a full review until I really feel like I understand the differentiating features of this, and I can tell you exactly where I think it fits in your stack. So, it's like early innings. It's got all the makings of something that's good. The actual work it does because it's powered by these OpenAI models is really good, and it's pretty smart with memory. I'm just not 100% sure uh yet. So, we're going to hold off a couple weeks. I'm going to come back to a review here.

All right. So, dots big review. What else is big in? I think one that's underhyped and I said this on X is chat GBT space. Space is a place in chat GBT where you can work on documents, artifacts, slides with humans and with agents. And I think this is what companies have been waiting for. We have been waiting for an AI native document slide HTML collaboration platform with all the right sharing built in with all the right data controls built in with all the like agent native kind of aspects of this built in. I mean honestly I built this at chat PRD early on because I knew that people and agents wanted to collaborate on the same document and so the openi team has built spaces into that same concept. So if you see what spaces can do, it can create pages and create documents by which you can collaborate with both your human friends and your dot your agent friends. Slides again another surface area for businesses where a lot of collaboration happens and I know that you know chat GPT these other AI tools are really great at creating slides but not really great at editing and collaborating with slides and so this is a big win and then all the like good stuff around sharing permissions etc. One of the things I love that my dot did one moment of delight that I didn't touch on in the overview before is it made a space for me called your personal scratch pad and it just created a space where my dot and I could just collaborate via dot on things we have to do and I don't know if they read my mind. I have a very um loosey goosey approach to my to-do list. I'm never going to be like an organized table notion boy. Um I just want a stream of consciousness list of bullet points and comments I can make on things I need to do that I can check off with me and my agents. And I actually built this um it's called Claire's notebook. Let me pull it up really quickly. Uh so what's really funny is I actually built this for myself a couple weeks ago. I built this collaborative um document where me and all my agents codeex grom muse could all collaborate on our to-do list because I was working across a bunch of agents. So I did invent this invented here um this collaborative doc. But a good example of a space again is this place where you and your agent can just keep a running list of to-dos, decisions, notes. It also created documents for me. I'm prepping for a talk. Um, I have how I AI recording plan. So, what's coming up next? And again, what's nice about this is it's not just a doc. It is um you you have the ability to appmentntion your doc at mention chat GBT and get updates to the doc. And then you can just imagine this across slides and other artifacts. So, this is really like a little bit of shots fired on Google, which again is just watching this AI opportunity go by. They have all the goodies and yet OpenAI meta have all beat them to the punch. We got to figure this out. But I do think spaces are one of those things that enterprises really want. So I'll really be watching to see who actually adopts spaces. Will you see companies move more and more to these OpenAI hosted documents or not? and how kind of the push and pull between some of these source of truth um business collaboration platforms like Slack, Google Drive, Dropbox, etc. play with spaces. The other thing that is related to spaces that I think is really interesting is the updates announced to sites. Sites, in case you don't know, is sort of like OpenAI's builtin website app builder that can be deployed. um you can sign in through chatbt and then anybody can access your web app. It's sort of like bridging that gap between deploying something locally and something you can share. You know, they pulled sites into space so it has nice sharing and permissions. But the one thing that they added to sites which I think is really really really smart is you can now bundle connectors and plugins into your site. So let me just give you an example. Let's say that you want to create a dashboard, a custom dashboard pulling data from your Snowflake. You have a Snowflake plugin through your Chad GPT app and you want to just use that same data in your shared reporting. Well, you can now use that plugin to power the data in your site. And then what's really interesting about this is if you share it to somebody on your team, they can log in and they can use their connector. So if they have access to the same data, great, they see the report. But if they don't have access to the data, then those controls come in and they can't view the same thing that you can. This is truly the number one problem I hear from enterprises that we work with trying to figure out what to do with all these vibecoded internal tools. What to do with all these vibecoded platform? What to do with all these vibcoded prototypes? What to do with all these vibecoded reports? Everybody's like, what do I do with this information? And I think this sites plus plugins is OpenAI's answer to this. I am really curious again. I think there's this competition between like clawed artifacts, OpenAI sites, where is the enterprise source of truth going to be? And I think you really have to compete on the things that matter to companies which are like easy access to data, governance and controls, permissions, you know, ease of sharing and then quality of these artifacts. So, I think this is really interesting. Um, coming up next week, we actually recorded it at the event. I did a live episode with Kath, the product lead on sites where she's going to show us all the like magic in sites. But this is something that I think you should if you are an enterprise considering chatbt as your platform for collaboration that I think is quite nice.

Now I've covered two three big things which is dots agent space like collaborative workspace sites just like a nice little quality of life thing here let's talk about the things that I actually care about model models and platform that's what I care about your girl is a software engineer at heart I am excited about models I am excited about infrastructure and I'm excited about platform components so let's talk about what those releases were. It was actually really funny. One of the first things Sam Alman said is, "You know what? We constantly get feedback on, we just want faster, models, cheaper." And it's like, of course, we all want models faster and cheaper. And they did release GPT61 Soul. This is almost Astra intelligence at sole speed and really, really cheap cost. Um, so it's $2 per million input tokens, $10 per output token, which is much different than Astro, which is $10 per million input tokens, and 50 per output tokens. So, it's a lot less, and it's a good one. And I I love the Soul Models. I have loved the Soul Model since it came out. It's just a workhorse. It's sort of like my middle-of-the-line most engineering tasks um model that I work with. I do love Astra the very most. And so because I will pay for the tokens, I'm like happy to be an Astra. But if I were being more costconscious, I would definitely spend time with Soul. Soul is super awesome and I think is just going to be a welcome um addition. But I'm excited about two other things that I want to show you. You all know that it's dev week this week on how I AI and you know what? It's also OpenAI dev week and they announced a decisions API. So they announced their little Jeff competitor. This is basically Luna with constraints and so it's going to have Luna intelligence on a predefined set of answers and it can return decisions very fast. Now that sounds like a fast follow to Jev, but there's more. I was able to test a little bit of the decisions API and it has vision. It has vision. So we get the speed cost of like a decisions API a Jev style model with eyeballs with computer eyeballs and I want to show you an example of where I think this is super super useful. So I have an episode with our friend John Linquist on Jev that we recorded. It's about a 40-minute episode and we need to make thumbnails. And every time I make thumbnails, I have to scrub through the video and find when we're not making awkward faces. And I can do this with some models that are actually quite bad at it. But it takes a lot of time because we basically like clip every 2, 5, 10 seconds. We try to find places where our faces aren't awkward. And so I built this with the decision API. It took less than I don't 10 seconds. It was like very very fast. And it went through 100 frames and found all the ones where we were not looking awkward. Like look, we look so happy. It's really cute. It's really hard to find photos where two live guests do not look weird. And the decisions API did that for me. And this has actually been a real pain point for me. Something I couldn't solve with Jeff cuz Jeff doesn't have vision. Okay. Also very important with Jev I built hot dog not hot dogs. Okay. So if I select this picture of a hot dog and I it it is 100% confidence in less than a second that it is a hot dog. So again AGI is here decisions API. You all know how bullish I am on Jev. I think this is like the missing piece in a lot of our AI architecture. And so I'm really excited to see OpenAI announce this. They haven't released it yet. So, TBD on when it's fully released. I got a little bit of early access testing, but when it's released, I will definitely do a head-to-head eval decisions API versus Jev and tell you what I think. So, those are kind of like two models that we're very excited about, uh, Soul 61 decisions API,

but what's the one that I really want? I can't afford it, but I want it ultra fast. So ultra fast you know there was fast it is ultra fast now I don't know if you know this you can run your models in codecs on normal mode on fast mode which is two times as fast or on ultra fast which is eight times as fast now it cost six times as much it's expensive but GP6 Astra ultraast is available in work codecs and the API and I got to test this also a a little bit early. And this is the one that is making me feel the future. And you all think I would put Astra on like extra high reasoning and let it rip through my bug backlog, rearchitect something, build something new. No, I wanted to see what someone could do if they were using ultraast via API as part of their stack. So what I think is really interesting is if you get very high intelligence at almost real-time experience, you can build some AI things that I think you couldn't build before. and like a good podcaster, you can build games. And so I want to show you two games that I did with Astro Ultraast that I think are just really interesting. I think the latency is still like a little bit um slow for what would feel like a realtime app, but it's pretty magical and I hope this just kind of like opens up your mind to what 8x, 10x, 20x, whatever is coming soon at the highest frontier intelligence can do. This is called the other pencil. It's an AI human collaborative sketch pad. And so if I go in here and do some waves, um, then Astra's going to come behind me and draw SVGs that, uh, match what I do. And so, let's see if I put like a circle here with a smile. Oh, no. I made a little made a little chicken. It's very cute. It's like almost too fast for me. Let's see if I do a flower here. Maybe what it will do next. And I don't know if you remember this from my Astra review, but Astra is really good at drawing SVGs. And again, this is something that I just feel like you couldn't do real time before that is super cute and very fun. Oh, look, I got a little ladybug. But let's take it a step further. So, this was real time SVG drawing with Astra. I think it's super cool and something you can do before, but let's take it a little further. So, this is Little Starship. This is a 3D app that I built. Um, it has three characters. It's Pip, Momo, and Sprout. Now, this was 3D built by Astra pretty fast. I didn't do um Ultra Fast, but it was built pretty pretty quickly. So, Astra can definitely code an app like this. But here's where Astra ultraast comes into play. So, if I type in here and I want to use those other rooms and I say, "Let's build a bathroom for our friends." Then what happens is it uses Astra Ultraast to render these new 3D objects super fast real time here in the app. So it's done a really detailed job here. There's a little rubber ducky. There's bubbles in the bath. Pip has gone over to the bath. So you can actually build this sort of like realtime game. My kids were doing this yesterday. They were doing so much stuff. So, let's say let's all go on vacation to a tropical island. Um, and it will completely change the scene of this little game. So, look, there's a little a little island it made with a coconut tree and um some sand and all my buddies are coming over to hang out on my tropical island. Now, you can also change the physics of the game. So, let's say remove gravity. It's time for a float party. and it will remove gravity. And like again, it's very, very fast. It's a couple seconds to do all these changes live. These are all rendered through Astra ultra fast. And my kids were playing this for like 20 minutes last night. Now, real talk, it cost me like $97 or something to run this for 30 minutes. So, it was not cheap. Um, this is a very expensive game. a little expensive demo for you all, but I just think how cool this is to be able to do these like real time code rendered games. Um, it has a keeps a big long log. These are all the changes that we made to um to the planet, all the changes we made to the ship. And again, like this is a a cute little example, but but I do think if you can start thinking about like what real time super intelligence would look like from a user experience perspective, there are some pretty cool things you can build. So from a model perspective, really excited about Soul 61 because it's fast and cheap. Really excited about vision in the decisions API. and really really really excited. My wallet is not excited about ultra fast.

There were a lot of other things announced. I mean again there were just so much like agents platform computer use in the agents API all the codeex primitives enabled for developers. There were updates to how plugins work. Rumor has it that you can even monetize at some point your use of plugins. They announced some changes to the Chad GBT plans, including a new Pro 500 plan. There was a bunch of developer primitives released so that you can build with the same infrastructure and tools that Codeex is built on, including using the agents API with computer use. There was just so much announced today. But if I were to tell you to pay attention to a couple things, it would be give dots a world, tell me what you think, full review coming later. Um it would be if you're at a company and you're really struggling with collaboration on AI generated tools either with humans or agents to take a look at space and sites. And then the models the models are just so good and as a developer somebody who both builds with these models and builds on these models I am very excited about the decisions API with vision and very very excited about ultraast just because it makes me really imagine what the future of AI products can look like. This has been my very fast OpenAI dev day recap highlights for me of episode with Kath from the sites team coming this week and and I can't wait to hear what you've tried, what you're excited about, and what you think I should cover next. Thanks for joining. Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcast, Spotify, or your favorite podcast app. Please consider leaving us a rating and review which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.