【ライブ】今週のAIニュース総まとめ：OpenAI DevDayのDotsとSpaces、Gemini 4 Argon、Mercury Voice ほか

Hi, Grant. Hi, Grant. Hi, Grant. How many people do we have in the chat? If you're watching, shout out from where you're tuning in from. We'd love to hear where you're tuning in from. This is kind of a newer format for us where we're going to try and recap everything that's been going on this week. We want to hear from you. What do you want to know about? What do you want our takes on? What are you confused about or what do you just want to gab about if you have opinions in the chat? Let us know, and we will talk about it all. And there's a lot this week. There's models. There's new tools. There's high drama. There's Google who kind of may or may not be back. There's all of this stuff going on that is really, really interesting. So we thought, why not just come on and gab and yak about it? We'd love to know what you all think too because we're curious. Just kind of scrolling through Quitter here.

It looks like Runway has introduced Project Continuum, a new operating research application built around real-time video interfaces. Ooh, you should pull that up. I should pull that up. Let's look at that. Let's check it out. Let's see what Runway's doing. All right, we've got three in the chat at the moment. I want to hear from you all. No lurkers. Hey, hey. We need your engagement. Okay. Now let's figure out what these words meant. Okay. Project Continuum, a new operating system, research application built around real-time video. We're sharing an early look at four new ways of interacting with your computer. Portals, visual thinking, responsive video interfaces, and interactive worlds. Whoa. Interface world models like Solaris are reimagining what an interface can be. We have an interview coming out about this very thing in a few days. About Solaris. my sincere apologies

oh what's up Ryan we're all in sure we're good yeah we're going to talk about spaces and dots yeah absolutely those are on the list I saw your message last night and we made sure to put those on because it was like yeah we should talk about this first off they're a big deal so we've been exploring what if your operating system was personalized, adaptive, responsive, fluid, dynamic, generated. All the words. So, continuum. Okay, real-time video interfaces. Ooh, I like that. That's really cool. Yeah. Generate portals for any topic. Wow, weird. I don't like the motion blur on it. I could do without that. Yeah. Interesting. simulate any world. That's funny. Can you do the back rooms? Yeah. Right. Do you want to give people a nightmare feel? Did you see that movie?

I watched the movie last weekend and I have spent the rest of lore. I watched all of the King Pixels videos and I'm now sifting through hour and a half long films by other people. It's so fun. Nice. Yeah. I didn't realize how big of a community it was, but yeah. Is this replaying? Is this just replaying? Yeah, yeah. It's just on a loop. Just on a loop. You've got me. All right. Shout out in the chat, those of y'all who just joined us, where you're tuning in from and what you want to talk about. I don't know if we have any developers in the chat, but Corey and I were just talking about this whole factory versus cognition. battle that's happening. I don't want to spend too much time on this, but that's wild, y'all. If anyone's been watching this, it is capital C. I'm going to bring up Matan's... I'll bring up his LinkedIn.

Okay. Because he had a lot to say. Yeah, for sure. That's really cool. The thing that you just showed, Continuum. Yeah. I like that. I like that research direction they're pursuing and I think it's really cool. I do too. And I think coming from Solaris it's extra cool. Solaris is their interface world model which is basically like it'll generate the software interface for you for things. It'll you know it's it's it's not your standard world model in the way you think of world models but it's actually very cool. Very very cool. Okay. Where are we at? I think he may have taken it down. Oh, man. Which is probably... Probably for the best. ...the gentlemanly thing to do. Yeah. And I like that.

Yeah. I don't know if anyone's interested in it. You can let us know. But basically there was some weirdness around an advisor that was advising both companies. and there's accusations that were made. My main takeaway from all of it, because I don't know what's true and what's not, is that it just shows you how much is at stake with AI coding tools. Both Factory and Cognition make AI coding tools. They're the two third-party model agnostic tools. That are not open source, right? Because there's open code, which is also open source. Yeah. And I mean, I guess that also doesn't include your Vercel, your Replit, I mean, you go down, there's still many. But these ones are, like, well-regarded. Developers appreciate, like, they have high standard amongst developers and are probably future acquisition targets, if I had to guess. We saw that Cursor sold to XAI for, what was it,

$60-something billion. So there's, like, a $60-something billion prize at the end of what you're building here. if precedent is set. Like if you're building models and don't have a killer agent, like suppose you're, I don't know, a Microsoft? Sure, yeah. A Google? A Google. What's funny about that is Google already acquihired the original people in charge at Tognition. Oh, that's right. Yeah, and then Scott Wu, I believe, took over the company. and then it's like, the word I'm thinking is that he survived it. What do you, he revived it. There we go. He revived the company and they've been crushing it ever since. But, you know, we'll see what happens. It'd be funny if Google buys them again. But, yeah, very messy situation. It's a messy situation. Yeah, it was essentially some corporate espionage accusations.

It was interesting. It's an interesting look into kind of the underbelly of a really, really competitive industry at a hot time. Yeah. But I say let's skedaddle on over to some of the exciting new things we have to play with. Let's do it. Let's do it. Let me bring up – oh, that's Muse. Hang on. I'm currently in what I'm calling my wizard phase. Because of your beard? I'm talking to – no. No, but fair. I need a pointy hat really bad. Uh, but, um, where yesterday I was using voice mold all day on my, my work chat GPT account. Simultaneously though, on my phone, I have, I have my dot Herman up and I have Frumpkin who is my muse, muse guy.

Uh, and, and I realized I was sending all of these things off to do different kinds of jobs at different times. And I was like, oh, wow, this is this is this is big. This is big. I and and like not stuff that's easy. If you know, I just shared this this new thing I'm writing for the neuron out that is this like science and math digest. I want to start releasing. And you should actually pull that up on screen. And yeah, that's a good idea. Actually, let's do that. But in doing so, I pulled up dots and, man, sorry, I'm talking slow. I pulled up dots and I was like, hey, here's this idea. I want to do this thing. And then I called it because you can call dots. I don't know that I like the call form factor. It's a little corny. I like my hands free. I think I prefer, I think it's fine if it's an option, but I also want a voice mode button in there really bad.

Why haven't they done that? Why don't you just have like a always on? I suggested this. They didn't take me up on it. But like there should just be a perpetual voice mode that's just in the codex harness that you can talk to anywhere. And it knows all of your chats. Like, yeah, that's like the easiest form factor. They should absolutely apply that. I don't know why I haven't done that. Now. OK, so I wanted to do this. So I pull up dots and I start explaining to it that I feel like only the crazy things in AI get attention right now. We're moving into a midterm, which makes it amplified ten times too. And I thought there's not enough attention to some of the amazing things that happen. Like I follow a ton of scientists and mathematicians on Twitter, and constantly you're just barraged with this neat stuff that these people are doing that's enabled by AI. So I sicked Herman, my dot – I'll show you Herman here in a few minutes too – on this and told him, you know, here's what I want.

I want these. Go gather every scientific study, funding announcement for science, anything where AI is touching science, where it's touching mathematics, and bring them back and, like, build me a spreadsheet. And it worked for 25 minutes or so. I like it. And it occasionally came back and was like, hey, how do you feel about dates? What if the science actually occurred in June? But they're writing about it in September. where I said if it was published in the month of September, we'll make that our guideline. If it's published in September, cool. And then we'll just do that moving forward, and then we make sure everything gets scooped up and organized. So it goes through. It does that. Let's scroll through this. Let's see what it looks like. Yeah, we'll scroll through the article first. Let me make this a little bigger, too. I have too many browser windows open. I am so sorry, people. It's okay. I also like this cat ad that they're very clearly targeting the neuron.

Oh, my gosh, yeah. That's a sick shirt. Actually, those are all kind of sick shirts. Okay, so they're targeting you. They're targeting me. Yeah. Here, I'm even going to do a thing I never do, and I'm going to right-click and open that tab in a new tab because I might order a weird shirt. I mean, that wrong one is pretty cool. So here we've got scientific findings and predictions, new research tools and methods, AI company funding, research, support. What will we be watching next? I also included a nifty form here so that in the event you see something that's missing or is not present or whatever, I want to make this as accurate as we can. It's not like an official record by any stretch of the imagination, but if we've missed something, I want to know about it. I don't care what company it's from. But if it's science, if it's math, if it's engineering, if AI touched it in a way that is good for those fields, I want to hear it and get it in here to make sure it's good. You know, Ryan, we've talked about that.

You should add a good AI news section to the newsletter that highlights these positive things. It's so easy to just see PDOAM or other drawings. God, it is. And it's everywhere right now. And the truth is this stuff excites me. I enjoy the – I've been a tech nerd for many, many, many eons, and I get excited about it. And we've talked about this before, like in an AI for good or AI in the real world kind of sections that were like ways AI is being used further. And we might have. I don't know. We might do something. Yeah. Yeah. I mean, you were talking about the kind of the impetus for this is you're like, there's not enough conversation around why we actually want to do all of this. Why are we doing all this AI stuff in the first place? It's not actually so, you know, the hyperskillers can make billions of more dollars. the end goal of this research is that so we can create at least in my opinion uh universal basic goods and services and not enough people talk about this this should be the end goal of all

this automation stuff it's not recursive self-improvement just to keep making better AIs just for the sake of making better AIs it's so that we can have AI do work and science and research that we couldn't do otherwise so that we can then free up our time to do other things like pursue net new knowledge. Not let people talk about this. Each one of these includes, it's broken into sections. Whoop, sorry. Yeah, don't try and copy anything. Our website's kind of messed up right now. And yeah, please forgive the ads. The way these ads are set up, it works fine on a normal article, but when you do something like this, their placement becomes way too often, and it's not been an easy thing to sort out. Exactly, Ryan. Science, medical, math, environment, quality of life. You know, all things that we can increase if we, you know, have autonomous research and autonomous, you know, labor basically working on it. So, yeah, agreed 100%. So, like, we'll have, like – God, it's – Corey, you can't try and copy anything. You can't try and select anything.

Oh, now I understand. I didn't know that. Yeah, it's a mess, so we've got to fix it. Oh, we've got to fix that. Yeah. So we've got a different way to recognize DNA, a different way for RNA to recognize DNA. That's that – and that's not the anthropic one, but there is an anthropic one. The date it happened, what it is – god, I didn't – You can't select anything. I'm not telling you, bro. I surf like this constantly, and I only live on the back end of this website. We need to send this to our tech team ASAP. I am. Yeah, we're going to cut this out of the video, and I'm going to send it over and be like, for God's sake, please. Yeah, what is happening? That's hilarious. But it's supposed to include what it is, right there, right there, and then what it means. It's like the work gives genome engineering teams another natural system to investigate for programmable gene control, and then links to the studies and any coverage. And it'll do that for every one of these.

There are like 56 entries on this page of things that were announced. Aging clocks tested inside of a trial of an AI-designed drug. Whoa. Peer-reviewed in Nature Biotech. Oh, I think they actually talked about this. Genomic classifier, Simclatron. Now, a lot of this may be over your head, but it's not all, And I'm hoping it explains it in a way that this last paragraph in each should be the one that is the, okay, what in the world does this mean? Yeah. Hopefully that's where it'll explain it. But it goes through many different sciences, astronomy and space, mathematics and formal reasoning. See, there's Navier Stokes. What's the coolest thing that you think happened this month from this list? this is a pretty big one yeah is a big one there's another one down here

that is an anthropic related one that is oh I'm going to miss it it's the what conjecture I can't remember Grant oh here's astronomy god this is terrible we're going to get this fixed strong secretary conjecture for linear matroids whatever in the world that means alpha genome atlas what is a linear matroid this is a really cool thing from google deep mind yes the alpha genome atlas yeah makes billions of DNA predictions searchable there's just so much there's so much and I still need to sit down and go through it much more closely paper to agent is awesome we did an article on that was cool the neurons deeper coverage yeah yeah yeah that's awesome yeah i did a full story on that and that's that whole idea like they've even got a feature on hugging face pages now that's like give this to your agent so you can

read a scientific paper now and be like oh that's a cool idea and take that paper and go to your agent and say, try this. Yeah. See what you think. And I was doing this earlier this year with a couple of papers out of OpenAI. Grant, do you remember that when we were talking about that? Well, so one of them was you did it with auto research where you actually gave your agent auto research, which is a repo, so not just a paper, from Tarpathy. and you were able to implement this thing that basically lets you work towards a goal using reinforcement learning and then just like continuously researches towards that and used it to make a tiny LLM. Yeah, and you made your own LLM with it, which was really cool. And then I think also, well, just for my own use, I did this with like three or four different optimization algorithms that came out, like DFLASH and some of these other ones,

where I just gave it to, you know, Claude at the time. And I said, hey, let's implement these in this open source agent. Let's write our own kernels. And let's try to make it run as quickly as possible on my MacBook, you know, 18 or 16 gigabyte MacBook M2. And I was able to get like to the point where I could run some things that were pretty big, you know, a couple tokens per second, but just implementing the code directly from the papers. And that's such a wild thing. Like I did – I don't think it was auto research. There was another one that was like a genuine ARCSV preprint. Preprint. Excuse me. I can't talk today. I'm not on my game. And I was at a Mexican restaurant eating lunch, and I was scrolling Twitter, and I saw the paper and i was like that's a really cool idea and i i sent it via open claw to my my home computer

and said hey let's do this and like it was working when i got home like a paper that had released in the past couple of hours was able to be sent to a house implemented and then tried now that won't work with all the other thing you know some papers are nonsense we have more nonsense papers than we've ever had right now. Some of that is AI. Some of that is that the science is happening faster than peer review could ever be expected to keep up right now. Yeah, for sure. And as a result, some things you get into are a little more niche. You want to really understand what you do because you could break things. But yeah, to get back to the point, I had Dots do this, was the idea. It put together. Now I'm going to pop over here and what we're going to do let me take this one down I'm going to show you what Dots did and this will show you both a little bit of this and a little bit of the value of of spaces as well

so I come over here to Herman this is Herman Herman is my buddy let me see if we can bring him up yeah I gave him a nifty hat he's got a little bowler and you get to pick their eyes, you design them however you want it's like here's some shapes can you make it look like your pet can you select you can also yeah you can also use your pets okay or you can just use the color colored wheel if you're like i'm not down with the cute fuzzies but uh you should be because they're i don't know it's just something about it that's different yeah yep and uh so we've got we've got herman here and last night it was like let's go up here this was the long discussion ah here it is go away I was just looking at this article in Nature and it made me think perhaps we should include science funded supported oh this isn't the beginning hey Herman can you put me together a list of as many AI science discoveries in the month of September as you can find

titles, links, any models or tools that are known feel free to break them down by branches of science or whatever is efficient a one sentence plain language explanation and a sentence regarding who it could impact would be wonderful. Feels like good work for Luna. And it says I'll use Luna to research September 26 across scientific fields. So what you should know here is I was able to tell it to use Luna, and it went and used Luna, which I think is really cool. Now, the one downside of using Luna versus Google is a model policy. As Ryan said, you got a wrong link on one of these. Oh, no. Oh, no. Luna does still make hallucinations and makes mistakes. Or at least mistakes, yeah. Yeah, yeah. Yeah, but he flagged which one it was. He said MIT engineers applied deep reinforcement learning to micro aerial vehicles. So you can literally copy and paste that comment and get the right link in there. Yeah, we'll get that fixed. Actually, here. Want to do it right now? Yeah.

Yeah, I need to know. You got a controller. Okay, but which one is missing? Is that what's – is it this one has the wrong link on it? Just kidding. He said it was his wrong link. Ryan. No, no, it's not an AI mistake. It's a very human problem. Happens to all of us. But here you go. So while it starts working, it's like, okay, I'll do that. I'll distinguish actual findings from tools, predictions, and announcements. And it's like the list is shaping up across medicine, materials, climate, and mathematics. One useful distinction. Remember, by the way, this is Luna doing the work. Several September headlines describe older research newly published in a journal. So I'm marking those dates, and I was like, I think announced in September is fine. It's like, perfect, I'll use that as the cutoff and note earlier research dates were relevant. I've gathered 47 entries, it comes back.

Here are three standouts that are really cool, like AI-guided heat storage salts. Wow. Hacken's aging research hypotheses and AI-assisted discovery of antimicrobial polymers. Wow. Plastics? Polymers? Plastics? Mm-hmm. Wow. This could aid antimicrobial material and drug research, though these aren't approved antibiotics. and then it comes back and it's like here's the full one now this lives in spaces sorry polymer is a large molecule made of many smaller repeating subunits called monomers so it's just like a chain of molecules not specific to plastic but you often hear it in the plastic context in plastic discussions we are not biologists no we are not alright so what you want to know is like here's where it did it So let's go over here. It saves it. Anytime it creates a document, anytime you upload a document, it's going to come right over here, and it's going to go into your spaces.

Your spaces is basically Google Drive, missing some features, but it's also, you know, 48 hours old, and I expect it will continue to improve and grow. they were asking for ideas last night already. So you come over here and here it is. It's saved there. It's going to open. You can share it with links. You can put links on however you want and adjust links, adjust access permissions to whatever you want them to be. You can come over here. You can download it, zoom in. What I haven't figured out is how to edit it. yeah looks like you can't this doesn't do anything but here's what it put me together originally it was like here's this it's pulled the scope how to read it what it means original 47 tools and you come down and this is the list it made with everything and like it's pretty cool this way but you know so we'll scroll past here and i had it come back later on and add in the

funding things. AI companies supported research results. And these are ones where the OpenAI Foundation has donated or Anthropic has donated. Google DeepMind has donated. And a lot of times, sometimes that's in cash. Sometimes that's, you know, $100,000 in API credit. You know, there's a lot of ways that donation can happen. But it's definitely a supportive thing. but I had it come back and put these all together. Yes. Ryan just said, can you use the chat field at the bottom and ask to modify it? Yeah, you can. Yeah. Yeah, you can absolutely do that. That'll work just fine. You can even select which model you want right there on the spot. In my opinion, if you're designing anything for, you know, collaboration between humans and agents, It should be cross, you know, anyone can edit it. You should be able to do the exact same thing.

So you and the agent should have parity in terms of what you can do. So if the agent can edit it, you should also be allowed to edit it. I agree. And I think that's not how they shut it up. You can from inside chat. If it opens up in the window on the right, you can type and work within it. But I think this may just be a thing that hasn't happened yet. That feels like such an obvious decision to me. But it's very much good feedback to give to the team. Now, here is what makes it a lot like Google Drive. You have pages, which are essentially your docs. You have slides. Slides and sheets are coming soon. They'll be here soon enough. You have site, which is your chat GPT sites. You have images. You're able to create folders. Additional spaces. You can start from a template. You can upload files to your space to have there and store together. You can connect all of the plug-ins you want to plug. So, like, if you come over here and you're like, I want to make a folder for, we're going to call it Neuron Space and Math.

It would have typed that better. another thing that ai apps need to do more often is they need to give the um dictation tool anywhere where you can type anything this is something that i've noticed a lot where it's like oh wow you have the dictation you can talk into a chat but then you if you're trying to comment on something you can't use dictation yeah it's like put dictation everywhere there is a writing interface that's my cell box please continue i agree with that and i think we're getting there. I think it's getting there. I just think it's maybe not yet. But you just drag and drop just like what you expect in any other drive type tool. And you can upload your own things here to make sure you have all your documents together. Instead of having to go out across the internet via MCP to get to your documents or something, you could now theoretically just have all of your work hosted right here in ChatGPT, which is pretty crazy. And the truth is, that's where I'm doing my work. I'm, you know, I'm doing most of my work

here. It's creating the Google Doc. It's creating the sheet. I go over there to make final changes. And that's really it for the most part. You know, make sure everything's good and adjust sharing permissions. But now, theoretically, you could just do all that in your ChatGPT account for what you're already paying. That's the kicker is like, you know, you're paying money to use you know, Google Drive, Microsoft Solution, Dropbox, all of the various things that are like cloud storage. Where right here you have it. They've not said like how much storage we have, but they do also, by the way, have Google Drive just right here. So if you do have stuff in Google Drive that you want, click the button, boom, it's right there. Wow. So Spaces is kind of like a replacement for Google Drive but then it also integrates with Google Drive Yeah and it well on its way to being one If you notice we got docs and websites and images Slides and sheets are coming

And that pretty much makes it, in my opinion, a real competitor for that kind of stuff. And frankly, I pay for a workspace account for myself that's not related to work. I mean, like for work, we have a lot of workspace accounts, but But for myself, I have one for a variety of small business ventures. And in trying to – I've just kept one base account. So I had my extra storage, the Gemini subscription, all of the crap that comes with it. But the truth is the only thing it has that this doesn't is the ability to stick a domain of my choosing on an email. Wow. So what you're saying is once OpenAI adds that, Google's cooked. I'm just kidding. I'm saying that if OpenAI has the ability to let you host your email through here as well, say you're a small business or an individual who just wants to, I'm saying that that's a real problem for companies who that's their core focus in what they do.

Because here, it's a simple add-on in a tool I'm in more than I'm in Google by such a wide margin. That's fair. And, like, it's, like I said, it's not there yet. We're very much looking at this. But, I mean, it's got a project management tool built in. It's got a to-do list. It'll run your weekly updates. It's got a feedback tracker. You can come here and sort through your own items, items that people have shared with you. I mean, it's, it is absolutely the beginning of the end of the traditional cloud storage. Wow. And here's the kicker. here's why this is such a brilliant freaking move by open ai in my opinion sorry i'm getting heated go for it because this is this is and from a business perspective this is really cool it's like here's the deal right now if if claude comes out if anthropic releases a model that's like three points better an individual is like oh my god i've got to have the latest model they switch and then three weeks later open ai comes out with a model that is the best model ever and they're

like oh my god i gotta have this and they go back businesses don't operate like that nope like that's just not how business works so but they can if they so chose but once you've worked all of your filing into here it's no longer we're going to use a different ai tool it's now we're going to migrate the entire operating system of our business meaning that you have much much better like plant lock-in longevity because they're more embedded. So what's Claude's answer to this? Claude doesn't have an answer to this. Claude doesn't have an answer to this, in my opinion. Artifacts? No, artifacts would be more like a site equivalent. Yeah. They have managed agents. They have cloud environments. Yeah.

day as I think it deserves. Like Grant, you and I kind of were tipped off to the fact that there was maybe a Google Drive competitor coming earlier this year, if you'll recall. Yeah, we talked about it and various things came out that I thought perhaps were it. Like we were looking at pieces? Yeah, this is really it though. Yeah, this is definitely the thing. Yeah. Yeah, it brings a lot of questions to mind. Like when does Google pull the plug on their access to chat dbt? At a certain point, do they say, you're creating a competitor to us, so you can't put Google Drive in here and you can't use our connectors? I think that would be stupid. I don't think they should do that. But if they feel threatened, they might. We saw something similar to this with Amazon and Muse, where basically Amazon had problems with Muse, people using their Muse bots to shop on Amazon, and so they blocked Muse from accessing. I've heard a good take on this, which is basically,

well, that's just how they get business deals done. So they block access, and then they get them to come to the table, and then they sign some sort of, like, rev share agreement, kind of like how, you know, Apple and Google have had this deal for years where, you know, Google is, like, the primary. It's the default. Yeah. Search engine. Yeah, exactly. Yeah. Yeah. Yeah, it was really funny when the Gemini thing happened. My favorite meme ever was Apple pays Google $1 billion a year for access to Gemini on iPhones. Google pays Apple $20 billion a year to be the default. Yeah. Such a Gemini on iPhones. Yeah. We'll give you one of your billion back. I want to call out a couple things Ryan said here in the chat because these are really great. In fact, I don't even go to Google to check email that often. I just ask Codex, now you're dot, what's in there? So bring the email, too. The beautiful thing about this is that what my dot did last night, and actually not just my dot. My muse does this, too, now. They both send these to me. They're like, hey, in your email, you got a notification from Vercel that you're at 100% of usage,

and one of your products ain't going to be served anymore if you don't come and upgrade. The nice thing is I get hundreds of emails a day, most of them garbage. and many of them not garbage but just not a priority that these things are keeping an eye on it and can be like hey urgent you know your lawnmower guy wants paid it's always someone who wants money by the way the urgent is someone wanting money but i have real like psychological misgivings about going to my email account. It gives me stress and anxiety. And there's nothing I love more than kind of pretending it's not there once in a while. And I really am working on that. But now to be able to have an agent whose job is like just looking out for what matters. And, you know, or, you know, if I say I want pizza and you're like, hey, wait, you got a coupon. You know, I want to know that. So here's the alternate take on that.

I actually enjoy reading my email. In fact, I actually do it for entertainment. That explains so much. I'm always like, Grant is so good with email. He, like, keeps up with it and stuff. I don't – well, the problem is I have multiple inboxes. I have my personal inbox, and I have the, you know, Neuron inbox. You don't want to know how many emails I have over the last 40 years, 35 years. Oh, I bet you I have you be. Do you? I'll bet you may. I've got probably a dozen that are active, but probably four that I check. I thought you meant physical emails. No, no, no. Emails, because I have, like, gigabytes of emails that I've got. But, no, I actually enjoy reading it, so I don't – yeah, let's see. Let's see my unread count. It's high. It's in the five digits. Oh, hang on. We're shazamming something. 36,021 unread emails. I bet you I've got to beat. Hold on. It was at 100 and something like a year ago

and I finally spent a Sunday afternoon just select, delete, select all, delete, select all, delete. And Open Claw went through and spent like an afternoon and got rid of so many. But like 36,000 is since... On my phone? Probably January, February? On my phone, it just tells me 99 plus so I can't even see it. You know, I'll have to pull it up on my personal computer later. But, yeah, so my point with that is that I actually don't use agents to scan my email. And I've tried it, and I tried it with Workspace agents. And I'd like to pull this back to dots because I have a lot of questions about dots. Okay. And I haven't really used them yet. Well, let's hop right back over here. Oh, hang on. Wait, wait, wait. Before we do that. Yeah. Let me. those aren't my images I need space those are templates this is Herman

Herman is your dot Herman was my grandpa and my great grandfather's middle name and I've always thought it was such a funny little name I was wondering why you called it Herman I spent an obscene amount of time dialing in two eyeballs and a hat. I realized that looks like it should have taken 30 seconds. Sure. I appreciate the context. Yeah. But – and since we're here, Muse. Muse is Frumpkin. He is – Totally different vibes. Totally different vibes. He's like a Viking kind of guy. I said, I want D&D Dwarven Fighter. Make yourself a Dwarven Fighter. Let's go. Let's go. And he said, how about some, and it suggested, how about Copper Armor? And I'm like, Copper Armor sounds rad. Let's do it. But, okay, back over here. Yeah, so my question about workspace agents.

So is a DOT a workspace agent? How are they differentiating these two? Are workspace agents retired? Can regular people use DOTs? Yeah, these are questions I have after that day. Dots are pro-account and enterprise right now. That's like the starting level. And for the month, they're giving a lot more usage. The good thing to know about usage is just talking to it is the same as talking to chat. You are not burning any of your usage by talking, just talking to it. But when it's like, sure, I'll go set up a calendar invite for you, then it is. By the way, here's my – Purcell now reports 100% of your included paid insurance. Oh, and Walgreens wants $1,304 for a prescription. Jeez. Yeah. That's the next thing. Like I said, it's always somebody wants money. That's all it is. Yeah. And Joseph's invite arrived. There's a good note here. Okay. You asked about workspace agents, and the truth is I don't know.

The way to think of these differently, though, is you have one dot for now. You can have many workspace agents. Your workspace agents are subagents to the dot. Oh, interesting. Is the way to picture that. Okay, yeah. As they would be. Now, as a tool, workspace agents might disappear. I don't know. I'm not sure what the plan is there. but the way I would think of that is as this is, think of your dot as kind of your chief of staff type agent, like this is the agent you talk to on it like yesterday there's this paper that came out that offered, that was like we want people to help us use some of your compute to run these Python scripts and help us find the next great semiconductor. Oh, wow. And it was like, you know, here's the script you download.

Here's how you run it. So I gave the paper over here to Herman and said, let's go run it and see if we can help them. I have like six computers around the house. Not all of them are brand new and fancy. Don't don't don't please don't be impressed. A lot of trash. But I was like, we could absolutely. I think it would be cool to kind of participate in community compute for a scientific endeavor. I thought it sounded like a cool idea. And it came over here to my Mac, set it up, ran the script, sent the info back, and has it set up to where we can do more. And I was at 95% of my limit still at bedtime last night on the $100 plan. So I'm much more encouraged with 6-1 Soul and the way they've done Dots as far as what they did to plans. I'm less perturbed than what I was Monday night when Tebow said it.

Maybe it was the wrong move to announce it prior to Dev Day then. Maybe. I don't know. I mean, I get the idea of, like, let's rip off this Band-Aid and then let them be angry tonight. So for people who missed it, the night before Dove Day, so Dove Day happened on Tuesday. That's when they announced spaces. That's when they announced dots. Some of the other things that you might have read about in the newsletter. And the night before, TiVo had announced that they were cutting the $200 Pro plan in half, right, in terms of the usage that you get? So it was down from 20X to 10X. Is that accurate? Yeah. Yeah. We mapped it out, though, and in order to offer a level of growth for what the $20 plan and the $200 plan are, it would have had to have been like 160x, which is not feasible. I do dislike it. However, from a business stance, I get it. And because I've not personally had the $200 plan, it's not as a war.

You didn't feel the loss. I don't personally feel the loss. Well, I kind of did because I was ready to upgrade. But still, the truth is, though, I mean, they cut caching 90% from Astra to Sol 6.1. 90%. 70-plus percent of your tokens are cached if you're doing agentic work on rough average. That's the number that they give. So far, Sol 6.1, I've found it to be very, yeah, like Ryan says, very efficient. I've been pretty significantly impressed with how far my $100 is going right now. That's great, yeah. That's really good. But I think they explained it poorly. Yeah, I think that's fair. And to Ryan's point, Ryan says, for the past day, I've been having my DOT manage a coding project as a way of an experiment. I call it on the phone, get an update, it tells me what's needed for approvals, and continues the coding thread.

So how have you liked that process? Like, how does that feel different from working with it in codex? I'd like to know. And Ryan also asks, did they give the $100 plan extra credits to use before the end of the year? That's a good question. Maybe they did a couple banks resets, but I don't think it was like you get extra usage. oh hang on I gotta go to settings for that not here I have one reset I used one because I I mowed through one Monday night in anticipation that they would do another one there would definitely be one for Tuesday so I blasted through the one I had left and here's where I'm at I've been using Dotson 6.1 like a lot, and I'm at 93% usage left on the $100 plan. Now, with that said, like I said, I've told this like specifically.

Opt for the right model for the job, not always the biggest. If a model can do it well, that's good enough. If there's any question, bump up. But like you don't have to do a web search with Astro at $50 per million tokens, okay? Let's just be reasonable was kind of the deal. Like I don't – I want to get as much AI out of my single $100 bill as I can get, and that's worked well for me. And I think is increasingly going to become a necessary way of using it probably is like sort of, you know, recognizing like thinking of your AI models as a toolbox and less like a like and I'm always going to grab the biggest hammer. Like think of it as like, oh, well, I need a quarter inch wrench. Well, why would I grab a two inch wrench? Why would I grab and create? I guess the question is, at what point does that become, like, decided for you by the tool, the box itself, like the harness?

It definitely will. Yeah. And I – you know, and I don't know how well just telling it to sticks. But I have noticed it comes back on other tasks later and is like I'm making sure to, you know, use the right model. And, you know, like referencing that I've said, please do that. and told it to think of it as like a policy. Also, another cool thing to do with these when you get them, when you start with one of these personal agents, pick it up, tell it what it needs to know, and then tell it to ask you questions about you. Tell it to interview you. I literally, yesterday, I did an article that way where I was in the work account using voice mode, and I was like, I want to write an article. I'm going to give you a giant brain dump of information here while I'm talking through what I want to write about, what I'm doing, what information I need to gather, what I have, and how I'm thinking about it. And then when I'm done, I want you to go through those notes, and I want you to interview me, punch holes in it.

And then when I was done, I was like, now drop that in a word in a Google Doc as like clean bullet-pointed notes and use that to write an article that, in my opinion, was one that worked as well as anything I've ever done with AI. But it was like, here's what I want to do. I want to write about this. I think there's something here. We should dig into this. Maybe we do a section on that. I want a piece on this, just like I'm a writer pitching a story to an editor almost. I think this is the way people are going to be writing with AI in the future, where it's like – I think so too. Look, I'm going to share all of my ideas and all of my opinions on this. I don't want you to infer what I think or come up with something on your own. Yeah. Um, but I'm, I, you know, it's hard for me to put the words together in the right way to make this message as compelling as possible when I'm just rambling. So let's take all of my ideas and let's organize them and structure them. And, um, and then that way you, and you could even tell it, like, use my own words as much

as possible. So it's your own voice. It's writing in your own voice. You just are – you are having an active brainstorm that you don't have to do on the page. You can do it out loud. It's very interesting. It is. It is. It's just a different way. And I've always used it a little bit in this way, like talking through an idea, especially for like opinion stuff where I'm like I want to make sure I understand this right before I lock myself in to what my take is. because, like, I have before locked in on a take that I just didn't understand as well as I thought. And, you know, you've got to be willing to walk that back, but I'd rather not have to walk them back. I'd rather just make sure I understand before I open my mouth. I wish everyone did not. You would think it's just a normal approach to the world, wouldn't you? Yeah, but sometimes, you know, you're pretty – sometimes when you start with an opinion, and you're pretty locked in and steady on your ground.

And sometimes it takes an exploratory process to figure out, oh, I don't actually know what I'm talking about on this. Or, oh, there's all these other avenues that I did not consider, which is another great way to use AI to write and to help your thinking. I remember a saying once that was like, smart people have strong opinions. wise people recognize when their opinion is wrong or something to that effect. I can't remember who said that, but I always thought, you know, you. It's an awkward, weird thing to be like, oh, like nobody wants to be wrong. It sucks. Yeah. But, you know, you've got to want to be right enough to be willing to change, I think. And I don't know where I'm going with this. Well, what's your overall take on dots? Is dots unnecessary? Are they how this interaction paradigm between human and agent is going to work going forward? Do you still like having separate chats?

yeah I do still like having separate chats I think this is as far as typing goes I think this is a significant portion of the paradigm moving forward I mean I think you'll always have chats to some extent I think you'll always be able to go direct into Codex and just do it yourself but at the same time and I credit Zuckerberg for this because he was the one talking about personal super intelligence months ago And actually – Well, over a year ago. Yeah, yeah. Yeah. And it was a big thing. He said, I think everybody's going to have their own personal superintelligence. And they really zeroed in on this thing. And by the way, like one interesting – okay. First, let me tell you what I think OpenAI's moat is here that they have that Muse doesn't have, that Grokbot doesn't have.

And that is the gigantic catalog of plug-ins and tools that are available directly in ChatGPT already. Right, right. They have so much. Like I think last night they said like they've had upwards of 2,000 added since Tuesday. Like it's crazy. I mean Firehouse Subs is in there. You can order sandwiches. You can get your Halloween costume now. You can order flour. Everything is in there. There's everything now. It's ridiculous. And I think smart people would be building toward that environment right now because that's an awful lot of people. Yeah. I want to give credit to who said this, Greg Eisenberg of the Startup Ideas podcast. Yes. Yeah. Yes, I did hear that. That's what I'm thinking of, and I couldn't remember the guy's name. Yeah, Greg Eisenberg, he said this is – the biggest risk-reward opportunity is to just make a plugin right now for whatever you're doing. Just make a chattyweb plugin for it.

And there's a guy he was – that he was sharing out though that talked about the volume of what it was. Let's pull it up. I'll pull it up because I flagged it yesterday. That's a good one. And I think that is – I think he's absolutely right. I think it's a brilliant idea. It's the way you should function right now and be thinking. It's interesting to see how the dot prompts the threat. Oh, wait. Before that, there's a comment of Ryan's I glazed over here that I think we should point out because we were talking about it earlier. And that was about Anthropic, and they haven't really followed with tools the same way OpenAI and Google did. Like, OpenAI and Google have ecosystems, Anthropic has Claude. And I think that's an interesting take there. I saw that. I don't know if I agree with that, though, because Anthropic has plugins and connectors and skill files.

So I'm wondering what you mean, Ryan. What are they missing that OpenAI has that Claude does not? A personal agent and a Google Workspace competitor today. Yeah, that's true. No, they don't have either of those things. Yeah. And I think chasing the model is important, but I think OpenAI learned from Claude Code's success. Yeah. That what they need is the right everything around it as well to really be relevant. And I think that's where they've really shined. Yeah, and they were behind. Remember, they were behind in that area, and they had to do a lot of catching up, and finally they did. Yeah. Yeah, this was the look at our growth. Yeah, so this is Piatro, right? Piatro has a really great tool called Magic Path, and he said, stop what you're doing and build a plugin extension for ChatGPT. we've grown over 2,000% since this announcement.

So, I don't know. We get to the full chart here. We don't know what this is measuring. No, no, we don't. But it does look impressive. Yeah. But then Greg, I think, added some more helpful commentary here. And he said, I literally can't think of a better risk-reward bet right now than building a ChatGPT plugin and getting it in front of their 1.2 billion weekly users. As of yesterday, ChatGPT recommends plugins in the middle of conversations. Millions of questions get asked every day with no plugin to recommend. Each one is an empty slot. Facebook apps in 2007, iPhone apps in 2008, TikTok in 2019. Every big platform has a few months where distribution is free and most people miss it. I saw this firsthand when I lived in San Francisco building venture-backed companies. They are sacred times. If I were starting this week, I would, number one, find a tool or workflow people already Google for or do by hand every week. PDF to Excel, invoice generator, weekly client reports.

Build it as a ChatGPT plugin. Write the description using the exact phrases people type into ChatGPT. Okay, so basically SEO for agents, which is, that's a whole other topic. That's super interesting. Ship five of them and see which one gets picked up and double down on the one with the most traction. In 12 months, these slots will be crowded, but right now they're pretty empty. Most people will realize ChatGPT plugins for the biggest free distribution channel of 2026 sometime in 2027 he's right i mean he's not wrong i think you need to have a little bit more than a plug-in and we can talk about that but but yeah i mean yeah because you really have to have something for the plug-in to be working from you know on the flip side of that you could have said the same thing of custom gpts three years ago but custom gpts didn't really take off right like no they didn't they didn't really work the way that they were intended so this could be like and i was thinking this this is like gpt store 2.0 this marketplace um that's a good call but and maybe it'll work this time because the ecosystem

is more uh evolved and yeah and we're definitely more accustomed to using plugins and tools with AI than we used to be. Yeah. Like, custom GPTs were a great idea, in my opinion. And I don't think they were even implemented well. I just think it didn't catch on the way they wanted, and they quit paying attention to it. They kind of just let them falter, for the most part. I think that probably is because the patterns of use were changing, and usage just wasn't there. And I think also they were difficult to build and manage and maintain. They didn't give a lot of control to the third-party providers. And then the MCP connection came out, and I think that changed it. They were huge for businesses, though. Like the ability for one person to manage the back end of a tool like that without really building a tool,

like we used those heavily at Technology Advice. and I mean we've pretty much leaned out of them now. Most everything that was one is now a skill or a project or something else. But those still don't generally give that same amount of control on the back end for someone to be like this is how it works and leave it alone. I guess a plug-in does that. Well, a plugin also, for people who don't know, what it does is it combines skills and connectors. So you can, and apps, ChatGPT calls them apps, right? So they're apps that are inside of ChatGPT. And actually, maybe I can pull up, let's see if I can pull up the marketplace here. Okay. So we go to plugins. So this is them over here.

You have all of these different options. And now they let you do this thing where you can have, like, a custom look and feel to them. Yeah. So if you install them, they have, like, custom pages. Let me just log in. Yeah, it's like you can go to the Adobe tool now inside ChatGPT and do your work there, but inside ChatGPT. And also connected to your new drive. And your engine. Yeah do you have yours installed Okay open plug Yeah Yeah I have my Adobe attached I believe Okay so check this out So you should be able to see this screen now. So this is like a custom interface that Adobe created inside their plugin. So this is something new, this feature that they just rolled out. So basically what can happen is they control what this page looks like. They control how to design it. they have a lot more, you know,

You have your own chat history within it as well. Yeah, so it's sort of like Adobe is controlling the experience to a certain degree. It's almost a website or tool hosted on, well, it is kind of a site and tool hosted on chat GPT. And I think the reason this is really important for people to understand and why they should be thinking about building these is because the browser, as we know it today, is going to be subsumed by the agent interface. So this is already happening. Corey, how many times a day are you searching the web through Codex versus Google Chrome? I missed the first couple words. What was the question? How much or how often during the day on a, you know, 100%, you know, It's like the full day, right? How much percent are you searching via Google Chrome or searching via Codex? Like you're using Codex Azure browser. How much percent of your day are you using each?

I don't use Google Search almost at all anymore. I do still go to Chrome a lot, mostly because I have a variety of tabs open. but I do a lot inside the Codex browser that just exists there now. So you're going to give a percentage to it how much of your time you're on a website, you're in Codex versus Chrome? If I'm building, I would say 85% in Codex. Okay. If I'm doing my normal work, probably 50%. That's a lot higher than it was a year ago, isn't it? Yeah. Yeah, it is. Yeah, I mean, first off, I used to do 50 Google searches a day. I mean, like a big part of our job was Google searches. And I just don't really use them. So my point is I'm – for as much as we try to be early adopters here, I'm actually quite a laggard in terms of the adoption curve where it takes me a while to actually change my habits and my patterns.

I still use Google Chrome quite a lot but more and more frequently am I using the agent interface as my browser it is happening, it's a slow transition but it's happening, so if you're doing it 50% of the day today and I would say you are an early adopter and let's call it the most AGI pills people are probably using it 100% of the time, if not 50-70% of the time and they're already there, then play this About two to five years from now, most people will probably not use a browser, and they will use their agent interface to access the web. Yeah, and probably won't even use it like a browser. Like, the truth is, I mean, when I want something, I'm just like – I ask questions. I don't even necessarily go. I use a browser on my desktop for Twitter, news sites, email.

Those are my main things. But as you start doing more of those things, either using an agent to intermediate between you and your email inbox, like you mentioned earlier, or you're already – Email's on its way out the door for me. Yeah, so that's established, right? That's stable stakes. Then let's say you start doing more of your Twitter surfing in Codex in a browser window while you're waiting for your agent to run or you're having your agent help you navigate x.com. As this starts to happen more and more, you slowly find less and less reason to open Google Chrome. At the same time, Google Chrome is going to become more agentic. So at a certain point, the browser will become an agent interface. So this means that if there's a native way for you to build an experience for your audience inside of the agent window, which in this case is like the ChatGPT app, right? That's what I mean by agent window. Then that will be one of the more predominant ways that people work with you.

Yes, you'll have your own application, but if people are just going on the website to work with you, they'll more and more likely be doing that through an agent tool. Yeah, absolutely. And, you know, like with me, I tried Gemini's like agentic element, like where you use the Gemini plug-in, and it's good for that. But then I realized the ChatGPT extension is the exact same thing and saves its history in the account I use all the time and can communicate with Kodaks. And I was like, oh, well, that's a no-brainer. So, like, I mean, I just use it as a side rail in my browser all the time. And, like, if I have a question, if I wanted to gather something, save something, scrape something, it does that pretty well, too. And it's just gotten so convenient. Like, it's now everywhere I am.

It's in Slack. It's in Chrome. It's in my phone. It's a widget on my home screen that you can't see. I mean, I even have now, I have Apple's widgets. Apple has a codex usage limit widget, and I've got Frumpkin, and I've got a little, like, Ask Chat GPT thing there with a voice mode button where I can just one-click right there on my screen. And it's just easy. It's gotten very easy is something that's different this year. So Ryan brings up a good point. Ryan says it's going to take a while for normies to get there. A lot of people don't get it or aren't interested. I think that's true, and I think that's a telltale sign that it's still not good enough. Yeah, I think you're right. It's still not easy enough. It's still not obvious enough for people. And I've said this multiple times, but I think in five years, this will all just be how computers work. But there's a couple negotiations that need to happen.

So the first negotiation is as much as possible, it needs to just be built into the computer, built into the device so that you don't feel like every time you type a message in, it's going over the cloud and the big tech companies are stealing all of your data just to serve you ads, right? The other thing, though, is that you also want it for free or as close to free as possible. So maybe you do trade ads for that. And there's a lot of people who do that every day, right, to use free services. So it either has to be local and free or it has to be local or free. So these are a couple of the things that need to happen before people adopt them at a much larger rate. And then the rest is the interface. Like how do you make this easy, obvious, and delightful? It's come a long way. Yeah. I mean it's definitely better than it once was. I don't know if you've noticed yet. What do you think of the sidebar? Is this on the plug-in page? Yeah, it's everywhere in ChatGPT now

You notice the sidebar on your left? That didn't used to be there They had this little fold-out thing No, you had a way to fold it away and get rid of it But you didn't have the options right there That let you, like, there's spaces This is actually kind of nice I like it It's a better pattern, I think, probably I've also gotten in the habit of taking all of my dropouts. What is all this extra crap over here? That's the junk drawer, Grant. Yeah, I was watching CBPN. Everyone has a junk drawer. Yeah, it's fair. I was watching CBPN's dev day coverage, and I guess in OpenAI there's a lot of talk of merging and unmerging, so merging things together and then unmerging them and splitting them up. I still think there's some of this that needs to get consolidated and cleaned up like what's happening because you've got all this

and then you've got multiple places where you can access it you can access it here I guess this is nice where it's like okay you're saying you can pin these things you can also pin chats what is maps? I've never even seen maps I use pin chats a lot yeah maps is new is that the moon? That was a loading screen or something. Yeah. I dig it, though. What are you and what do you do? What a grant prompt. I love it. No, okay. It doesn't know what... Okay, it doesn't know what we're talking about here. Okay. In what interface am I using this? This is Maps. What is Maps? Checking. hey that's my johnny voice i stole your bot i'm sorry it's all right feels dirty now just a moment

if you are seeing something that's a local discovery view you can ask like coffee shops near me or things to do nearby and explore results on a map it's more about finding and comparing places than turn-by-turn navigation. I can't see your screen from here, though. Interesting. He's so chill for you right there. He'll match your energy a little bit. I've noticed before if I come in hot and I'm like, Johnny, what's up? He'll be like, what's up, Corey? But see, even this, so even if there being a maps in Chatsviti, right? Like, okay, now I don't have to go to Google, right? I can just look. The number of times this week I've been like, I guess I have to go to Google for that, should really concern Google. Well, they know. And actually, there's plenty of people who use Gemini. Yeah, there are. Yeah. So it's not like – and speaking of that, we should cover Gemini since they have big news.

We should probably transition to that. We should, should, should. But yeah, it's just interesting. It's like they're trying to do everything, which is why I think they're trying to remake the operating system at a certain level. Yeah. Well, you know, a thing to know with Google is that it's not about the tools themselves. It's all about how does the data from those tools feed its ad business, which is what's made Google the monster company that it is. Search is great, but search became a whole lot greater when they figured out that they could sell ads and have had for 20 years the largest funnel of humans coming into their system of any company on earth. And for the first time ever, they've spent the last three years struggling with that. So what happens is as people don't go in and check their mail manually, as people don't go search as much, as people don't use a browser, those are data points that are losing on people.

And those are money for the ad business because suddenly they can't serve as many impressions as what they could historically. And I think that a lot of what Google is doing right now has to be centered around how do we maintain some footing of what we have. This is why I think that everyone should have this concept called data pods. So this is something that Tim Berners-Lee, who was one of the contributors and the original creators of the Internet, has suggested where you and I actually have all of those data points in a little like in our own file. Right. Everything about us, we control it. We can edit and curate what goes in there. And then we can decide who gets access to it, you know, on a prompt by prompt basis. As needed. it that way. Yeah. And that would be so much easier because then you could just say, sure, this company, I'll give you this information about me because I only want to see ads about

this information. I don't want to see this other stuff. And what will you give me in exchange for what's in my pot? Exactly. Then you can start charging for it. Yeah. This exists. People just don't use it. Yeah. They don't know about it. They don't know about That's wild. Yeah. That's wild. So let's talk Gemini. Now, to be clear, they didn't release it. They announced it, which is very different, and got access to whatever their program is called. They have a name for their program. Everybody has a program now that's the companies who get it before they break the Internet. Did you notice that – so all of the AI companies are now doing this where they have like branded names for each of their mom. Yes, I did notice that. So Google – so OpenAI is all star-based. Google seems to be heavy metal-based because they're calling it Gemini 4 Argon. Or gas.

Yeah. Could be gas. Argon is a gas. That's what I mean. Sorry. I mean like chemistry-based. Yeah. I mean, does that mean we're going to have oxygen and helium at some point? They're probably going to make them slightly cooler than that. You'd think. Yeah. Could you imagine a model called Gemini Oxygen? Yeah, Gemini Oxygen. Oh, my God. I would be willing to bet five actual real dollars on the fact that just based solely on seeing Gemini Argon, we will see a gemini oxygen yeah people people who know the periodic table of elements you gotta give us your take on what the coolest uh versions of of all the metals would be elements sorry i keep using the term metal i mean when you mean elements yeah yeah all right so argon will launch at an introductory price of two dollars per million input and ten dollars per million output so this would put it yeah this is like sole opus level this is not a fable or an

Astra. At least that's how they're positioning it price-wise. Okay, so a lot of noise was made about this whole quantum side of things. So, Gemini 4 Argon is already powering our internal workflows. Thousands of Googles highlight model strength. Quantum algorithmic optimization. Argon is helping our quantum computing researchers optimize the space-time resources, qubits times gates. Qubits, excuse me. Sorry. of subroutines that bottleneck important applications. In one example, it beat the published baseline by 40% in a matter of minutes. Then memory efficiency, a team of Argon agents analyzed fleet-wide profiling telemetry to autonomously identify and apply memory occupancy across Google's data centers, bringing up over 300 tib of memory once rolled out. Interesting. And an estimated 500 tib to 1 pib in total savings. Wow.

I assume TIB is not terabytes, terabits, tero. It's a... Yeah, I don't know this. I don't know what the I is for. If anyone knows, you should... I feel like I should. Why don't you look it up and we'll inform everyone. We'll all learn together. And then they talk about... A tebabyte. Tebabyte. A binary agent of digital data storage equal to 2 to the 40th power bytes. Wow. One, okay, a hundred thousands, millions, billions, a little over a trillion bytes. Wow. Well, here we've got the benchmarks, and everyone says, like, okay, we'll believe it when we see it when it comes to the Japanese benchmarks. benchmarks everyone thinks they're bench maxing and cory you can uh you can take over when we talk about that um bloomberg article that came out to try and body this release um i don't i don't know

if that was the author's intention but it really took a lot of wind out of this um yeah it really did but as you can see here it's like very high it's higher all these benchmarks are higher than the equivalent models and they even compare it to opus 5.5 so not all of them so So you can see here this one's Frontier Squeeze a little bit better. Terminal Bench, Post Train Bench, those are better for Opus. Terminal Bench Science, that's better for GPT Astra. OS World 2.0. So you can see even on Cyber Security Bench. So, you know, very, very promising model once it comes out. You know what one of the interesting things about this chart was? they also showed benchmarks that didn't exist in the marketing materials for soul or for five five oh really like which ones like uh you look through some of these and and specifically like and i only

think of opus five five because it's so fresh on my brain but like there are a number of areas where it didn't beat Astra, and it kind of surprised me. And then vice versa. There are a number where it did, like, of course, Terminal Bench, Post Train Bench, Terminal Bench Science. Actually, Terminal Bench Science was Astra. But it's so crazy when you realize how many of these there are and how they can pick and choose. But, I mean, this shows Google on top of a lot of them. But a good number of them I'm less familiar with. Agent's last exam. That's a new one. Yeah, and that's computer use. How does Claude do with computer use now? Pretty good. Nice. It's an area they were behind at one point. I would have to do a comparison.

Let's put it this way. I can do pretty much anything using other tools via MCP that I can do with Astra. I can do with Claude. So I would have to head-to-head them on a very specific computer use kind of challenge in order to really confirm which one's better. But does it have the – it doesn't list both of those here, huh? Yeah. No, it lists – oh, let's see. And OS World, I guess they didn't test 5.5. Yeah. I'd be curious to know, but – Same, same. I just – they seem to release a longer slate of benchmarks than the other models have recently. So I was kind of looking through them to see, like, okay, what wasn't in the marketing? Like, do we have any in here that wasn't, like – I think at this point, all of these benchmarks are just mostly to...

I don't think they're going to convince anyone who already is a hardcore Opus or Claude stan. I don't think they're going to convince anyone who is a GPT guy. I think it's just, hey, Gemini, we got the juice too. I think it's like speaking to their own people. It's like giving their own people credibility to say, if you use Gemini, Okay, this is where Gemini is really good. Do you know what I mean by that? I'm having trouble finding the right words. Yeah, like why would – like there's nothing here that says I should absolutely jump ship right now. Right. And go switch tools. Like I think some of that is we've just gotten more ingrained in them. Like now that I'm building stuff constantly, the idea of switching would be – it just feels like it would be a pain. And like I would be going to a scenario where it doesn't automatically know all of the things I do. It's not going to have everything that this tool has on me. I mean like, yeah, you can make switches, but it would have to be extremely compelling.

Like I would have to feel like I'm not going to switch back as soon as I see another one. Yeah, as far as Google. Now, in speaking of this, there was additional drama yesterday directly around this model. You want to pull it up? Oh, were you not finished? I'm sorry. No, no, go ahead and pull it up. Yeah, you have it on your screen. I just misheard you. Google staff were apparently questioned. I mean, and this felt like right about the time the model came out, like it was fast, that Google staff are questioning its performance despite its benchmarks, saying that they don't see those like, oh, what was the phrase? Oh, and of course we can't see the whole article here. Hang on. but that they don't see that turning into real returns. Like they're not seeing the level of improvement that the benchmarks imply there is. So I think that's worth at least calling out.

Now, the caveat to this that someone – multiple people pointed out on X was that they talked to three engineers. Google has thousands of employees. Absolutely fair. There's lots of broad swath of opinions across the company. So just because three people said, I don't know, I'm not really seeing the gains, doesn't necessarily mean anything. Which is why we call out both benchmarks and vibe checks, because you kind of, you know, to what Ryan said. So Ryan in the chat says, as far as Google, I'll believe it when I use it. This is how I think everyone needs to approach these at this point. He's like, yeah, you can try it, which is why I think the benchmarks are more like marketing to the people who are already Gemini fans to say, like, OK, you can use this for even more now. So it's sort of like giving Gemini users more gas in the tank and hyping it up to them. I don't think it's going to convince anyone else to come over to the Google camp. It's just giving the Gemini people more. And I'll definitely go try it.

I mean, you know, I'll check it out whenever it releases. Um, but yeah, and, uh, I'll take this down. Well, Grant, um, it's been fun. Are we, while we're wrapping up? Do you have to go? Check, check your DMs. I need to go. If you want to stay on, you're welcome. Oh, okay. Sorry. All good. Um, yeah. Do you want to, do you want to stay on for a little while? Yeah, I'll stay on just because I want to talk about a couple more things. Ryan, you're the only one engaging in the chat, so if you want to hang out, I'll keep it going. Love it. Love it. Well, guys, I'm going to slide out. I really appreciate you, and see you later. Please like and subscribe, all of the things. Yeah. Yeah. Have a good day. Thanks so much for engaging. It was a lot of fun, and I'll see you later. All right. Yeah, he had to go.

But we are still here and rocking it. Okay, so there are a couple other things I want to talk about that happened today. We talked about Gemini 4. You know, no official launch yet. I would say Gemini 4 is something to keep your eye on the horizon for, especially if you use Google. We talked about opening eye dots and spaces. One of the things that Corey really wanted to talk about, unfortunately we didn't have time to do it, is Mercury Voice. So I'm going to go ahead and share my screen and show y'all this, which I think is pretty interesting. All right. So Mercury Voice is a real-time reasoning model for voice agents. Now, for people who don't know, Mercury is a diffusion model. And they're an interesting diffusion model because basically it's a diffusion text-based model, or at least that's what they became known for.

We talked to Stefano Armand, who is the CEO and creator of Mercury, on our podcast. And when I first used this to create a very simple coding challenge, I asked it to create a Tetris game. it did it so fast it blew my mind and I told Corey I was like this is the fastest model I've ever used in my life like this is this is just a wild experience and so now what's interesting about this is that they have a voice tool now we're going to look at this together because this was something that Corey was telling me about that I have not had a ton of time to look into so we're going to learn about it together all right where did my screen go here it is okay so Two weeks ago, alongside Mercury 2.5, which is their new model that just came out, they previewed what was called Mercury Voice. Today, it's generally available for all enterprise customers. This is what they call it. It's a diffusion LLM, so a DLLM.

For people who don't know, LLM is a large language model, so that's what ChatGPT is. You talk to ChatGPT through a language model that converts language into tokens and vice versa. It's tubes to power voice agents. So it reasons, it calls tools, and it follows long system prompts while keeping latency low enough for a natural conversation. Now, on a test set of real customer service prompts, Mercury Voice returns its first answer token in under 320 milliseconds, while also beating models like GPT-6 Luna and Gemma 431B, this is the open source Google model, by the way, on a suite of agentic and voice benchmarks. So this is kind of crazy. 320 millisecond median, right? Time to answer. Time to first answer token, excuse me, on production voice prompts. That's 5.9x faster than GBT6 Luna. And that's without any reasoning, just to be clear. Reasoning is when you ask the agent for an answer and it thinks in its head before it responds to you.

That's overgeneralization. That's not exactly what it's doing. But basically it runs a bunch of reasoning steps before it gives you the final answer. Quality outperforms models including Gemma 431B, GPT-6 Luna, GLM 5.3 Flash, and QN 3.5397B. These are all open source agents, by the way, that you would use. on a composite of agentic and conversational benchmarks. So the reason they're comparing themselves to Luna instead of Opus or Astra or Sol is because they're competing on this speed and cost efficiency matrix, right? They're trying to say, like, we are faster and more affordable, I'm assuming. In this case, yeah, what is it? 40 cents per million input tokens and $1 and 50 cents per million output tokens. and for right now it's 50% off. That's pretty crazy. But yeah, then you can come in here and you can see these benchmarks. We just gave you a big disclaimer about how benchmarks are hard to trust all the time.

But for voice application to feel natural, the agent has to respond within about 500 milliseconds of the caller finishing. That's helpful in context. Any longer and the pause feels very awkward. That's why we optimize for time to first answer token. So they call that TTFAT. How long until the model starts producing the words the caller actually hears? After all, reasoning is done. So Mercury voice is low. The lowest, right? 320 milliseconds. Wow. 750 milliseconds for P50 and P95. I didn't get what that actually was. Let's look into that. Reasoning in real time. Okay, so this is voice application builders have traditionally had to choose between an intelligent reasoning model that is slow or a non-reasoning model that can only follow basic instructions but is fast. Due to tight latency budgets, non-reasoning models like Gemma 431B and GBT 4.1 are the default for voice agents.

But Mercury Voice removes a tradeoff, and here's how it compares. So it's showing us this speed metric here. Then it's got the quality, right? Then certain benchmarks. Let's see, T3 Bench Telecom. So that would be a telecommunications benchmark there. For airline, for retail, this is how well it works in these customer service environments, I believe, or as an agent in these environments. And it's crushing its competitors on most tasks. And it is cheaper than its competitors on most tasks. Yeah Gemma 4 is open so that means you can run it on your own computer on your own server or you can use a provider who provides it over the cloud and you can use it that way A lot of times the way that people use open models is they either use them through what called Open Router So that's a hosting provider where you just use one API key, and you can pick from a selection of different models, but you don't have any control over who the end host actually is.

So if the host is like some server provider in some other country, you really have no idea who is hosting the model for you. So that's just one thing to keep in mind about Open Router. Whereas if you go directly to a cloud provider, like say you go to Nebius or you go to Fireworks, you know that Fireworks is the provider and they're hosting it for you. And so that's just one thing to keep in mind there. You can go directly to the provider and get an API key, or you can use Open Router where you don't have control over who the provider is, but it's a very streamlined experience. Different situations, you know, you might want to use different versions. But, yeah, this is very, very cool. And we just talked to Andrew Edinger, I believe that's his name, over at Hume AI yesterday. And I will share really quick that. So we just talked to him all about Voice AI's blind spot. And this was a really great conversation where he basically talks about their company benchmarks tools and benchmark tools.

So we're really curious to hear what they – maybe Mercury worked or Inception maybe worked with them on this. I don't know. I'd have to ask him that. But that's a great conversation about all of the ins and outs of working with voice agents because they're very nuanced, right? It's a nuanced situation. But yeah, this is looking pretty good. Like it says right here, at this level of pricing, for a typical voice agent profile, that works out to about 0.009 cents per minute of conversations, which is 5X cheaper than GPT 4.1, which 4.1, I believe, is their live agent. so definitely check this out if you are thinking about using voice agents in your products okay let's see what else we got on the doc oh yeah the white house super intelligence get together yeah this was interesting let's take a look at this I'm going to go ahead and share my screen again

alright so So inaugurating the era of super intelligence. Now, I don't know if we can officially, I mean, this is an official government document. I don't know if we can call what we're dealing with today actual true super intelligence, but this is the term that we're using now and that all government documents have to use. So by the authority vested in me as president by the Constitution of laws of the United States of America, It is hereby ordered, and then talks about the purpose. The terminology used by the federal government should reflect the transformative capabilities of these technologies and the little opportunities they create for the American people. Accordingly, the term superintelligence more appropriately captures the promise, potential, and rapidly advancing capabilities of these technologies. It is therefore the policy of my administration that to the maximum extent permitted by law, the executive branch shall use the term superintelligence and SI in place of artificial intelligence and AI and will not acknowledge the usage of artificial intelligence.

Interesting. So implementation to the maximum extent permitted by law, executive departments and agencies shall use superintelligence and SI in place of artificial intelligence and AI. So when you start seeing this term SI being thrown around, especially by the government, that's what they're referring to. They're just referring to AI. It doesn't necessarily mean that it's superintelligence in the way that it's defined by the industry. And the industry typically defines superintelligence as AI that is more powerful than, let's call it, any human, or more intelligent than any human. So it can do any task that a human can do. Some people also refer to this, it can do any task better than all the humans can do, right? That's another definition of superintelligence. Now, typically what they will, they'll refer to this as ASI, so they'll call it artificial superintelligence to separate it from artificial general intelligence. And, you know, if you listen to the neuron, you know what that means, but artificial general intelligence,

You can just think of it as an AI that can do any task that a human can do, like so it can generalize across any task. We're pretty much there. I would say like we're, you know, it's not perfect, but the systems in which we use AIs are kind of like AGI systems. And we're very close to something that meets that definition. So it makes sense to a certain degree to start using the term artificial superintelligence to start, you know, differentiating between where we are now and where we could go in the future. but I don't know if just calling everything artificial, everything superintelligence is useful, if you know what I mean. It kind of just makes it already a little bit more confusing because people already didn't know what the difference between AGI is or ASI. And I've been trying to explain to people that when it comes to the existential risk of this technology, there's kind of two different versions. There is the version where, you know, we have a less than super intelligent system that is embedded throughout all of our technology that can make dumb mistakes and, you know, could misinterpret a command in some form or fashion.

And, you know, that could lead to a catastrophic disaster. Now, does that catastrophic disaster destroy all of humanity? I find that hard to believe. It depends on the scale of the disaster. but certainly a less than super intelligent AI embedded into our technology could mess a lot of stuff up, right, if we rely on it too much and it makes a mistake. So that's one risk. The other risk is this concept called artificial general super intelligence, and that is very different from just a basic artificial super intelligence. You could technically call many systems today, like traditional neural nets, recurrent neural networks, deep learning machines, however you want to refer to the different types. You could refer to a lot of these specialist systems as artificial super intelligences because they are better at any human at certain classification tasks. There are a lot of models that are super intelligent by that definition. But the risk from an AGSI system comes from when a super intelligent system is general enough that it could be applied to any domain. And it's better than humans in every way, essentially.

That is really, the general part is what's really the risky part, because then whether it's a bad actor controlling it or a rogue AI system that is controlling itself, it can sort of apply its super intelligence to any sort of task and any sort of goal. And if that system is not aligned, which is what a lot of people talk about when they talk about AI safety and guardrails is alignment, and there's different types of alignment, but we've talked about that a lot, so I won't get into it. then it could be used for very nefarious purposes, some of which could be an Immatrix-style deal extincting all human life. Now, that's a really far-out risk to a certain extent, but it's not as wild as you think because of all this agent swarm stuff that's been happening lately. And as Ryan said, I can think of much better things for the government to spend time on than renaming AI. Yeah, I mean, this is not maybe what I would have made an executive order to do, but what they actually agreed on at this AI summit was interesting.

So beyond the legal executive order here, let's look at what they actually agreed on. I got to pull it up. Where did it go? Okay, yeah, right here. I'm going to pull this up okay this is the first link I found we talked about this in the newsletter yesterday just so people know okay this is like a radio summary let's just go to our coverage and we can chat about this very briefly So, there was this accord, right? And so Elon Musk, Mark Zuckerberg, Dario Amadei, who if you don't know him, he runs Anthropic, Jensen Wong, Sundar Pichai, and OpenAI's Craig Brockman, who's like the CTO

at OpenAI, signed a one-page pledge. So let's look at this pledge. So it's a morally binding, voluntary controlled pledge that's almost like a constitution. Not 100% a constitution, but it's not binding in any case, but it's morally binding. And then basically the tech company is committed to four layers of controls and audits to ensure the AI models are acting as intended, including internal evaluations, audits by an external firm, and reviews by each company's board of directors, according to a copy of the one-page accord. They also agreed to meet regularly to establish standards and best practices to improve the safety of their systems. Over time, this is important, over time it may make sense to codify these steps into laws or regulations. Regardless of whether this is required of companies, we believe that implementing these controls, and audits is critical to ensuring a safe future for everyone, and each of our companies are committed to doing this.

So Mr. Trump also said the AI leaders are also discussed forming a 10-person committee to watch over the enterprise, and the president said he plans to name an AI czar in the next three or four days. Also, last I heard, I think Jay Clayton, who was head of the Small Business Association at one point, is going to be the AI czar. I don't know if that's true, but that was the latest rumor. So let's talk about this. So I actually do think that morally binding is a good phrase for this because essentially what's happening is let's just play. Let's just rewind the tape here for a couple of weeks. Number one, the AI CEOs and their researchers have come out and said, AI is dangerous. We are in a sort of prisoner's dilemma. we need some sort of external third party to help us regulate the industry because we don't have any incentives to slow down. But we're worried that if we keep going at the current pace that everyone will become extinct or not, everyone will become extinct, but it could lead to a catastrophic disaster that potentially ends all life on earth.

Right. That's the worst case scenario of what happens if they keep racing at the current pace. They've all come out and said that or they've more or less all come out and said, like, hey, this is a risk factor and we're concerned about this. But mostly OpenAI and Anthropic have been worried about this. Now, the subtext behind this, right, is that they need to keep racing and showing progress to keep up with the hyperscalers like XAI and Meta and Google, who are, you know, can easily afford to keep doing the same level of research. But the AI companies like OpenAI and Anthropic have to continually doing funding rounds of increasingly large amounts to keep doing this research. And in order to raise that next round of money to pay for everything that they're doing, they have to show a lot of progress. So the speed, and then there's the whole China angle of it, but we can just leave that out of it for the moment, right? The concern here is really, like, can they stay in business by raising the money they need to keep up with everybody else in this system? So they've all said that they want some form of regulation,

right? Well, they all got together, and they all said, well, we're going to self-police ourselves, and we're going to follow these guardrails. So at this point, if they say, hey, we need some sort of formal regulation in order to, you know, make sure that we don't outpace our ability to understand what we're building. And number two, okay, we're going to follow these rules to make sure that we're all pacing at an appropriate rate or, you know, we're all reviewing what we're doing. Then it is kind of morally binding if they don't follow it. Because basically, if you're saying we need these rules and then everybody gets together and says, okay, we're all going to follow these rules, and then you break those rules, well, it's kind of on you all the way. Like, it's entirely on you. Like, no one is making you do it at this point. Everyone's saying we're going to follow these same rules. So I do think that is the appropriate terminology for this. It's interesting. But yeah, so we'll see how long this holds. Obviously, we have an election coming up,

and if there's any change in the current government, then there might actually be some stricter rules placed on all these companies. But it's like they're saying, hey, we're going to follow this agreement that we put together and we'll see what they do. And I think Corey's take on this, we talked about this a little bit before we got on, was he wants to give this agreement a chance to fail. I think that's not a bad stance. I think it depends on how big the failure is, right? Like we don't want catastrophic failure. but he does think that we should give the kind of like a gentleman's agreement of this accord a chance to play out and see if they can actually follow the rules and if you know internal pressures are such that they can't or they feel compelled that like hey we actually can't um you know we we you know somebody somewhere breaks the rule then yeah then it shows okay you you can't you don't have you don't have our trust anymore and we do have to regulate you and i think at a certain point there does need to be regulations over this stuff um in some form or fashion it's just a matter

of like okay are do the regulations make sense um and you know do they do they actually benefit the larger society um or do they just benefit the few players who you know are pushing for it So, I don't know. It's an interesting topic. Lots of nuance there. I would much prefer to have a conversation with someone about it than just, like, ramble. But, yeah, I did want to share that because that is a very big deal. All right, before I wrap up here, I want to just check Twitter and see what the latest is. So, all right. We talked about Project Continuum. Okay, this is interesting. You can now mod Cloud Code. Cool. You can now mod Cloud Code. You can change how it behaves. You can customize the UI, and you can swap in your own features. So you can write one with a few lines of TypeScripts or have Cloud build it for you.

Mods ship inside plugins, so you can install them with backslash plugin in the CLI or desktop app. And so here's a few examples. Let's look at these. Interesting. I guess they're just, yeah, changing the look. Let's see what else we got here. Okay, so this would be the terminal version. interesting this might be a little overly technical um okay but it seems like you could keep mods to yourself

or you can submit them to the clod directory. Okay, that's cool. Get started with clod code mods. And we've also got clod code mods in TypeScript. Full blob about it. Okay, cool. All right, well, that seems fun. So if you're into modding out your environment when you're working with clod, that is a new way to do it. Very, very fun. Let's see what else is going on here. fable 5.5 is going to be anthropics ko on open ai we'll see they never know okay let's see oh speaking of the government check this out trump puts a u.s equity stake in open ai and anthropic like its intel deal so in a time interview so u.s president donald trump rejected national leading ai companies but suggested the government could negotiate equity stakes, similar to the Intel agreement, where a 10% stake generated massive returns for taxpayers without new spending.

He emphasized doing it for the country's benefit, drawing from past deals like one in Alaska. Yeah, people have talked about this. This is not his, I don't think this is his only idea here that a lot of people have advocated for this. Interesting, very interesting. That's very cool. Okay. New open tools. There's a lot of implication for that. Do we think the government should do that? I don't know. You can weigh in. I really don't know, but it's interesting. New open tools cut data needs for training AI language models. Okay. Alexander Doria. So this. Ulmo Core. Ulmo's cool. So this is from the AI, the Allen Institute. they release a lot of open technology so I'm going to actually check that out for later let's see here a banger paper from meta context language models

oh I saw this yesterday yeah this is cool yeah let's see introducing clef our open source decision models yeah I will say Cloudflare has been crushing it they have released so much stuff lately that you should definitely be taking a look at that, especially if you self-host your own websites and work with agents to code. Cloudflare is a great resource for this. We really like using them. Oh, yeah, something we didn't talk about yet but did happen this week because this week has been wild is CloudSonic 5.5. So this is a really interesting chart. So this is from Epoch.ai, so we should probably go to the original. But I'm going to go ahead and just look at Chubby's post first because Chubby has good commentary on these. So, Claude Sonnet 5.5 has roughly matched Claude Fable 5.1. Wow. On what benchmark? Well, we'll have to look at this. Anthropix midsize model now matches its state-of-the-art

and probably the best model in the world within just a few months of releases. Wow. So, ECI. Okay, we're going to have to look at the original post to get what that's talking about. But as you can see here, here's Opus 5.5, here's Sonnet 5.5, and there's Fable 5.1. So it's better on this particular metric than both 5.1 and Astra. Wow. Okay, what's this metric? This is the Epoch Capabilities Index. Okay. Wow, I guess I've been sleeping on Sonnet. Okay, for both Opus and Sonnet 5.5, the gap between Epoch's capabilities index and the SWE ECI is minimal, less than one ECI point. In comparison, GBT6 is Astra. SWE ECI is around two points lower. Okay. Oh, this is interesting. Let's see if we can learn this together, huh?

Okay, the Epoch Capabilities Index. The EPOX Capability Index combines scores from many different AI benchmarks into a single general capability scale. So this is basically their version of the Artificial Analysis Intelligence Index, if y'all are familiar with that. Hmm. Okay, so you can see Astra. There's Opus. Wow. interpreting the ECI. An ECI value on its own is an abstract number. The chart below converts it back into concrete benchmark performance. For each benchmark, the curve shows the expected score a model with a given ECI would receive. Each curve is a logistic fit into the actual data. Every dot in the background is a real ECI score. okay so yeah so the curve is just like showing you the shape of it but the

dots are the actual scores here okay so this is showing you every model and where it lands on the ECI capabilities index and the percent score they got on the benchmark oh yeah and it breaks it down by the different types of benchmarks so So GSMAK, it's got Otis Makami, which doesn't mean anything to anybody. Obviously, there are, you know, in Frontier Maths, okay, so we know what that is. Obviously, GPT-6 Astra is ahead on that. Makes sense. Interesting, interesting. So what does ECI represent? How should I interpret ECI values? but is it just those three benchmarks or what all benchmarks do they use maybe these are sorry I forget what's going on here these might be composite benchmarks all of these together oh this is cool

we can filter by domains this is awesome I'm going to save this so people can play with this very very interesting I can go back over here So yeah, so for, I guess we should wrap on Sonnet since we're getting close to time here. Let's talk about Sonnet. Is there anything else that's like really important that just happens? Let's see. Oh yeah, Anthropoc IPO, they're targeting $2 trillion valuation by mid-November. I think they moved it to November so they can see if the Democrats are going to sweep and uphold their regulatory regime that they want. We'll see if that happens. It seems likely. Going public before Thanksgiving. That's interesting. I've yet to see any of the 2026 revenue numbers leak from the prospectus.

Maybe this put it together. Oh, Andrew Curran is a really good source. Let's see what Andrew says. According to this report, formal marketing for the IPO could begin the week of November 9th. The rumor for the last few weeks were that they were aiming to IPO before the midterms, so this is accurate. It's been pushed back slightly. Yeah, probably. I would if I were them. Let's see. Yeah, we still have yet to see S1. I think S1 is official. The next one's like their idea, I think. That's going to be really, really interesting. But speaking of them, let's wrap on Claude Sonnet. This is the model that supposedly was equal to or better than Fable 5.1, even though it's a small model. This came out on Monday. We didn't get to talk about it because there was more important things happening, but maybe not. I mean, if Sonnet is actually as good as everyone's saying, then maybe this is one of the most important things going on.

Certainly for agentic coding, it seems like Opus is still the better model for most tasks, but it's showing you that it's comparable here. So like I said, I feel like these marketing charts are more for people who already use these tools to say like, okay, when do I use Sonnet versus when do I use Opus, you know? So that seems to be when this comes more into play. But you can see here that if you use Sonnet on max, it can actually outperform Opus on this task set here, which is TerminalBench 4.0, which measures how well a model can complete complex, multi-step professional tasks within a command line interface. So this would be like you're working with it in cloud code in the terminal, and you need it to do a series of tasks for you, you would assign Sonnet 5.5 to do that. And then custom quotes from different companies who use it, which they can highlight here.

And they talk about it for knowledge work, you know, really cheap. Wow. For cash reads, that's pretty good. This is still pretty expensive. this is like Google's Google's Argon is actually the same cost of this sauna here. So it's still expensive but that's the entropic steal is that they're like we're the best and then they charge the most. Interesting. But yeah, so keep sauna on your radar. I would say if you're already using entropic models you can use Opus 5.5 as your main model for like thinking through stuff and then when you actually need to like go implement the code It seems like Sonnet 5.5 can do that perfectly well for you. I never trusted Sonnet models that much, just because they were still from hallucinations. And actually, I'm going to write it on this, because this was the number one question I had in this release. I was like, how well does this actually do with hallucinations?

Let's see. Sonnet 5.5. Okay, so hallucination is more complicated than it. Sonic 5.5 system card had the hallucination information in it, and I was wondering. It makes it sound like I did the same thing before the court was doing. Okay. So they tested it on what's called artificial analysis omniscience, which is a 41 topic closed book factuality benchmark. And Claude gets no web search or knowledge base. Each answer is graded correct, incorrect, or an abstention. The net score is cut to minus incorrect answers. And so Sonnet 5.5 is for 0.35, which beats Sonnet 5, but it trails every other Claude model in Anthropics comparison. So my takeaway here is that I would still have anything that Sonic does for you reviewed by Opus 5 and 5 to make sure it didn't make anything up.

And I think that's the rest of it. Ryan, thanks for hanging out. I enjoyed hanging out with you, too. Next time, hopefully, more folks will join in the chat. But, hey, we didn't advertise this. We just went ahead and went live. And we're going to try and do more of these. So definitely come hang out on Thursdays. we're going to try and wrap up and talk about all the things that are happening throughout the week. We're not talking about a given topic. And so I'll go ahead and stop the stream here. If you haven't yet, if you've stayed this whole time or you're just joining, make sure to like and subscribe. If you can subscribe to the channel, that's really helpful to us. It helps us keep doing these types of live streams where we can chat and hang out and talk about the news and get our unfiltered takes, whether or not you want that. yeah if you like this video it'll help more people learn about what we're doing and help us educate people about what's really going on in the AI industry we try to demystify it and learn together to try and understand this

wild and often fast accelerating AI industry where we're living so these are wild times y'all try to help make sense of it as much as we can It's our job to do that. Thanks for hanging out, and farewell for now, humans. Thank you.

Thank you.