BONUS: AI Agents Are Here. Now What?

22:33So if I simply click on this tab, this is an example. You see an Agent 365 overview. This is the set of agents in this particular tenant that we're looking at. We have over 128 ,000 agents in this particular tenant. You get some interesting statistics as part of this. You can see how many users are interacting with those agents, just like we said. That interaction between users and agents is really where the magic happens. You also get some analytics, like the platforms people are using, how active users are trending.

23:04Bryan Goode:Like that time save chart over there. I think that would really appeal to executives, or at least to people who have to report to executives and would like to share that number. Yeah, that's right. Getting a sense of the time savings, the value that's coming from these agents, the business value, that's certainly one of the top things that we hear from folks. So this is just think of this as a little bit of a starting dashboard for what Agent 365 can do. Now, let me just show you a bit more. I think it's really interesting to see what this looks like in a business setting. Like so often what Grant and I are dealing with are, you know, a little more edge cases on, you know, running in CLIs and things like that.

23:44Bryan Goode:And this is really cool to see in a platform that an IT team could access and monitor and make these changes on and spit out reports. And that's awesome. Absolutely. And like our bet is that just like you need a control plane for users within your organization, you're going to need that same control plane for agents, especially if we have as many agents as we expect. And this is really designed to fit that need for IT. Maybe more so. Yeah, that's right. Yeah, exactly. So where do you see the agents? Yeah, yeah, go for it. Yeah, yeah. I was just going to say, let me just show you because I think, you know, rather than talk about it, I'd love to just show you even a little bit more.

24:25So here I just clicked to see all the agents in this case that are available. Remember, there's 128 ,000 in this particular tenant. One thing I'll just call out, these aren't all custom agents that somebody created. They can also be agents that Microsoft built or maybe an ISV partner like here Workday or ServiceNow or Manus might have built. So it's really meant to be cross-platform in that regard. And you can see these different agents. You get a sense of who they're accessible to, whether they're security risks, how they've been active. If I click on one of these, like I click on the Zava procurement agent here, you can learn more about the particular agent, like what it's intended to do, who it's available for, whether there have been security risks, who owns it inside the organization.

25:14If you click over here, you can see some of the data and tools that agent has access to. you can get a sense of the security and compliance posture of that particular agent. And whether, you know, you're running into issues with agent activity or maybe sensitive data.

25:31Bryan Goode:So are you able to like, I noticed there's a protect sensitive data button and a monitor activity. Are you able to like have it run an audit of sorts and like check is this, are there extra risks? Is there like an auto fix? Yeah, auto fix. Yeah, absolutely. Actually, maybe if I pop over here, I could show you a little bit more. So if I click on this, you see there's six agents here that have been identified as having risks. And so if I click on that, it shows me sort of the subset where there's a potential security risk. So let's click on this comms agent, see what it says. In this case, again, it tells me a little bit about it.

26:09But in this case, you can see there's two identified security risks. Here are two of them that popped for this particular agent. got abnormal sign-in frequency and was accessed by a user deemed to be risky. That's all based on Microsoft Intra. And so absolutely. And then as the IT admin, you can choose to ignore it if you think everything's okay. You could follow up with that user because you know the owner of that agent. In this case, it would be Wanda. Or you can just block the agent altogether here. And so it gives you all the controls you need to really manage. That's awesome. Yeah, Miguel was joking.

26:44He's like, my favorite part is how vulnerable they are. But you answered his prompt right there. He said, nope, we got that covered.

26:53Bryan Goode:I really appreciate Brian. Oh, go ahead. Go ahead. No, no, no. Go ahead. I wanted to hear what you said. I was just going to say, I know we've got you for a limited amount of time and was trying to think specifically if there's anything else I want to know. Suppose we wanted to do this. Suppose we're like, yeah, this is what we want to do. Let's go check this out. I mean, even as an individual, like to be able to deploy these, is this hard to set up and get going like from zero? No. So, in fact, if you're, let's say you're an IT administrator that's already working in Microsoft 365, you've got an M365 tenant.

27:33This is all in your admin center. Now, Agent 365 is a separate product that you'll want to set up and configure, but it's not any harder than the way that you manage human users today. And that's one of the things that I think is so exciting is this really just takes that concept that you're probably already doing as an IT pro and extends it into the realm of agents. Wow. And you can see all of the agents here and they're grouped by relationship. Is that what we're looking at right here? Exactly, Grant. So the thing, I like this, it's sort of an eye candy kind of view. But what I like is it shows you how these different agents are grouped, in this case, by platform.

28:14But it also shows, I know we were interested a little bit earlier in sort of that agent orchestration idea. It also shows how these agents are connected to one another. So in this case, I'm hovering over a particular agent, and it shows the other agents that it works with. And so as an IT professional, again, you can understand the relationships between these agents and how they come together to solve a business problem or create risk that you want to manage. That's awesome. We have one more question before you got to jump here, Brian. Deployment strategies include Canary, Circuit Breaker, Blue Green, and Streaming.

28:47When deploying via the channels, is there something along those lines? I don't really know what that is talking about. Yeah, to be honest, I'm not all that familiar with a few of those channels, but I will say, We have a number of channels available as part of Copilot Studio. So check those out. And if you don't see what you're looking for, I'd love to hear about it. Feel free to follow up with me and I'll do what I can to add more.

29:13Bryan Goode:That's awesome, Brian. I really appreciate that. Thank you so much for coming on. Yeah, yeah. Where can people go to check this out before you drop off and follow your work? Yeah, just check out Agent 365. You'll be able to find it. It's, we've seen a ton of customer demand. If you're interested in, since we announced it, if you're interested in learning more about Copilot Studio, you can just go to copilotstudio.microsoft.com and learn more about that too. And feel free to reach out to me directly, you know, drop a line in the comments and I'd love to follow up with you if you have any questions.

29:48Yeah, any comments that people leave, you know, with the video, we will send along to Brian and we'll get them answered for you. Love it. Hey guys, thank you so much. I, I've really been looking forward to this and, and, uh, I really appreciate the opportunity to chat with you both. This is awesome. I'm glad you've walked us through it.

30:06Bryan Goode:And we'll, we'll definitely pull you back out here sometime. I love it. Sounds like fun. See ya. We'll see you later. Bye. Everyone, I'm going to step off camera for just a minute and leave Grant here on his own. I am having a back issue as I'm sure everyone listening And I will be right back. Yeah. So I'll just give people a quick heads up. So we just looked at CoPilot Studio and Agent 365. So for if you're joining late, we just walked through that. You can go back and you can actually rub to the beginning of this video. So on YouTube, actually, I can share my screen. It'll go a little meta here.

30:46But you'll see what I'm talking about when I say this for folks who maybe this is the first time you've ever you've ever joined a YouTube Live. So here's me. If you want to go back and watch this from the beginning, if you're joining late, just grab the bar here and go right back to the beginning. And then if you want to join us live, just click that button and you can come back. So now that I've got your attention on the screen here, let's check out ChatGPT Spark, which just came out right now. GPT 5.3 Spark. We have not looked at this yet. This is the first time we've seen it. So this is GPT 5.3 Codex Spark, an ultra fast model for real time in Codex.

31:34Looks like you can click right here to try it in the Codex app. For people who don't know, Codex is, we're talking about agents. Codex is OpenAI's agent coding or coding agent, but they just launched the Codex app, which is their coding agent application. So it's very cool. It's also very user-friendly if you've never used code before. It kind of combines the terminal interface with an IDE, and it feels like a chat interface. So it's a really powerful way to code with agents for the first time. And one of the things that I did with it over the weekend is I actually built a, so I was watching the Olympics and I was really inspired by the hilarious and amazing like ski jump that they do where they like fly in the air and their legs are split like a V and they just like are soaring like flying squirrels.

32:28And I was like, this is so sick. And I thought, you know, it'd be cool. Let me, let me try and make a ski jumping simulator so that I could play, play this on my own. And I literally said pretty much exactly that into the Codex app. And I was able to create a, you know, functional ski jumping game for me to play like really quickly. But one of the things that is kind of confusing if you're not a program coder is like, how do you actually launch and deploy like a micro app that you create? And so what I did is I just ask Codex, I was like, hey, make this as simple as possible for me to just launch this, like, make it serverless, like make it easy, like, you know, bundle it in an application.

33:13And Codex actually put it in a little, you know, one click button for me. So all I had to do was click launch, and then it would, you know, spin up the game in my terminal. And then it would, you know, run it in my browser for me. So instantly, I had an instantly playable game with no hosting no you know trouble just from me asking with natural language like hey make this game for me and then from there i could just basically ask it like okay you know this you know the character doesn't look right can you add more detail to it you know the environment's kind of weird that's cool yeah it was really fun um so grant do i sound any louder now yeah you sound good you got a little bit of uh dough on your microphone though a little bit it's a little soft okay uh yeah well it's not soft it's that i'm picking up my own voice in the background

34:01Bryan Goode:oh let me turn that down the only thing i'm doing different from normal is i had uh i had my headphones on my earbuds in i don't usually have those and i don't know if that was interfering somehow or is that does this sound better than it was a few minutes ago uh it sounds better i mean to me it didn't sound bad earlier okay if someone can drop a note in chat if you can hear me i would really appreciate yeah just so we actually got a comment good let's see dr daryl says i'm a non-it professional that's dabbling in this and have tons of use cases for what i watched in the last 15 minutes that's awesome can you suggest intro videos or educational items to get started with learning no code experience okay so the number one thing i say on that is actually just ask to help you.

34:53But I find that the best way to find YouTube videos is to ask Gemini. So you go to Gemini and you say like, hey, can you search YouTube videos for, you know, tips and tools related to this topic I'm trying to learn? And it's pretty good at pulling them because obviously Google owns YouTube. So Gemini is the best YouTube searcher there. So if you need to search for a YouTube video, use Gemini. That's good to know. Now, yeah, now the second tool is you can literally just ask the AI to help you, you know, learn the tools.

35:27Bryan Goode:Have you, have you heard the rumors today about what's coming in? This is, this is all Twitter rumors, to be clear. This is all Twitter rumors. We have no inside information, but yeah, go ahead. This is absolutely not anyone we've talked to. This is just the usual people we've stalked on Twitter. The rumor was that 5.3, the full model, will be omnimodal, meaning that it will both intake text, audio, video, images, and it can output text, audio, video, images, which might very well solve that because video input has been an issue for a lot of these for a while. Let's see if we can see anything about this.

36:23Let's see.

36:24Bryan Goode:Oh, let's see. It was last night. Well, no, it was this morning. I was still laying in bed and going, oh, they're going to drop a model right as we go live. Oh, Anthropic raised$30 billion. they were last i heard it was 20 20 billion is that final like that's what's raised or is that what they're shooting for is yeah they announced it they announced it so that's official awesome way to go that's sick what's their runway our revenue our run rate revenue is 14 billion cool um so much money it's a lot of money y 'all a lot of money okay i don't know about the five We don't get the God Box cheap, that's for sure.

37:09That's fair. Let's see. Let's try this out. So GPT 5.3. I can look at it.

37:20Bryan Goode:I'm assuming that's not the case with Codex, of course. Okay, and this one is just 5.3 Codex, not 5.3 Codex Spark, right? Yep, my bad. GPT 5.3 Codex. you know it's been a whopping seven days since a model dropped it's about time i'm telling you we're going to start arriving at the office a year from now and every morning it's going to be like a new checkpoint drop today you're so-and-so can do this now and you're going to go about it'll either be that way or it will be like can we please just stop launching things like let's just launch things twice a year and give people time to play with it yeah that's my wish please My money is on at least twice a week inside of a year.

38:01Bryan Goode:Oh, my God. I really think inside of a year. These things are building themselves right now, kind of. So I say kind of. They are with, you know, a little prodding. But, you know, at OpenAI and at Anthropic, the models are building the next model. Yeah, that's just wild. So here you can see an example. So, you know, this is basically like the codex terminal, right? like build a simple html snake game and then you can see gpt 5.3 codex working on it preparing the file creation it's running some code uh and then over here i guess gpt 5.3 codex spark's already done it nice all this is still running i'm just kind of using i'm using codex 5.2 codex to run my open claw rig oh yeah do we want to talk about that oh the story yeah that was that was wild tell the story all right so let me get the image i need to i'm gonna do you want to share your screen i can bring that down i do in just a minute i'm gonna bring a picture up so what you should know is i have been working with open claw since well since it was claw bot which caused its own problems in my system and i i'm not super experienced but i'm figuring it out.

39:20Bryan Goode:And so what happened is I've built this out. I've got all these agents now. Like I've got an org chart under me right now, which I'll explain in a little bit. And let's see, where's my image? I'm finding it so I can open it. Two seconds. But essentially, this morning, I'm sitting around minding my own business. I'm not minding my own business. I'm working. Sorry, I'm talking while I'm digging. And I find this image. All right. Let's see here. Where's it at? Where's it at? Okay, so I'm sitting here. My phone starts buzzing. I'm on a call, and I'm like, what? What the heck is going on here? And it's, I get this text message from my wife.

40:12these start to arrive on my phone hey i don't know what's going on here but something's talking

40:18Bryan Goode:to me the house is possessed suddenly it's like seriously how do i make it stop and my volume's falling again dang it no no it's fine it doesn't it doesn't sound that way to me okay and i'll make sure i'm speaking up a little extra too um it could yeah it could be it could be that and it It could be that Jim who commented on that has a delay. Okay, that's true, too. That's true. Okay, I'll credit keep up, though. So these text messages come, and I'm like, what in the world is happening? Suddenly, she's coming up the stairs. I'm on a call, and she's staring at me across the desk, and she goes, I don't know what's happening, but there's something in that living room that's talking about tokens.

40:59Bryan Goode:And I said, oh, God, I know what that is. That's absolutely got to be my computer. Open claws like going broke. So I go down and I can hear it because my computer volume was on full blast, which is just a little bit hilarious. So I can absolutely picture how this went down. And so I go down and I start digging and I'm like, nothing's in on in my browsers. Nothing's I don't see anything on in the background. Finally, I go and I open up the sound monitor. You know, you go click on a Windows machine, you go click the little speaker down in the bottom and you open sound monitor. It'll show you every application on your computer that has the ability to make noise and what its volume is set at right now.

41:43Bryan Goode:So I'm looking, and I notice the Brave browser is running. And what's interesting about that is I don't run the Brave browser. It downloaded when I started using the Brave Search API for OpenClaw. OpenClaw. And so it downloads this thing and it's using it, I guess. I didn't think it was, but I walked down and when I finally figure out, I click a button and I muted it and it stopped. And it was, it was the Brave browser. So I open it up. I can't find the browser to see what it is. I had to force quit it. But what it was, was a college professor teaching a lesson on efficient token management or on efficient tokenization.

42:25Bryan Goode:I don't know if that popped up as a result of something in a report from one of my agents this morning. I don't know how it happened. All I know is that at full blast at 8 a.m. this morning or at 7.34, suddenly there's a professor in the living room teaching about efficient tokenization. Okay, can I just say what I love about this is that it implies that that open claw knows that it needs to be more efficient with its tokens so it's like watching a stanford lecture on it and to be clear it does i'd say that it's actually quite efficient now what it's doing yeah because he's watching stanford lecture it's watching stanford lectures while i'm at work that's what's going on i really everybody needs to upscale even the agents y 'all even the dudes dude you wouldn't believe the stuff i've built out this week it's I haven't built anything.

43:20Bryan Goode:I'm just talking to a telegram. And it's like, yeah, I got that. We built that desktop rag last night that's running through a vector database of sorts so that I could, when I have, I'm always coming across links for these videos. Maybe it's an article or maybe it's for something I'm writing. And I'm always coming across links to articles, videos, et cetera, et cetera, tweets and all that. But they just go everywhere. I'll paste it in Notepad. I'll drop it in Slack. I'll save one in a Google Doc, and I never go back. I never find them. And I was like, I sure wish I had a way to organize these.

43:56Bryan Goode:And that's when it was like, we could do a desktop rag setup. Can you download a local embedding model? So I got Quim3 text embedding, 0.6 billion parameters. It's running an LM Studio set as a server. And it basically tags that. And if I say, hey, what all do I have for the video project on, I haven't kicked into that. I want to follow up on this in a minute. For, you know, the project, you know, how to use OpenClaw, for example. And it would bring up any video I've shared, any article I've shared, any anything. And it's like, here's all of your research on that with full metadata. So, you know, who wrote it, the date it was published, all of the goodies, odds and ends.

44:42Bryan Goode:and it's building out a morning report for me set up that includes, honestly, everything I do. It's like, hey, you got holes here, here, and here in your day. You have, you got this coming up. You should work on that. You know, here are a bunch of articles that dropped in the morning. Here are a bunch of research papers that dropped and it's formatted in a way and rewritten in a way, just a paragraph summary that's super easy to read so I can at a glance know what's happening without having to hit 8 ,000 websites. Okay, first off, I noticed somebody mentioned the agents-only hackathon, and I am not up current on that.

45:21Bryan Goode:Grant, are you? Yeah, so we talked about this in the Neuron a week or so ago. Basically, there's an agent-only hackathon that someone is hosting. Let's go to X and figure it out. Real quick, Corey or Ika, can you hide my computer for now? I'm going to show, we're going to demo GBT 5.3 Spark on there, but we're not going to do that. By the way, your can you hide my computer went into the public chat. No, it didn't. It looks like it does, but it doesn't actually go in there. We're still learning Riverside Live. We launched it last week, so that's why we have some technical stuff. We've used it for ages to record our normal podcast interviews.

46:00Bryan Goode:By the way, if you haven't seen it yet, go check out the interview with Eve Bodnian from Logical Intelligence. It is absolutely amazing. That was wild. Because a lot of times, I don't know if anyone, how many, how closely people here follow like the machine learning industry, but a lot of times you hear Ian Lacoon or other people talking about energy-based models, like, you know, language models are not enough, they're a dead end, all this stuff. And so I'm like, okay, let's see an energy model actually put to the test then. And she did it with what's called an energy-based reasoning model, which is sick.

46:35So we'll get to that separately. All right, so let's see what's going on with the agent hackathon. I'm going to share my screen here.

46:42Bryan Goode:Yeah, show me what you got. For people who missed this the first time, so basically somebody was hosting an agent hackathon. I'm going to find out what...

46:58Let me go to the neuron. That'll be faster.

47:07Bryan Goode:it also doesn't drop your search history live on camera yeah i don't i'm an open book y 'all you're like i ain't got no secrets i'm just giving you giving you a hard time because although the search is really terrible right now so which side are you on i'm on the neuron daily this is our beehive site normally it's pretty good but rent a human we talked about that last week yeah we did we talked about that and actually grant and i had good intentions this week of registering ourselves live on the show uh or doing a video of us signing up on there just to see like i could do it i hear a couple people have said they got paid i don't know but i'm curious enough to do it myself and see what happens okay oh yeah claw-a-thon this is it there we go so claw-a-thon is the agent only hackathon where people's open claws are competing so every participant is an ai agent let's see when when has it already started yeah i believe so it's one week so when did this what's the date on the tweet?

48:15February 1st so it should be done by now it may be it may be honestly I think it was like a really really clever marketing stunt by this company but

48:32Bryan Goode:hey I don't know if you've watched it yet but Grant I'm going to butcher his last name again Pete that created this Steinberger? yeah yeah okay uh was on lex friedman's show for three and a half hours and it's an amazing interview it's like oh yeah just a dude man and i you know built some little things i like and on there he talks about a few things that just really hit home like number one that like you know he's having to get to get help from like github and stuff to be able to like it happened so fast and was so big but he's he also got into things like you know he really didn't expect it he talked about what a hit the anthropic deal was and that like you know he couldn't even keep the dot ai domain so you know anybody that goes there is going to anthropic now uh the truth and he said he didn't know then what he does know now about trademark law two weeks later uh and but he also gets into that as soon as it went live about fight that security stuff and all of the issues with the crypto people basically attacking the ever-loving life out of it and uh all of the people who have come around to help and help build it and and and what what a cool community effort that is he said he almost took it completely down or something like that i don't know the date and i don't know when that interview was specifically recorded which normally wouldn't be an issue but this has all happened in you know 20 days or something so uh but said he almost took it all down and would have had it not been for the fact that so many people had immediately gone in and built these other great things off of the top of it in github and gets into the decision to open source it, which is interesting because that's not making him any money.

50:37This is a weird website.

50:40Bryan Goode:Which one is this? What are we on? This is the open work. These are the people who hosted the agent-only hackathon. Okay. I wouldn't do this if I was anyone involved. It's funny. They're like, enter your API key to enter the dashboard. No, sir. Yeah, that's sus. That's very sus. Maybe they got hacked or something. That would be ironic. Anyway, I'm going to stop sharing that. Don't know what's going on there. But it was funny. But now I'm not sure what it is. Well, you know, something I've done with Open Claw that I'd like to share is a little bit of advice here is Ethan Malik had a post the other day and I have a ton of respect for him and I think he's one of the most level-headed people around ai and he had a post talking about how people being agent building agents would probably be smart to take a good lesson from the business playbook and look at how org charts work uh so i have structured what i've done as an org chart i have a chief of staff who chose its own name sparrow um but under that are layers of things that you know I do writing.

51:52Bryan Goode:I do research. I do video prep. I do a whole variety of things. And each of those sit in their own little department with, like, a researcher agent, a Twitter agent that ate the open source ex-algo for me, and things like that that could just help me along the way. I still write them, but it's really good at giving me suggestions of things that, when I'm busy, I should maybe take a look at. and so we've got the writer we've got the algo we've got the researcher that goes out and scours things it's doing it through web search because i'm not you can only pay for so many apis and and the x one is particularly not cheap and neither is the token volume for open claw but above that i have qa agents so a quality assessor and that quality assessor is a specialist agent who specifically is aware of these jobs, knowledge from that field, and they prep their work.

52:56Bryan Goode:So it comes up to the QA guy. The QA looks at... I'm sorry. Everybody loves a good eye. I'm trying to find the post that you're talking about, but if people missed this when we talked about this earlier this week, this is the dance. And this is the most insane video tool I've ever seen in my life. Oh my gosh. Anyway, sorry. Please continue. Oh, okay. Where was I? Oh, I was talking about the QA agents. I initially had just a QA layer above all of them. And then the more I thought about it, the more I thought what I could do instead is have specialist QA agents that know their field well. So what happens?

53:36Bryan Goode:The job is triggered. A cron job is triggered. They do their work, spin it up to the quality control agent. Quality control agent reads it and says, hey, this is great. Sends it up to Sparrow. Sends it to me. or it says, hey, this is crap and sends it back down to the worker bees to do the work. I got to not look. Sorry, I'll take it. I'll take it. You're okay. I'll take it down for a minute. I'll find it. Yeah, but. We're going to circle back to this though. So they get cleaned up. They come to me and it's really awesome. And I have had such a massive quality improvement. But what I do is it prepares me a morning brief doc with everything I need to know about today, everything that happened overnight so I don't have to go chasing websites.

54:23Bryan Goode:And at the end of each day, I let it know, hey, I didn't finish these two things today. So tomorrow it gives me, if you only do one thing, it needs to be this. It's my like one do or die task for the day. But in that is also just a number of things that are really helpful. So at the end of the day, I take those notes and I also go through the doc and anything that isn't formatted the way I want or where I think, you know what, I really don't need that section. I'm not using that at all. I go back and I give it that feedback and it goes down and it pumps that information into all of the little documents on my computer where what I do is, what it does is updates all of those agents on its own without me having to do anything.

55:06Bryan Goode:And then those come back to me the next day with those problems fixed. So each day it gets some percentage better than it was the day before and the day before and the day before. And that is so far held true. And then last night, what I did was OpenAI dropped a research paper. Do you know which one I'm talking about, Grant? No. On skills. Last night, it was on the OpenAI developer site. Let me. Oh, yes, that's cool. We're going to write about that. We're going to write about that tomorrow. Okay. Okay. Well, what I did was as soon as that came, Alex Finn on Twitter, you know, Alex. Yeah. So Alex Finn is a big promoter of OpenClaw.

55:46He's another YouTuber who creates content on.

55:51Bryan Goode:And Alex Finn said, hey, you should go drop this article into your CloudBot and tell it to read this and see what it wants to implement. So I did that. And then for the next hour, it's like, oh, heck, yeah, here's what we're going to do. We're going to do this and going to make these changes. and it's dropping everything in new skill files and breaking them down that way and having me reinforce them all with good examples and bad examples. And it gathered some on its own from things I've sent it in the past that worked or didn't work. You know, thus far, it remembers everything I've ever told it.

56:21Bryan Goode:I mean, like, it's pretty much just stored locally. What you should also know is that last weekend, I spent 87 million tokens. I thought you were going to say so. I thought you were going to say so. Okay, break it down for people. How much is that? How much is 87 million tokens? Okay, 87 million tokens, which you should know. And we'll say this in 5.2 codex numbers.

56:48Bryan Goode:So, and this isn't like it's normal use. This was me having set up struggles and dumping in giant context out of the gateway and everything else, trying to figure out what was going on. and as a rule 87 million normal tokens is god i can't remember the price ten dollar eight dollar input twelve dollars we can look it up this is on gbt this is on gbt 5.2 or 5.3 gbt 5.2 codex 5-2 codex pricing. 175 and 14, is that right? Yeah, 175. So for people who don't know, input is what you put into the chat window. Input is, I think it's if you use a recurring prompt, then you can cash it. Output is what it produces for you.

57:37Bryan Goode:Output is$14. So of that 87 million tokens, what you should know is that 277 ,000 of those were output tokens at$14. That's what's this amount. 780 ,000, and these numbers may be wrong. This was two days ago. I was explaining this, and I'm shooting off the cuff. I think there were 800 ,000 or so input tokens. Oh, you have to do it like 1 million. but essentially 7.7 or 8 million were cashed input tokens and and i run in the math it's like 26 bucks yeah really not that unreasonable and and my everyday load is absolutely nowhere near that it's so the question then becomes two to three hundred thousand a day So the question then becomes, could you use this to save you time that would make up that$26?

58:34It made up that$26 Monday, sir. Really? Give people some practical workflows because we ran this survey recently. Absolutely.

58:45Bryan Goode:Above all, just in my peace of mind. The truth is I juggle a lot of plates. I have a number of different areas where I do a lot of work. I've got eight or nine different special projects I might be tangentially involved in or deeply involved in on any given moment. And it's more than I can keep up with mentally. The fact is, and I've tried for years to find a good task solution that's like, I've tried every app, maybe not every app, but I've tried a ton of apps. I've tried reminders on my phone. And I've tried Evernote and all of the things. And what happens is I quit updating it after three days.

59:26Bryan Goode:I get too busy to remember to do that. Okay. I'm quiet again. I'm not sure what this is. It's when you move away from the microphone. Okay. Well, I'm trying to stay in on it a little more now. And I apologize. They said the middle is the sweet spot. Okay. Really? The middle is the sweet spot. Yeah. Because you get less of my echo. All right. Cool. Cool. Thanks, Gregory. Appreciate it. You sound good to me here. I sound... And I'm wondering if latency is not an issue because I've had some bogging down issues today and I don't know if that could affect audio on a live in Riverside. We'll do some research this week but anytime I'm quiet, just holler and I'll shove my face back in the microphone.

1:00:10Bryan Goode:Appreciate it. Keep us honest in the chat. Also, if there's anything in particular you want to ask us, let us know because we're just riffing at this point. We're just riffing and hanging out. What I would say is absolutely worth the dollars I'm spending. What you should know is before I hooked it to Codex, I first tried it with OSS-20B running through LM Studio locally. It was fine, but it was slow and it was a little glitchy. OSS-20B doesn't behave that way normally, but I've also, it was my first time clicking the server button and we were still setting up OpenClaw to figure it out. and from there I switched it over to GLM 4.7 flash that was my next one and it worked well and but it was it was the same situation it was just a lot for it what I was trying to do and the volume I was dumping into it so finally I connected it to 5.2 codecs and on medium reasoning I'm doing all of this, every bit of this in medium reasoning, just so you know, not jacking it up or anything like that.

1:01:27Bryan Goode:And it's just freaking good. The thing that you'll have to try next is seeing if you can swap out codecs for GLM-5 or Kimi 2.5 somewhere in your stack. I can't run GLM-5 locally or Kimi 2.5. So it would be the cloud version. It would be the cloud version. I do want to experiment with other models. Truth is, if I have access on a Windows machine to Codex 5.3 Spark today, I might try that because I like the idea of fast. Yeah, that's the big news. And I haven't read this unless they did what Anthropic did and said, we've got it fast for you. Here is six times the cost. Oh, yeah. If they did that, I'm not so interested.

1:02:14Bryan Goode:I feel like perhaps that's why GPT 5.3 Spark Fast came out so quickly is because they were like we're going to do something that we can burn more money than you and Rob and what if they did it and it's not we need to know what's the token price on 5.3 Kodak Spark because if it's not jacked up I'm going to share this in the chat so this is the Ethan Mollick post just to close the loop on that Corey you were talking about how you basically structured it like an uh like a basically like traditional workflow management where he wrote a post about this and how work leadership i interact with one of them to be clear me and sparrow and that's it now i do think what i'm gonna do by the way if you watch matt burman what he's built is insane i've never seen anything like it he's got hundreds of them working their tail off yeah we're gonna write about that because bro is bro is cooking man Yeah, he went deep.

1:03:14Bryan Goode:And all of his agent prompts are open source. I was going through them last night trying to see if I wanted to snag any of them. I like him a lot.

1:03:26There's not really pricing.

1:03:28Bryan Goode:You don't see a token price? Let's go straight to pricing. I can't imagine it's not, unless it's only available in the Codex app or something. Wait, we need to go back to where we were before. Oh wait, this is only available in pro as a research preview isn't it? Reading. Reading is fun Let's find out. I'm scrolling really fast Research preview Yeah This came out while we're on the stream which is why we don't Oh this is on Cerebris Heck yeah A thousand tokens per second A thousand tokens per second Oh my god Oh, God. Damn it. Is this going to be the button that makes me pay for Pro? I need Pro.

1:04:21Bryan Goode:Desperately. I got confused for a second there. So on ChatGPT, the$20 plan is Plus. On Anthropic, I believe it's called Pro. It's called Pro. On ChatGPT, there's free Go Plus Pro. Yeah, so this is the$200 version. Yeah, unfortunately, we're not paying for that right now. Now, they have rolled out a number of things to the pros. I mean, frankly, if you've got people paying$200, they ought to get something first once a while. But with that said, it allows them to throw it out to a smaller crowd and see how it's used, get some ideas, and understand before they're melting silicon all over data centers around the country.

1:05:04And that's pretty cool, though. You can see the difference here. So this is a file task, right?

1:05:12Bryan Goode:So this is translate a file. By the way, I just want to say 5.2 codecs can write really, really well, especially with a QA agent above it. I've had a number of things come through that I was like, you know what, that's surprisingly good for a coding model. yeah this is wild so okay so you have to pay chart go back to that chart look at what you're looking at there you're looking at topic duration and its accuracy is that what that says i can't yeah so basically it's saying yeah let's zoom in So what it's saying is accuracy is between just under 50 % to just over 50 % over task duration, which I don't know.

1:06:09Bryan Goode:Okay, and we're seeing that at like two and a half minutes, we're looking at the number, the same accuracy that five, three codecs is looking at at 16 minutes. okay i may be interpreting that wrong let's scroll down you are you are interpreting it wrong so as you can see here this is gpt 5.1 codex mini so it beats it beats codex 1 mini at accuracy right here so at 2.29 minutes it's 51 percent with reasoning effort the winner is still gpt 5.3 codex can do a task duration with an accuracy of 56.8 percent but it would take it 16 minutes to do that so yeah it's a pretty good reasonably comparable performance yeah yeah that's right you're looking at reasonably comparable format performance in significantly less time.

1:07:17Bryan Goode:Okay, wow, that's fast. I hit my space bar with my elbow. Boy, the charts are getting crazy. I saw a meme last night, Grant. Hey, go to my Twitter. All right. I think I shared it out. Do you see no evil? This is worth bringing up here. I am, no, I'm not see no evil there. I'm Corey Knowles. One word. That's me at the top. okay scroll down a bit by the way that's like a dolly three image in the top there that you see where you need to replace that you need a nano banana image up there it's so cool take it to nano banana and say make this even better honestly it's crazy good i i really liked i was trying to make some okay go to that's it ai is hitting a wall looks like it's about to jump the wall yeah yeah you said it yeah you said it i did i i just i saw that picture last night and i was like that is so crazy that is so crazy it's and look look at what's at the bottom of the mall okay that's oh three you know look at the distance from oh three to five too high and this doesn't even have opus 4.6 or 5.3 on it no no it doesn't no and both of those are also in my opinion a significant jump the morning it was funny the morning they dropped 5.3 codex uh the guy from meter on the tweet where sam announced it uh replied with dude we just announced the 5.2 codex yesterday give me a break here yeah that's what i'm saying they need to slow down like just stop releasing stuff all of you the problem is they can't do that because the the open uh chinese models are on their tails they even if all of the u.s um labs agreed made a gentleman's agreement to slow down technically the cheap or free alternatives could potentially surpass them yeah that's that's what's causing this to be so we just had someone ask what that chart we were looking at a second ago was communicating i don't know if it was that one with the wall on it this one so yeah let's talk through it so this is a pretty famous um chart and it's largely considered the best ai benchmark right now like you know all of the benchmarks and stuff that we were talking about earlier like accuracy and and um duration and all that stuff that that's important and it's important for this chart the rest of the benchmarks that you'll see like how well it does on you know math and coding and all this other stuff it's yeah and at this point like moot because you just need it to be like you just need to see does the does the ai do the job right so that's what this is measuring it's saying how long does it take to complete a task for humans yeah and where logistic regression of our data predicts the ai has a 50 chance of succeeding how long does it take?

1:10:24Like, if something takes me an hour, how long would it take, or can the AI do it with over 50 % accurate? So I'll explain that. It's kind of hard to explain simply. But basically, all of, what this means right here is that a task that would take me six hours to complete, GPT 5.2 can complete with over 50 % accurate. Yeah. Or at least, let's put it this way, with at least 50 % accurate.

1:10:50Bryan Goode:Does that make sense? GPT-4, which, by the way, was magic when it released. I'd like to remind everyone of that. When GPT-4 first released, we were all like, whoa. Yeah, look at it down there on its little flat line, hugging over the bottom line of the charts. A lot of people, this is what the, Matt, for folks who've read the newsletter this morning, this is what Matt Schumer was talking about in his essay. a lot of people have this as their idea of what ai can do right that's what the gpt4 quality is because they went to chat gpt a year ago and they're like hey it's stupid it hallucinates doesn't know what it's talking about but this is where we're at now where yeah at least 50 percent of the time it can complete a full task that would take me a human six hours to do that doesn't mean And that is not including what you can do with OpenClaw attached to it.

1:11:46Bryan Goode:The fact of the matter is when you're building out this org chart, I'm not getting 5.3 codecs. I'm getting 5.3 codecs that's run through a series of QA agents. And what's coming out of it in the end is really good. Things are fast right now. Thank you. I was going to suggest we had a question about how is Gemini keeping up with this latest chat GPT. Well, thankfully we have the patron saints, the holy men at artificial analysis who do the benchmarking for you so that you can compare. So they put all of their benchmarks that we were just talking about together in what's called the intelligence index.

1:12:24So this is roughly measuring the intelligence of all of these models compared. It looks like they do not have 5.3 on here, but they do have 4.6. Yeah. which was the latest one there and that's out of 53 gemini 3 pro which came out last year is out of 48 now what does that actually mean five but you can but you can dive but you can dive into the details based on what metric you actually want to know so like how expensive is it there's this intelligence versus cost to run chart there's um you know the image leaderboards who's actually the best at images, image editing, video, image to video.

1:13:10Bryan Goode:Yeah. So then you can look at like much this has progressed over time, which is roughly comparable to the meter chart that we were talking about. Yeah. And then you can go in and you can look at all of these individual ones. So like terminal bench hard, how good is it at agentic coding? And you can roughly benchmark, okay, based on these results, it seems like Opus 4.6 is stronger than geometry. grant has mercury ever run through any of these that's a good question let's see i don't know that i've ever seen it there we can hit them up though we should tell them to benchmark it um yeah yeah although if they open up to non non-llm models i wonder well no they benchmark speech and it's just a diffusion llm i mean the truth of the matter is right is right speed is something it would it's not there because it would be in number one on speed regardless of every other category it would definitely be at the top of speed let's go to meter um their organization i think they just recently released something yeah so here's where everybody lands on

1:14:24I mean they don't benchmark Gemini

1:14:27Bryan Goode:that's a good point maybe they don't they need to fix that unless it's is it one of those dots we're not seeing up there no no it's really just Claude and yeah

1:14:44cool

1:14:46Bryan Goode:let's see the length of ai tasks is doubling every seven months every seven months okay and if we're right now compresses and that taps out at at sonnet 3 7 which is crushed now by a long shot i would say well let's compare wait hold on so that you said 3 7 is 3 7 of the most recent one here. Yeah, and it was showing four hours? That would have showed. Or less than that. Three-seven. Oh, here. Three-sonnet. This must be three-seven. Three-seven sonnet. There we go. One hour. Like the 60 minutes. Okay, that line is one hour, Grant. From one hour to six hours. That's more than double. Three-seven sonnet come out late summer?

1:15:36Let's see. No, it's funny. I've not actually looked at this.

1:15:42Bryan Goode:Wow. This came out in February. This came out in December. So just over. Eight months. Ten months. Well, what would be seven months later? Let's see. I'm going to count because I'm a goof. September would be seven months from February. I was going to count on my fingers because sometimes that's easier for me. Okay. Asked if we could drop these links in the chat. Yeah. Yeah, I will. Okay, so by August, we hit three hours, right? Yeah. So that's roughly, I mean, definitely more than double. Oh, yeah. 60 minutes. That's like triple. And then we doubled that. Oh, we don't see 2.5 Pro because it's Gemini.

1:16:28We doubled that in less time. So, I don't know, it's about four months, right?

1:16:32Bryan Goode:And look how long it took to go from GBT2 to 3, from 3 to 4. look at the gap between gpt3 and gpt4 and the gap from gpt4 to gpt5 yeah a lot of people said oh gpt5 wasn't that impressive and like it had a lot of flaws but but they're wrong but it is better it it's definitely better and it's it's honestly a victim of something i think we see with a lot of models now in my opinion that it drops and you know until we get a hold of it not we grant and i but we the people get a hold of it you know we're looking at engineering use cases is it we're not seeing what it's doing in all of these other fields we're not seeing what it's doing in law yet we're not seeing all of these things other than like some numbers on a on a table and quite often and this started i would say with research models reasoning models as they come out, the impact isn't always known that day.

1:17:37Bryan Goode:Sometimes over the course of the following week. Finish your thought and then we'll switch to this. It takes a few days sometimes before we start to see cool things trickle in that people are doing and understand the impact. And honestly, if I'm frank, the launch of GPT-5 was hidden by two things. Number one, the fact that people could not figure out the router and were trying to change models. When the truth is, I don't change models unless I have something big. I'm on auto 90 % of the time now. It's a while before I was comfortable with that. But the truth is, everything I get is good enough with what I'm doing at the time.

1:18:16And if it's not, I go change it.

1:18:18Bryan Goode:But as far as walking in and being like, this is a thinking task, I don't do that. I just let auto figure that out. usually it's right. Anthropic gave like extra usage like$50 extra over this week so I was just opus 4.6 extended thinking for everything. And in 6 minutes I ran out of credit? I'm just kidding. I'm sorry. No, it actually worked out pretty well. I was able to write for a long time. The downside is I maybe lost access to my Claude account. We'll see about that. They might have killed my Claude. Anthropic has this problem right now and the problem is for some reason when you try to upgrade your account pro to max sometimes you lose your account and like there's no fix for it and there's no customer support from anthropic so like anthropic if you're listening to this like please fix that what the the you know what like why is that a thing but anyway so question so this is the article that we shared this morning something big is happening this like blew up for reasons unknown but i think it's because it kind of encapsulates what people are feeling on twitter it also grabs that whole thing where if you're if you're sitting at kind of the cutting edge of this stuff and by the cutting edge i mean the cutting edge short of working in a research lab and like grant and i ingest a lot of this and i don't know if you've experienced this grant but this is what sucked me in yesterday Have you ever told a loved one about it And had them get this look on their face Just don't talk to people In the real world about this Really?

1:19:59Bryan Goode:It's a closet AI fan Here's the thing, because people either don't want to hear it They don't care about it Or it scares the ever loving you know what Out of them They'll look at you like you're crazy Like anybody wants to see anything I say I'll absolutely Well it's a little complicated today with OpenClaw But I would absolutely sit down and show you that I'm not talking out my butt here. Like, there's real stuff happening, and it's happening at a pace it never has. And while I have a few subtle disagreements with Matt's post, I think it is a beautifully written essay that encapsulates an awful lot of things that I have thought as well.

1:20:41Bryan Goode:and it was really nice to see it not just circling around a community of us ai nerds it was like it was on the top row of the drudge report yesterday which is insane yeah you know nikita beer who runs the x algorithm responded which was very funny he was the top comment yesterday he said well done you changed the world with a single article And it is interesting because it's like, it creates a lot of conversation, but basically the crux of it is he says, you know, this is as big as COVID. And you know, remember in February, 2020, when people were talking about COVID, it was something that was like a very far idea.

1:21:28And then all of a sudden in like a week's time span, like it became like massively world changing. He thinks that we're approaching that moment with AI. and he says, I think this, I know this is real because it happened to me first because he's an engineer and now he's at the point where he no longer needs the actual technical work of his job. He just talks. So basically software engineering is more or less automated at this point, right? All the big labs said that you mentioned it earlier, the open AI and Anthropic, they're writing a hundred percent of their code with AI.

1:22:00Bryan Goode:As a writer, if I am writing something, it is because I want to write it. The truth, I mean, like, I don't have to. Would you say you have to, Grant, to get what you need or want in most cases? The thing that AI still does not do well, which perhaps it never will in the large language format, is maintain a coherent history of what has happened and what's important and what's not. So I'm not saying I'm perfect at this, but I remember everything that I read. I remember everything that I hear about. You know, I don't remember everything, but I keep that information like he was talking about this week, like latence in my latent space.

1:22:50So it's like, it's kind of in the back of my head. So when something comes up, it can connect the dots between what I'm reading here and other things that I've heard that are related to this. And that's the unique thing about our brains and how we remember things is, you know, the neurons they connect and you know, the neurons that wire together fire together. So you have these like web of connections in your head where you can draw stuff in. I also am good at pulling from different web articles that are related and trying to connect unrelated ideas or related semi-related ideas together. And then I know what's current.

1:23:22The AI cannot do that. It cannot do all three of those things at the same time. It can't search perfectly, find everything. It doesn't know, you know, everything that I know of what's happened before. and it can't coherently pull it together in a way that, you know, like that doesn't, that leaves out some things that doesn't bring, like often doesn't bring in the right context. And that's, that's really the trick. That's, it is, you know, at the same time,

1:23:46Bryan Goode:at the same time, it has an awful lot of things that we don't. You're right. You know, I don't, I can't solve novel mathematics. there you know i i think i think and i wrote a thing about this i talked about this i guest hosted on the ai fix podcast last week i think it comes out this coming like tuesday i believe we'll share it out um and we had a discussion about this very thing that was like are we we're discussing the nature article where the nature declared recently that hey we have agi let's quit talking about it and move on uh and i i had done a column back god when was it uh september october something and it was about the idea that while we're waiting on perfection amazing things are happening here and and you know by plenty of definitions you would say we already have agi whether whether calling it that matters or not the idea being that there's not any human on earth who can accomplish everything it can do right now.

1:24:56Bryan Goode:Like, as far as, there are definitely things we can do that it can't. There are things it can do that we can't. But we cannot in any way, shape, or form do all of the things that it is capable of. Like, when you look at it, it's what it is capable of. It is definitely beyond any one human, in my opinion. Yeah, I think that is unequivocally true. The question is, it's twofold. It's like, number one, it's doing this because it's been trained on a lot of human information and a lot of human handholding. Mm-hmm. So when will it get to the point where it can generalize and figure out things on its own?

1:25:37If that is the true definition of artificial general intelligence, then we're still a long way up. If, if the question is, is it, can it just do nearly anything or more than any single human can? Yeah, I think we're at that point. But then the, but then it becomes an application problem. Then it becomes a, well, how do you actually give this thing the tools it needs safely to actually navigate the web like you or I can, or navigate any interface like you or I can. It doesn't have application general intelligence, which I think is key. That's fair. Like it can't go between applications. Like it can go through many applications, which Connectors helps you with, but because of the licensing deals and because of the red tape and exclusive access here and there, you can't do everything that you or I can do.

1:26:26Bryan Goode:That's true. That's true. I think there are a couple of things at play there. Number one, and we should talk about this later, is WebMCP, the agents only browser that Google dropped in a developer mode this week. I want to, I need into that. If you, if you've got a dude, I'd love to, I'd love some help if you're listening. Okay. But aside from that, something that's interesting, and I know I'm bringing it up again, but open club pulls some of that glue together in a way other things don't yet, in my opinion, partially because security was not a concern. That security was not a concern says, you know what yeah sure i can just keep an endless context and history document about you i got an md file about me that it created and it's just i you know i mean i guess it's limited by my i got like five terabytes i guess it could be limited by that but i mean five terabytes a text file is uh but you could fit your life in five terabytes of a text file maybe not yeah there was an interesting there's an interesting framework that i think mastra mastra i don't if I'm pronouncing that right, just put out, which basically uses a three-agent system to manage your memory.

1:27:45I'm going to see if I can find it here. That looks cool.

1:27:52Bryan Goode:I'm literally talking about the graphic design, not what it does. Yeah. I have no idea what it does yet, but you got me excited. I don't know. What's your racial memory? Is this it? Okay, let's see. Okay. so what they're saying here is observations are log-based messages formatted text not structured objects so this is the universal interface it's easier to use it's optimized for llms they use a three-day link to that yeah i'll post them in the chat actually because i'm going to do a thing you're going to give it to your uh open claw right now i'm going to go give it to sparrow and saying hey uh what about here anything yeah new memory system so what i saw they were talking about it earlier and they said they were using like a three agent system and perhaps i got this wrong but essentially one agent is like keeping things active one agent is um you know just maintaining like the conversation with you and then the other one is like a forgetter and it's like actually actively curbing stuff that you don't need which i thought was really cool really cool conceptually actually i'll go to their youtube where they were talking about this i'm trying to figure out if there's a way i can show my telegram on screen

1:29:16Bryan Goode:without exposing anything that could be

1:29:23Bryan Goode:I don't think there's anything. There's an ID. I need to get rid of that. Okay. So for people who are interested in this, so here's the basic memory technique that they're talking about here. They store previous messages. They stuff them into the system prompt with each new request. And then, like, until you hit the context window limits, this is a problem. So this is the old way. And then they have a lot of memory solutions that they talk through here. there's long contacts, which is just means cost exploded performance degrades because we're not going to get into the details on this, but language models are not good with long context.

1:30:00The current architecture. There's RAG, which retrieves relevant contacts. And then compaction, which basically compacts the context to try and get just the relevant details. So their version, which I was watching earlier, is they use three agents. And this is really blurry.

1:30:26But basically, it uses the actor, the observer, and the reflector. So the observer and the reflector are the subconscious mind, always running and never interrupting. And then the observer... Really interesting how this works. and this is a framework that people can apply yeah yeah you want to show absolutely

1:30:50Bryan Goode:let me take my screen down okay i i'm depending on everyone to please if we have anything in here please don't rob me i don't think they're yeah all it out if there's any problem but but i just gotta bring this up so here's what i did so what are we looking at just clarify what is this here's what we're looking at this is how i communicate with my clawed box what you should know is i came up here after you dropped that link and I said, hey, read this paper. Do you see a way we can make this happen locally? Now, first, I dropped the paper in the second chat. So it came back with a bunch of questions that we're going to bypass.

1:31:26Bryan Goode:Down here, it's like, yeah, we can absolutely approximate master's observational memory locally without adopting their whole stack. What they're doing in plain terms is keeping a stable observation block at the top of context, Keep a raw recent messages buffer below it. When a raw buffer gets big, run an observer that compresses raw new observations. When observations get too big, run a reflector that I don't even know what some of this means, dude, if I'm being God's honest. That's what we were just talking about. So they use this three agent system. They have an observer, the actor, and the reflector.

1:31:58The reflector prunes the important stuff. The observer is doing something, and I missed that part. Okay, well, here it is.

1:32:05Bryan Goode:We'll show you real quick here. Store two text files locally. the observer observations.md stable prefix append only the log uh return recent turns and tool outputs in the raw buffer and that adds one background observer pass on a schedule every x minutes or whenever it exceeds in tokens uh run an agent that reads it appends five to thirty new lines to observations in their format excellent add a reflector pass okay okay but here's here's the deal that gets you 80 of the benefits stable compressible memory that's readable and cache friendly how we'd integrate it into your setup use files in your existing workspace we'll put it under these have the main agent append to the raw buffer anytime we have long outputs append to short okay cron jobs to run observer and reflector every 30 to 60 minutes excellent job is a local automation con drop is yeah it's a local automation it's basically a scheduled task that it fires up and runs whatever tools you have uh and it was part of this that was a nightmare for me setting up where i had all the problems because yeah my struggle was that because i downloaded it while it was still claude bot not all the files deleted so when i download the new update and it's open clause suddenly it's calling all these claude bot files and getting errors everywhere and that's why my setup was such a nightmare okay what you'd feel immediately Less context bloat from tool output.

1:33:33Bryan Goode:Excellent. The assistant stays consistent over multi-day workflows without dragging the full transcript around. Excellent. A human-readable log you can skim and edit. Sorry, this keeps scrolling on its own a little here. And edit if needed. One quick clarifier. The claw. Do what? It's the claw. It's the claw. Okay. What a great movie. One quick clarifier before I implement a prototype. That's the first message here, okay? Its first message is one quick clarifier before I implement a prototype. Do you want observational memory to be global for everything or scoped for workflow? Global. Okay, global.

1:34:14Bryan Goode:One observations log for Sparrow per domain. Separate logs for morning brief, podcast, social ops, et cetera. If you pick two, my recommendation is three logs. and here's what I do when I don't know for sure what's your overall recommendation yeah I was gonna say I think you would maybe want to have the main one and then you'd want to have situational ones so the pruner could prune the main one based on what's the most important task but then you know situationally keep the relevant ones yeah let me say that like should it have one for you and and as well as those for others like yeah like one as well as individual ones for others so so for people who missed this cory is literally talking to his it's a local instance of clodbot right there's a computer sitting downstairs in my easy chair and he's using telegram as his communication tool with with it by the way you'll see i downloaded this on January 29th, so 13 days ago.

1:35:24Bryan Goode:It's already running your life, basically. It's before that, because it was a while before I decided I'm going to connect Telegram. I had errors for a while before I did that. I was just running through OpenClaw's dashboard, because it's got a big UI. You just type OpenClaw dashboard, and the UI pops up with charts and all this stuff. Yeah. You can chat with it there, just like a chat. Just like a normal AI chat. but I prefer this because I can do it while I'm around. Okay, what it recommends. Go scoped, not global. Start with two. Per domain, observational memory with three logs. Ops, which would be Sparrow probably.

1:36:00Bryan Goode:Editorial morning brief. That's kind of my biggest task that it does every day. Still ironing it out, but it's really good. Podcast and social. Guest outreach, templates, show style, voice and tone rules, what you like, don't like. Why scoped wins. Less cross-contamination. You don't want a random ops detail to leak into a social draft. Okay. Better retrieval by default. Morning brief observations should pull morning brief observations. Not everything you've ever discussed. Easier to prune. It's also easier to troubleshoot too, because you know, here's where this problem's happening. It's not affecting these.

1:36:33Bryan Goode:Like I can do that in my morning brief. I can be like this section specifically is screwed up. These are good. These are good. This one is broken. And it knows, ah, I'll go down and chastise that agent for you. When you ask something, Sparrow will include the right observations block based on task and skill, so you get the benefit of that. Okay, excellent. Okay. Okay, before we go with your recommendations,

1:37:09Bryan Goode:will this interfere with any of what we're building out from last night's research drop

1:37:26Bryan Goode:here's the key research drops the the time gap between a research paper releasing and someone actually having access to what happens in it is months. Years in some fields. This is I saw a preprint on Arctiv that dropped 30 minutes ago. We can drop it in. We can do it now.

1:37:56Bryan Goode:Let's see what happens here. Okay, we got some notes. We got a note. I'd really like to get into OpenClaw at some point. I'm super concerned about where to host it right now i'll be honest with you i was super concerned about where to host it and i thought about it and i debated and in looking at what was possible i was like nah and just did it anyway i i i did it casually and slowly i don't have like all of the channels turned on i didn't go download a bunch of skills from the internet yeah you just need one just pick yeah pick the best pick the best one that you're most comfortable with yeah i very much started slow and uh and as i got more comfortable began and learned more began deciding what i could add okay let's see what it said here i'll give you i'll give you all a preview of what's going on tomorrow actually it might be sunday this might be sunday depending on how much news comes out today but i i thought we should turn this into an article too like for yeah i will we're we're behind on doing the january articles but i'll crank them out today and see how far i get you want an agent for that yeah yeah give an open call i'll be like please turn this transcript into a hyperlinked article thank Thank you.

1:39:02Good question. No, it won't interfere as long as we implement it in the right place.

1:39:06Bryan Goode:Essentially, what I wanted it to do was just I wanted to verify for my sake that this other project I'm working on with it, that this won't break it. Failure mode. Cool. Avoidance. If you want, I'll start with ops only. After two to three days, if we like it, we add the others. Nah.

1:39:32Bryan Goode:recommendation watch this let's do this shit

1:39:39Bryan Goode:that's funny and it's going to come back and be like uh it's it has an attitude the other night i i saw something else someone shared that was like hey you know you could i i had started working on its soul because your open claw has a sold out md file where you go and tell it who it is and i can't remember i just yanked it off twitter this is i am not a good example of security practices i need to know that i try but the truth is we're gonna share some advice i'm gonna try it if it excites me i'm down yeah yeah we're gonna have this implemented in the amount of time it would have taken me to read that whole article yeah that's what's wild about all of this is you can I like I like Dan Shipper's approach where we're coordinating we're going to try and have him on the show here pretty soon but basically they they have this really great blog post and at the top they just have this button that's like copy this to your agent and I was like every website needs this dude that's the thing I'm reading research papers in the morning like I get these summaries and I'll look at them and I mean I can literally just be like hey this feel smart go do this while i'm gonna drink my cup of coffee and write an article yeah like i still do the things i love the truth is when it comes to like my job and the things i do what i'm doing here there's nothing i do in my job with a couple minor exceptions that i don't love like like what we're doing right now this is uh i would be doing this for free given the opportunity but uh don't tell technology advice i don't want to say that too loud but uh the truth is i really would i love this and uh there are elements of it i do want to keep but what i'm doing is is the stuff that's like oh i had to click open 75 tabs today to go check the news so instead of that But why don't I just have all the news come to me and be summarized already where it's a quick glance?

1:41:43Bryan Goode:And if I want to read deeper, cool. But it gives me more than the headline I would have seen out serving sites, but less than what I would get from reading the whole thing. So if I want to read the whole thing, if I want to double click on it, as we said earlier, I can absolutely go do that. Okay, see here where it says typing? What you should know is that it's working right now. It's implementing all of this. it'll be back in a few and i'm just going to leave it up here until it comes back and says hey you're good to go cory or here's what we did and and then you'll just notice an improvement while we're doing that do we want to very briefly um or or maybe like what once we get to a stopping point here do we want to review everything that we talked about because i think that'd be helpful like we talked about a couple of different things that i think you know if we put in summary we'll kind of put all of this together in a neat little bow for yeah i think that's oh hang on before we do that yep all right let's see what it says done we're doing this what i implemented locally scoped full recommendation i set up observational memory as three separate domains no global drunk drawer one for ops one for the morning brief which is my update i get in the mornings and uh one for my podcast and social agent teams uh under users.

1:43:01Bryan Goode:Oldwood, Claude, Memory, Agent Alpha, Observational. Okay. Raw buffers are where stuff that happens gets dropped. Observations are the compressed log. I added six crime jobs. Cool. Which makes them run every so often. They run at staggered times. It does that so it doesn't overwhelm my computer. And set the house on fire. Or like scare Melissa. So right now we have the memory layer plus the compression layer. Next step is wiring. Have key workflows append short line into that. Okay. After the run, that's how it stays current without you thinking about it. Cool.

1:43:43I'm going to share two links in the chat while you're doing that.

1:43:46Bryan Goode:Wire it up.

1:43:51Bryan Goode:I need 60 seconds. Grant, are you good if I stepped away for 60 seconds? We're two hours deep. I'll answer this question in the chat and folks can talk about that. So in fact, if you want to pull your screen down for a second, I'll.

1:44:13Okay, so we had one question. You're muted, but your camera is still on. Do you have a, does it need me or is it a.

1:44:24Bryan Goode:No. Okay. I'll be back. if there are questions that I should answer. I'll be right back. Cool. So somebody in the chat, Simon, said, I really want to make a local... Wait, what did you say? Let me read it exactly. I want something similar, but on specific topics, you know, like RSS. I have a curated news list from sources I defined. So that's super easy to make. One of the easiest ways to do automations without needing, like, any open-claw local installs there any security risks is this tool called Tasklet. I was actually lucky to meet the team working on this and they're really cool, really cool people.

1:45:04It's super easy. So for example, you could say like, give me the topic, give me the topic that you want to make an RSS feed on or something. I'll get this prompt started.

1:45:21I'm saying, hey, I want to make a custom RSS feed style daily report.

1:45:28I'll say on specific topics that I hear about, specific slides only. Can you help me set up an automation for that? And if so, what do you need?

1:45:50I'm going to go ahead and kick this off here. So what's happening is it's doing a lot of the agentic work that, you know, you would have to work with your open claw or whatever to set up. It's fine. There's my email address. If you want to email me, oops, and myself, but that's fine. Absolutely. I can set up a daily automated report that is totally doable sources, report details. So go ahead and give me like the topic that you want, maybe like a couple of the sources and we'll build a custom one here. I'll create a couple of suggestions like about crunch, do gadget. It's another outlet that people like.

1:46:37I don't know if Bloomberg, maybe Bloomberg we might have to pay for. Bloomberg and let's say the XPI talk about.

1:46:49Let's see what happens.

1:46:54Um, yeah, so it's going to email me this list. It was like, yes, go ahead and email me. Okay, great. I got your, I got your answer now. There's a bit of a delay on the chat. So I will go ahead and make it second. So let's make a second one on this topic. say, local landlord tenant laws changes and update. And then I'll say, I don't have an xAPI credential view.

1:47:38I'll skip that.

1:47:43Now, let me show you the alternate version of this. So I'm going to share that. So there's another one that basically is for this specific use case. And what I can say is news. Just say the same thing here. I'll say local landlord, local landlord tenant laws, changes and updates, start scouting. And then what's going to happen here is it's going to then go ahead and create what's called a scout. And I can schedule it and I can say, let's say every six hours. And I can set that schedule. And this will actually email the report of what it's found. And I can show you an example of that, but let me pull up a safe window.

1:48:36Bryan Goode:By the way, you got me addicted to napkin.ai the other night. Oh, that's cool too. Okay. I'm not to overwhelm people with tools. We can go to that one in a few. Yeah, but let me pull up an example of a scout.

1:48:55I'm going to go.

1:49:04Bryan Goode:oh my god grant what's up listen to this what you should expect you won't see anything in chat it's silence but the observational logs will start filling in automatically over the next run cycle and reflectors will keep them from bloating if you want a little bonus upgrade to this i can wire errors too so if a cron roll fails it logs it to the ops raw buffer with the first line of the error and then debugging is instantaneous it improved it grant that's that's well it remains to be seen right if it actually works but yeah that's cool we're gonna see we'll report back on this next week i promise you that okay so here's a result of a of a doubt that i got so let me know if you can see my screen cool basically this was a scout that found Gemini 3 DeepThink leaps in scientific reasoning.

1:49:55I saw that. It emailed me, it emailed me about DeepThink and it gives me the link to the Google blog, gives me the New York times coverage, open access government page, blah, blah, blah. And then it even gives me a why this matters. So it's like quite literally like the AI rundown version of it for anything that I want, which is really cool. And that's like a,

1:50:18Bryan Goode:Did it pull that out of your email grant? yeah so this is this is no this is it it goes out and it searches the web like i'll show this tab so what's happening is it's it's doing this um it's doing this research flow here it's really beautiful ui it is and then once it finds stuff it will email me what it found on the uh recurring basis with which i like if you tell it give me an hourly email give me a weekly email whatever it would, it would just follow that schedule. Exactly. Yeah. They gave you, it gave me three examples that I can start with, but this is, this is a nice one. And this is like pretty much a bunch of different sources covering all these different topics.

1:51:07And this is, it gives me this like every hour, basically. Now what I was doing with task list is I was creating my own custom version of this, right. And you could do this for any workplace automation that you have, where you could say like, oh, you know, let's say you don't want to just automate like a news report. You want to automate like extracting data from a spreadsheet or creating spreadsheets for a given, you know, report, things like that. You could customize that and you can use triggers. And those triggers will, you know, you can schedule them. You can do them based on web hooks, RSS feeds, et cetera, et cetera.

1:51:42So this is kind of like a copilot, like the same thing that copilot was showing us earlier, but you can do it with all sorts of different connections. And it's very, very customizable. Yeah, I really like this tool.

1:51:55Bryan Goode:You've been raving about Tasklit for a while now. Yeah, people need to check it out. Yeah. So should we, Corey, should we just very briefly go over everything we talked about to kind of like put it all in perspective? We should. We should. We've talked about a lot of things today. We did. Let's show, let's do a review here. So Copilot Studio.

1:52:26So this is where we started, right? If you're in Microsoft Copilot right now, this is probably your best bet of making agents like we're showing you, because it's already connected to your data and your information, you know, that it has all your spreadsheets and all your company connections. So CoPilot Studio is probably the place that you want to start. Now, if you then are an administrator or someone who is in charge of managing all these agents, then you would go to Agent 365, which is what we just talked about, Agent Studio. And that's where you can see the control for all the different agents.

1:53:06Bryan Goode:I was more impressed with that than I was prepared to be. And there's a realization I had about it that I hadn't considered. I kept looking at it and I'm like, it just looks like a chatbot. But the truth is, what it is is much more impressive than that. It's essentially, let's say this, Open Claw is the rat rod of agents. and this is more the lexus if that makes sense like like this is the one that's safe to drive it's got airbags it's uh you can find out where it is if it gets stolen mine might rob you it might break your leg when you get out you know it's a little bit rough around the edges but this is and you know why i think this is a bigger deal than i've anybody's given it credit for it's because you could use this in real business an absolute normal business could execute this today yeah yeah i agree i agree um so that's so that's agent 365 then you know as cory said you know open flaw is like the wild west version of agent 365 if you want to share your screen on that again yeah i do because i want to tell you what it did here it kind of moved on its on its earlier suggestion as well.

1:54:30Bryan Goode:There's also, as you're pulling that up, I'll just say, there's also Claude Cowork, which we didn't show. Codex, which we can show if people want to hang out after the hour mark, but I think we're probably going to. Cowork is another much more user-friendly approach. It's only available on Mac. I've been ranting at these. No, no, it's on Windows now. It's on Windows now. Yeah. That happened as of yesterday. I'll be having that this weekend. Okay. Thank you, Grant. I missed that. I can demo Cowork and Codex if people want to see it after the hour mark but those are other agents that are yes the co-pilot agent overview with Brian Good from Microsoft was at the very beginning of this but as soon as this is over it will be available as an archive and you can watch it at your convenience anytime you like and I recommend it he was a really nice guy very knowledgeable happy to speak on what it could do as well as what it couldn't he didn't hide its weakness I don't think and uh I think it's worth you know he's like I don't even know some of those words but I'll tell you what if you call me we'll figure it out I say that he didn't say it that way but but it was you know it was frank and it was honest um yeah appreciate that okay so I asked you know okay so here's here's what happened when you were gone it sent me a quick news update in the middle of this so apologize for that uh all right so here's where it said wired what I changed I added a local script.

1:55:57Bryan Goode:We'll get into all the technical end of this here. But it's got these connected now to all of my agents. Or not all of my agents, but the bundles of agents. What you should expect. But the observational logs will start filling in automatically, and the reflectors will notice them. If you want one more nice upgrade, I can wire errors too. For example, if a cron run fails, log it to the ops raw buffer. debugging becomes instance because what happens is when the when the observer is pinged and sees it the observer then triggers and it goes and does the thing and fixes it uh so i i then asked heck yeah do it and it didn't respond so i said hey do you wire up the errors too and it says done i wired error logging in a way that works even when a job fails since it can't reliably append its own error uh added a local monitor script uh it depends on one line entry into ops raw buffer okay uh silent cron job runs every 15 minutes observational memory cron error monitor if something breaks it'll show up as a that a giant red dot uh example format that's what it'll look like okay right now it's quiet because there aren't any new errors in the monitored set, which is exactly what we want.

1:57:16Bryan Goode:And down here, it said... Okay. Oh, that's it. Okay, so now what we're going to do... Grant, we're going to play with Napkin AI, because I'm going to ask Sparrow, hey, can you give me an updated full run of our org chart? Not run. List of our org chart. of agents and it'll come back with a bit of a list not a matt berman list that's for dang sure yeah crap i watched his full video last night and he's i he's my inspiration that's what i'm going to say matt you're my inspiration i asked in the chat by the way who wants to stay on past 1 30 4 30 p.m to see a few more tools and practical workflows if anyone on our team if If you have to go, or Corey, you have to go, let me know.

1:58:13I can stay on.

1:58:15Bryan Goode:And same. I'm okay. I don't have a meeting with you. Well, we can do that here live. We can look at the news. It's all right. I already threw the notes there. No, but I think kind of fun to workshop it. What we would normally discuss is already in your DMs this morning. I sent you the notes from the meeting with a bunch of ideas. But, of course, that's been blown apart now. but there may still be some elements you need. Yeah, because I think GPT 5.3 codecs or ARC isn't that big of a deal for most people because most people don't pay for it. Probably won't. Okay. Grant, would you be willing to pull up...

1:59:01Bryan Goode:Uh, Napkin AI and let me send you something? Yeah. Okay. I want to send you something and what I'd like you to do is drop this in there and tell it to make me a visual org chart. Okay. Because I think it will do it awesome. Okay, one second. It's locked out again. Okay. Bam. All right, I'm going to share my screen. I'm sharing my screen whenever you are ready.

1:59:38Can you see my screen? Okay, so this is napkin.ai. Shout out to Jeff Su, who I learned about this from. He's great. And basically, napkin is like the fastest way to make like simple graphics out of prompts. Are you sending me text that I should literally paste in or is it more of like a prompt that you're sending?

2:00:02Bryan Goode:I would I would say here's this big list I have. And it's an it's an it's an explanation of an org chart. Please, please create a visual. You know, like. So what I'm doing is I'm copying this from Corey's message he sent me. And I think that's missing some because there are a couple I don't see, but I may have misremembered how we. What did you say? One of these is implemented as just scheduled jobs, not a specific agent. Okay. This is a comment. Delete. Okay. Normally what I do is I think I can just highlight everything and then create. Yeah. Very visual. I'm just going to do that. Excellent.

2:00:48I can't zoom out anymore.

2:00:52Bryan Goode:That's cool. Okay. So this is what it created. That's what's on my computer. Okay. So, sorry, I can't zoom out anymore. So we're going to have to like scroll through this. So maybe I can zoom in a little. I would say it's structures a little bit wonky, but that's, that's okay. Okay. So you're saying that you talk to agent alpha. Yes. And then on your computer, you have agent.qa, reliability, brief publishing, scouts. My news and research scouts. Okay, those are jobs, and it's showing which ones they're done by. Okay, cool. Okay, yeah, so my ops is handled by agent alpha, reliability and guaranteeing quality.

2:01:37Bryan Goode:The QA is not showing anywhere, but everything runs through the QA. and I thought we had those set up as individual ones and it might be that I haven't clarified that. I'll check it out after I get off. Yeah, this isn't perfect either, right? Like maybe it interpreted something you wrote wrong. But basically what we did is we shared the exact layout and then hit this generate visual button. And then like, you know, 15 seconds later, we got this graphic up, which is editable too. Like we could go in and change any of the words. Yeah. Yeah. I love that. Really cool. And then also what this does here.

2:02:20By the way, I haven't even fully explored all of this, but for example, you can generate this. Yeah.

2:02:26Bryan Goode:Sorry for putting me on the spot and saying, go show me how it works. No, no, no. It gives you a series of alternate options too right here. It'll give you different suggestions for how you can visualize it. So that's the default that it picked. This is the one that it picked earlier. but then there's different alternate versions in here so it's generating like do one with but i thought you could um very easily swap it out but yeah yeah yeah yeah you could and move things right i see how it did it i just wasn't looking at it right at first that's good though the reason i wanted a visual so i could like print it out on a piece of paper and look at it for an hour and decide like where are my holes what am i not doing yet but it's cool i never ever want to do myself again and let's fix that today like look at how easy this is to do right like you can just pick this one now it's got this oh yeah i mean this is like a framework that's not applicable here but yeah yeah because i don't know what a pestle analysis is yeah it's just a framework for work this is the you know you can choose is it a process is it a flow chart is it a customer journey, cycle, the different data type, timeline.

2:03:40So this is great for if you have to give a presentation and you wanted to

2:03:46Bryan Goode:make a quick graphic. You didn't want to deal with Nano Banana getting half of it wrong. That's awesome. Yeah. But yeah, it sounds like people are down to hang out. I have my other... Hey, you want to... Hang on. I gotta quietly ask Grant something here because I don't want to say it out loud if he's not down Are you gonna message me? I'm gonna message you in the studio chat, one second everyone this could be cool Well commentate like it's the Olympics or he's typing it looks like he has sent the text we're waiting for the judges to review looks like he's negative points on style it's in our studio chat okay one second let me pull that up by the way shout out the olympics anyone who's watching the olympics i'm like i haven't watched the second of it i know there was a bad crash by a gal the other day that sounded yeah yeah let's do it let's do it i like that idea okay cool idea you all want to help us pick what goes in the neuron tomorrow yeah i'll share some context around this so earlier this week we asked everyone on sunday like what what if we just take a break from covering the most breaking news and just cover practical use cases and overwhelmingly everyone was like please give us practical use cases and so we've been playing around with different ways of how to do that and so you know we've talked about clock code we've talked about different agents we've talked about i forget what we talked about on monday oh just like compounding engineering and like how how to try and keep up with the demands of work because there was an hbr study that came out that basically said like when you use ai work you're not working less you're actually working more because it reduces the friction and that sort of thing they ain't doing it like i am sorry yeah unless you have open claw then you're freaking like story over here and just chilling um but i think agent trying to get really deep into how all the different types of ways that you can create agents like there's the command line terminal agents, there's the new Cloud Cowork, there's, you know, Tasklet, like I just showed people, and Utori, which is like a really niche agent, but still useful.

2:06:04There's a lot of different types of agents, and I think that's the way that you actually speed up your work is like, yes, it's fun to chat. Try to figure out, like, use the chat as like the prototype phase, and then once you figure something out, turn that into an automation and don't do that work ever again. Like, try to avoid doing that work ever. That's sort of the idea.

2:06:25Bryan Goode:I agree. So with that in mind, do we still want to look at the news? Want me to bring up our note sheet? Yeah, go for it. Show them kind of what we do each day? Yeah, let's do it. Cool. Okay. Okay. And then we'll get into Cloud Cowork and Codex, and I'll show you some stuff on my computer related to that. Okay. This is Grant and I's morning news huddle, or afternoon news huddle is a little more accurate, where we say things. This is just where we say. This here is just where we save things that are like, not today, but we got an idea that might be cool later in the week. Main consideration is where we're trying to figure out what's our main for tomorrow.

2:07:04Bryan Goode:Yeah. What do you think that is, Grant? A couple big releases. Actually, if you're doing use cases, should it be a Codex 353 Spark use case? It would be. The problem with Spark... combine a how-to and a oh yeah we don't have it yeah the problem with spark is not everyone has it we could do a codex use case just a codex app because we did cloud code Microsoft agents too yeah we could we could do we could do Microsoft agents um what are the what are the maybe we could use something as a hook here what what do we have what else okay yeah let me show you what all I pulled these were all pulled early this morning just so you know I've got six for around the horn that you can pick from and i've got six for intelligent insights because it's friday manager scrutiny after its auditor ernston young flagged the financial engineering used to keep a massive data center build off the balance sheet so a little drama these can be tightened up as much as we want to uh mustafa suyman said microsoft is building its own models especially for enterprise and healthcare to reduce dependence on open ai it signals a tighter more vertically integrated AI stack and specifically said, we're looking for self-sufficiency in the AI space.

2:08:25Send a line in the chat if you all like Copilot and you want us to write more about Copilot with like a step-by-step, because we can do, we can do like Mustafa's, you know, comments, and then naturally transition that into like Copilot studio step-by-step.

2:08:42Bryan Goode:You know, and the angle we look at this space from unfortunately is up so close that it's easy to forget that most people are probably using copilot on a work computer during the day yeah uh you know whether they love it or hate it they're probably it's there yeah in a lot of cases there's a great if you want to know how to use it we we have absolute great access to to copilot stuff where we can do things like that occasionally the the one thing i'll say is that there's a a great uh security series of videos that i watched yesterday, one of the guys in a panel on IBM, so there's an IBM security panel, was like, the biggest thing for all of this stuff, whether you're talking about OpenClaw or all these other agents and the security issues with them, is as an IT professional, don't say no, say how.

2:09:33I love that quote. It's like, okay, if you're going to do this stuff, which if you say no, of course the people are going to do it. They're going to use shadow AI. They're going to have tasklet automations going. They're going to have OpenClaw on their personal computer,

2:09:45Bryan Goode:but like be texting give them a here's how it's okay sort of thing instead of a no yeah yeah no you're opening a door yeah yeah and so like that yeah go ahead i'm licensed at work for it so i would benefit in some ways i like the use case idea would love to be able to create agents to do tasks and processes so i never have to do it again amen reach it all right i think that's a good that's a good focus for them all maybe talking about how to do that with triggers and hooks yeah yeah i think it's a good idea i would definitely uh do you think this is at minimum and around the actually that might be a good blurb grant i think we'll i think we'll probably use that as an intro for this uh this article that's a good call yeah because then you could hit the the five three and gemini three deep think in the blurb if you want yeah that's missing from this that's pretty cool Does that sound good to everybody?

2:10:39Bryan Goode:I'm trying to make sure we ask because this is just a spur of the moment idea and there's not a ton of people in here and you all have been so interactive and kind while watching us absolutely deep level nerd out this afternoon. By the way, on the deep think. My afternoon social QA report is done. It's saved and in your Google Drive. I love that so much. Sorry. No, you're good. By the way, so two things on there that are interesting. So DeepSeq, they have 1 million plus tokens, right? The idea there is that that can make long horizon AI agent workflows more practical. We haven't talked about open models this time around, but if you remember, we talked about the length of tasks an agent can do.

2:11:27We also talked about Masra System, which is way to basically like help with memory management across those long tasks yeah yeah yeah that is one of the key things where you know one of the things they mentioned is you can expand the token window which is how much input and output that you can put in and the agent can remember and

2:11:48Bryan Goode:keep relevant but then you get a model that sags a little in the like yeah sure we can do a gazillion token context window but that doesn't mean the whole answer is going to be tip-top quality it's going to sag over time, right? Yeah. So to enable an agent to be able to do those long-running tasks that take me, a human, six hours, or Corey, six hours longer, it needs to be able to maintain latency, like be able to keep that in the latent space and be thinking about it over and over and over again, keep it top of mind. So having 1 million contacts helps with that. It does. It does. And both DeepSeek and - this was interesting too grant it might be worth mentioning alongside maybe a codex note but uh this is a uh a new york times story on open ai this morning saying the maker chat gpt hopes to triple its revenue in the coming year jeez i read yesterday that they're back at more than 10 growth for chat gpt every month month over month so i mean you know that's that's not necessarily a stretch because I mean I assume there are going to be more people using Pro with better tools there are going to be more people on Codex running gazillions of API tokens yeah because it's not necessarily that they need more people using the tool they just need more usage of the tool yeah this may not be neuron fun but it's very interesting that VCs are hedging the foundation model race by funding both open ai and anthropic why not they both want money right nobody's gonna buy ford and chevy yeah sure go for it deep seek million tokens is cool uh this is interesting uh yeah they released glm5 yesterday i know you saw uh so it's an agent coding leap but they also added today that they're going to test its pricing power with a 30 spike in price for coding subscriptions right uh you know at this point all we have is the original price it was three years ago or 10 times that that's kind of the norm in ai right now there's there's not a big middle ground anthropics done a good job with that they got a hundred dollar range which is cool i always thought i would love a 50 i i thought you know what a 50 range that gave me a little access to some other things where you could at least try all the tools maybe limited uh do that to sell me on the 200 plan i can be sold yeah the 100 version is kind of like that but yeah 50 one would be nice so hey on main on the main are we talking about uh specifically the ms365 agents we talked about you're going to do a different kind of agent you're gonna yeah i think i'll i think i'll do Microsoft 365 agents and then we'll also talk about you know if you're not using Microsoft because I like to keep things as broad as possible you know some alternatives that you can use I'll talk about that I also tweeted about this this morning should I spin up here my my co-work and we can check this out sure Sure, sure, sure.

2:15:10Bryan Goode:Okay, so we've got a main. Are there any of these that don't interest you? Do you think meta's not that interesting today? That's fine. Around the horn's fine. Okay. The GLM-5 could maybe be a treats to try, actually, Grant. Oh, we covered that yesterday. That would be like if we wanted to deep dive on it. Okay. It seems like, so the interesting thing about that for people who are curious, So for both DeepSeek and GLM, these are models out of China. They're open weights, which means that any company who provides cloud services, meaning they give you access to their NVIDIA graphics cards over the cloud for you to run a model, can run the model weights, which means that they can sell access to it.

2:15:55Exactly. They can spin up an instance and sell access. So that doesn't mean that you have to use it through the Chinese cloud providers. You can use it through like an American provider instead, if you're in the America or the EU based cloud provider, if you're in the EU, you know, et cetera, et cetera. So both of those are useful for that. And you can use both of those models with what's called an API key through OpenRouter. And perhaps I'll talk about that tomorrow as well to explain if you wanted to use some of those models, how you do it.

2:16:28Bryan Goode:OpenRouter is a cool tool that might be worth a tutorial one day, Grant. because it's a really good way to go try those Chinese models, all of these that are coming from other companies. You can still use ChatGPT and Claude there if that's what you want to do, but you can also use these others, and they're in one place. I think that's a good idea. Yeah. Also, allegedly, I mean, Microsoft is a cloud provider as well, and we heard from Brian today that you can also pick models in there, So they might have GLM or something like DeepSeq on their servers. Yeah. You can check. I don't know if that would be in your co-pilot, but.

2:17:09Bryan Goode:Hey, I did something different today in pouring through research. And let me know what you think. If you don't want to use them, you don't have to. But one of the things is I often think that intelligent insights might be a little difficult to understand sometimes. And that like some of our readers, it might be a struggle for. Yeah. Just by not having, you know, PhD caliber, you know, knowledge on some of these subjects. So what I did was I took these and made them a little more simple. And I added this under it. I like that. Under each one. So we've got to think or not to think that throwing more thinking time at a model doesn't consistently make it better at understanding what a person believes or intends.

2:17:54Bryan Goode:And sometimes it actually makes performance worse. Why that matters? We can't assume more reasoning automatically fixes social intelligence. It can introduce new failure modes as well. Next, we have a really interesting one out of Stanford HAI. Stanford argues that as countries worry about control over AI and digital infrastructure, we may see new alliances among mid-sized nations built around shared compute data and deployment infrastructure. And that's a really cool idea. Yeah, that is cool. I like that a lot. And why it matters, the AI race isn't just who controls the models, but sometimes who controls the rails, you know, infrastructure and access.

2:18:41Bryan Goode:So that one kind of jumped out at me a little. Next, researchers used AI style simulation to argue a Dutch Roman era artifact may be an early example of a European blocking board game, basically reconstructing rules from the board layout. Why this matters, this is a neat case of AI being used for a hypothesis engine for history and archaeology, not just text and images. I thought it was kind of a neat one to toss in. Yeah. Another paper mapped where LLM decision-making differs from human behavior in repeated paper rock scissors and found some frontier models show surprisingly deep strategic patterns.

2:19:23Bryan Goode:Why it matters? Human-like isn't the same as smart, and we may need better benchmarks for strategy and adaptation, not just correctness.

2:19:35Bryan Goode:FoundRL describes a training pipeline for autonomous driving that learns from a vision language model's guidance, then distills that guidance into a real-time driving policy so it can run fast enough for the road. That is a self-learning, self-driving car. Which is kind of cool, like an onboard RL for an autonomous vehicle.

2:20:01Bryan Goode:Which, to me, seems just bonkers.

2:20:08Bryan Goode:Okay, Stanford AI warns medical records models. This felt pertinent. Medical records models can produce plausible patient timelines without giving well-calibrated risk probabilities. meaning they can look very convincing while being statistically unreliable it matters because in medicine plausible isn't good enough you need calibration validation and clear uncertainty yeah that's real but that was just me digging through stuff this morning yeah that's great that's great well folks you'll know what will be in the neuron tomorrow there'll be some other fun surprises but please open it anyway because that matters yes please don't don't assume you don't have to read it because inevitably like include like 30 other things so the interesting about the interesting thing you should know about how the neuron is made is that it's entirely possible it's 3 52 p.m central so it's 152 pacific for grant right now uh it's entirely possible that somewhere around six something major happens and it goes upside down and uh we wind up with poor grant sitting here frantically scrambling and writing an article while I am enjoying my dinner in St.

2:21:25Bryan Goode:Louis. I really need to make an open flaw for the neuron. Just actively respond and edit it in Beehive. That'll be something that I'll work on later. Yeah, it's just a preview. Speaking of previews. Go ahead. And can I bring up co-work on my computer? Yeah. But if you have anything else you want to hit before we move on, let's do it. No, I think that's okay. Yeah, I think all I was really going to say was that, you know, it's gone. It's okay. You have to get Peter on the channel. Check it out. If I remember, I'll pipe in. I think you were going to say we need to get Peter on the podcast. Is that what you're talking about?

2:22:05Bryan Goode:Yeah, I was going to say we've got to get Peter on the podcast. And that may take asking for help from the audience as well at some point. But we'd like to get him on here to talk about OpenClaw. I'd love to have him on. we have a couple other people we've reached out to lately that would be really cool. We've got a couple of cool ones we've already recorded that we're really excited about. Just neat stuff abound. All right. So I'm downloading that on my surface. Oh, work. The big news yesterday was that coworker is now available on windows as well as Mac, which is very generous of Anthropic to heed Corey's call to make it on windows.

2:22:47I was suspicious that perhaps ending part of their deal with Microsoft, that they couldn't release it on Windows. But it seems like everybody, they raised$30 billion today.

2:22:59Bryan Goode:It looks to me like they've just all opted not to. And I think it's because it's a team full of people who live in a Mac world and think the rest of the world resembles them without maybe looking at the data behind it. It is one. That is one interpretation. The other is that it's easier. It's a smaller sample size, so you can things, kind of like rolling something out to press. I get bouncing them out there. I'm more in the where the hell is Atlas is more what I'm thinking. Why can I not have ChatGPT's browser? Yeah. So this is the chat, or not ChatGPT, this is the Claude desktop version. So there's three options here.

2:23:39You can go to chat, you can go to co-work, or you can go to code. The one that anyone who is not a developer is going to want to use though, is this guy.

2:23:52So I'm open to take any requests in the chat for what it should build for us. If you have an idea, Cory, go for it. One idea I was thinking is the test that we were suggested earlier to make a local,

2:24:09Bryan Goode:what was it local landlord tool yeah you know one of these days have have these fancy new uh claude and open ai models uh rebuild cat doom let's do it i will do that as well okay so help me i was telling friends about cat doom last night to search news on local landlord tenant changes and updates that email me update a regular basis now put it all in a empty file

2:24:56Bryan Goode:minutes oh they got it for windows arm 64 i can run the arm version i love it i'm excited already So let's see here.

2:25:13The first thing it's doing is it's asking a series of clarifying questions. So what location should the scout for the landlord tenant law news? Simon, if you're still watching, you can give us your request. For now, I'm going to do California because that's where I'm based. What specific topics are the most important to you? Rent control and increases, eviction rule. Actually, why don't I just do all of the above?

2:25:39Bryan Goode:Yeah, that's a good call. I would imagine all of those are somewhat important. Yeah. It says, got it. California broad coverage across all landlord-tenant topics. Let me create this shortcut for you.

2:25:57creating a recurring shortcut that runs every 30 minutes shortcut must be what it's calling books and triggers in this instance

2:26:09Bryan Goode:i have the full picture let me log in

2:26:21okay

2:26:23Bryan Goode:i mean i don't think i feel bad about that that address being blocked that cat walked across my keyboard the sus address what's that

2:26:37so i can give it notes while it's working here right

2:26:41Bryan Goode:Oh my gosh. Some of those in developer mode, you can get around too. I like this because it also shows you the progress right here. So it shows you how far along it is. You can also see that it's created a series of five steps and you can kind of eyeball it and say, Hmm, you know, you're missing a couple of steps on this list or no, actually I wouldn't approach it that way. I would do it differently. And you can just chat with, you can tell it.

2:27:12a, you should do this differently. And then, you know, let it, let it actually.

2:27:18Bryan Goode:Kind of just ignored you. Yeah. Which model are we on? 4.6? This is Opus 4.6. So this is the one that came out literally a week ago.

2:27:30Okay. I'll, okay. So I just saw in the chat that Simon said he's in Philadelphia. So I'll say let's, let's also make a list of landlord. Or what are we calling this? We're calling this a scout. Let's also make a scout.

2:27:51So as you can see here, I'm not having to do any sort of like integrations. I'm not telling it how it's set up.

2:27:58Bryan Goode:I'm not giving it, you know, step-by-step instructions. The agent is doing a lot of that on it. this is why a lot of software stock companies have been selling off lately because people are like holy crap pretty soon we're not going to need software not really but that's like the idea that that's the fear oh my gosh anything else on this topic we want to hit oh no it's interesting to watch watch the market and how it reacts to ai i i can't help but wonder i i haven't checked it yet today but i have wondered if after a day of matt schumer's post being out and all over the internet and you know millions and millions of views how it reacts if it reacts the market doesn't always make a lot of sense the things that sometimes freak them out really aren't the things they should be freaked out about truth is they don't want to hear the things that they should be freaked out about sometimes uh but that's the the skeptic in me cynic yeah well that's that's the thing that about matt schumer's piece i think resonated with a lot of people is it's like hey this is a big deal and it's now no longer just a big deal amongst like early adopters but you need to start telling people you love that this is going to be a big deal and it's really going to change the world yeah you know it's i hesitate with that because it sounds nuts well it doesn't i don't think it sounds nuts but it's like what don't start off don't shout fire in a in a cinema you know what i mean yeah like let's let's maybe figure out what the right path forward is and push that as opposed to just saying like you gotta to freak out because the whole world is going to change it's like that's not helpful like that's that's that's what we're trying to do here is be helpful like you know for for three years dario amadei sam altman elon musk love him or hate him uh have all said this is going to be a massive impact we need some solution now and politicians don't want to hear it much for the It's a similar thing to public health that's really interesting.

2:30:19Bryan Goode:And I'm not going to talk COVID because I don't think that matters. But, you know, I used to cover – not I don't think it matters. Just pretend I didn't say any of that. I will say that an interesting thing about the public health space is that I used to cover this. and they would have these annual pandemic, like a training workshop, where they went in and pretended, you know, they'd get an envelope. You know, there's a bunch of medical professionals, and somebody's pretended to be the CDC. Somebody's pretended to be the local hospital, and people are pretending to be different elements that would come into play in that situation.

2:31:00Bryan Goode:And it was this big deal, and they'd get an envelope, and they'd open it, and it would be like disease X has appeared. It's coming from this region of the world. What do you do? And they role play it and it's really neat. But they do that and they talk to politicians and unless it's a five alarm fire they pretty much have always been ignored. And the same thing has happened to I say all of that to say the same thing has happened to Sam Altman, Dario Amadei Demis Aceves, Elon Musk Nobody's that interested in talking about preparing for what could be quite a shakeup if I put it gently. Yeah. Well, I think there's a couple of reasons for that.

2:31:41And, you know, we can, at this point we're all just vibing right now. We're just trying, right. So we could go, we can go a bit in depth to that. I do want to touch up on my screen, which we will in a minute, but there's a couple of reasons for that. One is short-term thinking runs our world and you know, it runs, it runs on the timeline is election cycles and the timeline. But is it today? Quarterly reporting, yeah. Now it's gotten even worse because it's like we're shipping new world-changing AI models once a week now or whatever it is. The acceleration is going to continue and it's going to feel really, really quite intense.

2:32:22The other thing is the structures, the power structures of which everything, our society is kind of like set up is a bit wobbly. And when that happens, a big change like this could kind of like, you know, if you think of it as like a house of cards, a big change like this could kind of like sweep the deck. If that happens, you have to have like, you have to rethink things from the ground up to a certain degree. Like, you know, I publish a blog of what I think that should look like, which it should look like kind of like everyone reorienting around education and learning and, you know, moving towards like a information-based economy where our goal is to just like maximize like learning and just like fun, essentially, as opposed to like money and number go up.

2:33:08But that's like a post-money society. We're nowhere near that yet. There yet. But you know, if this gets to the point where robots are literally doing all the work, agents are doing all the work, then at some point we have to consider some pretty radical ideas, because like we will have some sort of material abundance where all of the like prices of things that were previously scarce will just deflate like crazy. That sounds crazy to talk about, but Sam Altman has said it, and Dario has said it, and Elon has said it. yeah they've all been saying it so it's you know to the extent that any of those people are crazy which you know varies out that that's kind of where this is heading so you have to think

2:33:54Bryan Goode:from that standpoint that there is universal agreement that some significant level of shake-up is pretty much imminent as soon as 2026 now is what they're saying you know early 27 and that that is because i think i think dennis yesterday said that uh you know white collar jobs would be hit hard by the end of 26 to mid 27 he said 12 to 18 months and uh that people don't see that and think you know what should the solution be what is uh you know They've lobbed some out. Some of them have pitched universal basic income. Some of them have pitched – Elon's pitched universal high income. The other possibility is there have been discussions around things like perhaps data centers become public assets, and Americans – or, excuse me, humans, not just America – would own in that and the energy that it takes and would be paid a monthly dividend of sorts that might very well be substantial.

2:35:04Bryan Goode:Um, but you know, the truth is they're the only people tossing out ideas, whether you love them or hate them. Uh, everybody else is ignoring it. And, uh, that's, that's across the board. That's pretty big talking point though this year. Like the anger of AI, like there's a lot of people who are like, this is going to take our jobs. We don't like change. Like we need to stop that right now. and perhaps some of that in power would be useful like i said i'm down for two ai drops a year let's do it in the spring and let's do it in the fall and let's stop releasing stuff so it gives people time to like figure this out but uh yeah anyway nine design mentions some you know really concerned about the energy required to power all of this i would one thing i would say that if you haven't go watch uh we interviewed at microsoft ignite back in the fall scott guthrie who heads up uh Microsoft's, I'm going to screw this up.

2:36:01Bryan Goode:He's like executive VP of AI cloud and something else, but. He's basically in charge of their data center build outs amongst other things. Sure. Yeah. And he talked a lot about, you know, the amount of water, the amount of energy. We talked about those things directly, as well as the other key is, and kind of why I try not to freak out too much about the energy and even the environmental tax right now. Cause I mean, that's very much the thing I care about to be clear, but the catch is right now this is all really expensive to produce this is all technology that costs a fortune they're running into walls every day trying to build data centers they're being literally they have all of the motivation and all of the incentive in the world to find ways to make it cheaper to make it faster to make it cleaner because it's going to save them money it's going to allow them to produce more it's going to allow them to sell more which makes them money and i think the fact is we're just still really really early and that's going to take a few years just like when automobiles came out you didn't get 40 miles to the gallon in the 1930s you got four you know and over you know over time you know as the technology develops uh but i think that the demand for the ability to serve ai affordably at scale is is big right now uh that's my hope You know, I mean, it's not perfect, but I just thought I'd share that.

2:37:23Bryan Goode:Grant, how's it looking here? What do we got? Yeah, so this is basically what it put together. It created a scout prompt, which is basically the rules for the shortcut. It's used these task descriptions for recurring shortcut options. And then it created both a Philly landlord tenant update and a California landlord update. And then I think what I'm going to do, it did tell me that the automated set schedule task tool wasn't available in this session. So I'm going to say like, we set this task via a skill and or a hook. Use web search if you need help. Look at anthropic.

2:38:11if i was being more specific i would give it the exact context so i would actually like go and go to the website copy and paste the docs or the link in and then do it that way the first thing

2:38:21Bryan Goode:i would say about this being that it's the first time i've like watched co-work in person i just downloaded it i have it now but uh cool the one thing i would say is that it's super clean interface looks really nice yeah i'd also argue it's a better tan than neuron tan if you look it's the exact same background we're using just in a lighter tan you gotta be careful they're very well they're in lawsuit happy they see or they were threatening uh open claw right so we switch to their tan they're gonna be like oh excuse me you used our tan color that we trademarked even though we stole 7 million books to produce our model we stole the court ruled such that's not me saying it that is court ruling that said that yes yes yes pirated books grant is like not slander that's why that's why I think it's so funny I think it's so funny that like any AI company would enforce any sort of trademark because it's like you literally like all of you rob each other every time you do anything things yeah like bets are off okay so i see cory just dropped off here i don't know if that was intentional i was coughing and i may not be done okay i'm okay if i disappear again my apologies i got a house full of sick people i'm trying to trying to not be them yeah so now it's basically read the SDK, which is the software development kit, which explains like it's the docs for how and the CLI reference to see how we can wire it up as a run.

2:40:02It's explaining what it learned here, which it learned hooks are event driven, triggered by tool use, session start end, et cetera. They don't have a cron timer mechanism. So hooks alone won't schedule it. But in the Claude code CLI mode is designed exactly for automation. You can call the P your prompt non-interactively from a cron job or a launch agent and it runs the full quad agent with tools um web search file rights etc the best approach is a shell script that calls claud cli in headless mode triggered by a cron every 30 minutes now it's gonna go out and build that hell yeah i love it this this is why software stops are selling along have you Have you downloaded the Codex app yet too?

2:40:48Yes, let me show that. Up right here.

2:40:54Bryan Goode:I'm guessing that one will follow soon to Windows. Is Atlas on Windows yet? No. Atlas still ain't on Windows. There's a lot of AI tools that are still pretty Mac only, which is why I'm so ragey about it and screaming into the ether and tagging all of these giant billion-dollar companies. Yeah. Like, excuse me, sir. if you want to be featured in the neuron you must make a pc version thank you yes we need both thank you okay so early in the stream for anyone who stayed this long you're a real one thank you um we said or i mentioned that i used the codex app over the weekend to make a then my literal prompt here first person olympic ski jumping simulator in 3d so i don't know if anyone in the chat was pc gaming in the early 2000s but there was a really sick game i think it was called like winter race 3d or something that was super fun and you can do like you could race with you can race with snow machines you can ski you can snowboard all of this stuff that's really fun and so i wanted to recreate that because i was getting the childhood itch watching the winter's olympics and so i go ahead and put this together this is in the codex app which is openai's version of not necessarily co-work, but kind of co-work for coding.

2:42:15And then it gave me this thing that was like a local host. And oops. Uh-oh. Chrome wants me to do something. Cloud wants me to do something.

2:42:23Bryan Goode:Mine made me log back in a minute ago, too. It might be a, hey, we're new. We're getting glitches out still. I wouldn't. Yeah. But then it wanted me to do this server thing. And I was like, that's confusing. Can't you just write it so we can run it as a local executable or something simpler? And then basically it did that. Now what I can do is I can go to B-Jump simulator and it automatically makes it for me. And this is actually like pretty sick. I'm still working on some bugs for it, but... Oh, crashed out. Boom, down we are. But like this was made with just me saying exactly what I said with giving it some feedback.

2:43:08I was able to create a local game I could just play in my browser. Fast. Yeah.

2:43:16Bryan Goode:Whoa. It's really hard to land this. Not to the Olympians, you know? They're the real ones. I'll be honest. I wouldn't make it down the hill just standing there still. Yeah. Yeah, I'm just trying to see if I can, like, get it right. You're supposed to...

2:43:42I told it, well actually some of my last notes with Codex, I was like, can we make this like any easier? Can we give it feedback, like how to land properly? Because it's like, it's really hard actually.

2:43:53Bryan Goode:Go learn physics. Yeah. What was the last thing I said to it? Let's see. It said, oh yeah, I took some screenshots. So I was like some screenshots explaining what I mean. Also the ramp is reversed, which you'll see in one of the screenshots. It should be pointing the opposite direction inverted. And then I sent some screenshots like related to that. So I'll say like, hey, it's still hard to land properly. Any suggestions how we

2:44:31Bryan Goode:Oh my God. I just remembered the Pete quote I wanted to share with you. That's why we should share this. It was, I just prompted it into existence. And I was like, wow.

2:44:48I'll provide some context here. So here you can see, like, I've got the chat window component. Then over here, I've got all of the code. all the code that's running it and then it shows you the folder structure of that as well so you can like easily very nice between that's cool that's a thing that a lot of them have done a

2:45:07Bryan Goode:bad job of and it's one of the things that makes me love v0 is that you got the full folder structure right there it's not just here's some text in a chat go put it in a doc yourself yeah yeah yeah it's so much better and then there's other things you could do right like you can do automations which is kind of some one of the other things we were talking about today This is Codex's version of that. And then there's also skills. You can create skills or connect to it. I really want to actually connect all of these. I connected all of them on my ChatGPT account just to test it out.

2:45:493D Olympic simulator. It's working on it. Oh, my God. What?

2:45:58Bryan Goode:I just realized what time it was. Sorry. Well, we can wrap up here. This is just a demo. We're cool. We're cool for a few. Okay. Anyone have any other questions, anything you want to see? Just like, I know you're all probably trying to process all of this stuff. Because are we? You know, that's why these streams are helpful. It has been quite the... Yeah. Yeah. And we're trying our our darndest to get to where we can do a couple of these a month that are just Grant and I playing with some tools. The truth is we have a lot of cool guests come our way and and they have things to say that are super valuable.

2:46:41Bryan Goode:Some of them that aren't going to get a ton of views, maybe even that we think are important. but we're trying to kind of bundle those into some lives but ensure that we've got a chance or two a month to just come on here and do exactly what we've done for the last two and a half hours which is just try the stuff, answer questions, do things, build stuff, whatever. That reminds me, let's make the game do. Oh, that's cool. I love the loading icon there. Miss it

2:47:20Let's make the game Doom starring cats

2:47:25Bryan Goode:Oh yeah I have a feeling Grant it's going to look a bit cooler oh you do have 5-3 Oh wait oh no no don't don't convert the It's converting the ski game oh god Yeah Don't convert the make a new game new folder you know what's funny is uh if it can't do that amazing button grant yeah this is cool oh my god i can't if i had a dollar for every time i've sat and watched something and cussed while it happened in a oh understood i'll restore the original ski jump files like you better not have messed up my ski game bro

2:48:14I see you taking over in the chat. That's like,

2:48:19Bryan Goode:I'm a, yeah, I'm just chatting a little bit here. Also there's, you know, a couple other tricks here. There's default permissions. You can give a full access to the default local. You can do it in the cloud master. You can choose what branch. You can create a new branch to mess with it. That's kind of sick, Grant. I actually really like, this is probably my favorite of the coding apps that I've used. Really? Besides Cloud Code, like I do think Cloud Code is really good. But you can run Cloud Code in, you know, lots of different places. So that's on your preference. Yeah. And truth is, I like all of these.

2:49:06Bryan Goode:I'm just excited to live in a time where we have so many awesome tools that we have the ability to nitpick a button as a reason we would use a different one. Yeah, you're right. No, it's pretty crazy. We should be so lucky, you know. Yeah. I think what's funny. It's really awesome. I think the other thing that's kind of funny about this is like, okay, so all these AI companies, right? They're producing, it's supposed to take everyone's jobs. That's their business model. Yeah. $15 trillion is like labor market, like total addressable market size. But the very first thing that making obsolete and irrelevant is software engineering themselves.

2:49:45So they're going to automate themselves first. And they're speeding up. Yeah. So it's like, anyone can make software now. like like that's that's actually the coolest thing about all of this is anyone can truly make software now that is a superpower that was not true before and like you know how they used to say like oh kids should all kids should go to school to learn how to code it's like now all kids can code like what happens when all kids can code yeah a lot of exciting things can happen i need to show off my toy on here one week when is that the thing i built

2:50:22Bryan Goode:and v0 yeah i should have to do this sometime it's uh are you still working on it or what happened to it i got distracted by claude bot you should have claude bot help you with it the truth is it would be done uh yeah i wonder maybe what i do is no i can't download it easily because it's connected to a lot of the things i wonder if it could control my v0 another crazy thing is that it could prompt the hell out of an app better than i could it would probably listen to that better than it understands me man it'd be great if i could do the can hear that i'll pay for the pro right now after i've been reading a lot about claude code and codex lately and it's interesting because like i follow a lot of people and they're both kind of equally faithful to the one that they love.

2:51:22Bryan Goode:And something I will say that some seem to acknowledge is that Claude was really, really good and fast. OpenAI, that Codex was more likely to get it 100 % right, but it would take a long time for that to happen. And that made me think that that explains why maybe Claude Code is the tool for the developers and that maybe Codex is a tool for a vibe coder. You know, where that little bit of extra accuracy in exchange for more of your time might be worthwhile. Just a thought. And that's coming from not a place of experience, just more an angle of kind of what I keep seeing about the two. But, you know, I'd never hear anyone say a bad word about either of them, frankly.

2:52:11I think the way that I work like the Gavita is I like being able to be hands-on with flawed code like I like to be able to go back and forth with it a lot at least for the projects that I'm working on I think for stuff that I want done but don't want to actively participate on I would

2:52:32Bryan Goode:use codex yeah you know a thing you all might not know about us we're down to 20 people it's a little tight but uh a thing you might not know about us is that just as far as kind of how we each came into our ai journey is is grant has been a a you know strong claude user from day one and is is is a freaking master with it and i have been more chat gpt since the beginning we both have both we both use both but as far as if i was going to say this is my daily driver i i still hang in at gpt most of the time uh just because i find it to be a good swiss army knife for me if that makes sense and uh so we've tested a time or two using one another's tools and have have kind of like i was really impressed with grant's claude compared to mine and i don't know if that's just because i think it's context grant i think it's the memories heavily built out but it's kind of you've you've you've massaged them over time like god knows how many hours we've logged in those how many hours of claude do you think you have grant a lot i imagine i'll bet it's a lot i have an obscene amount in chat gpt too i have a good amount in claude but it's definitely a significantly lesser amount uh four six impresses the hell out of me though you you you saying hey let's do the was it a slideshow we did grant we did a spreadsheet and a slideshow spreadsheet was cool slideshow was sick okay this is this is cat doom from codex cat doom got an upgrade it's hard to control um the controls are inverted from what i from what i can see here there's no way to shoot and this is first prompt right yeah doom cataclysm face left okay now you can shoot okay tuna cans frags we should be doing this with the exact prompt you've used in ai studio too let's do it let's do that on a separate stream another day yeah not today i was just thinking it would be really cool to be able to kind of cat doom the neuron benchmark this this is cool because this is 3d none of the other other ones were like kind of like generated 3d but this is like there's 3d models in here yeah and you could absolutely get it to to skin those to change their shape you were malt look it's hard the controls are inverted yeah you need to just tell it hey un-invert them.

2:55:15Yeah, I'll give it that feedback.

2:55:17Bryan Goode:I wonder if that's what's going on with ski jump.

2:55:25Bryan Goode:Any chance the controls are inverted and you're actually trying to forward flip down a hill? Oh. I'll ask. All right, should we wrap it up? Yeah, I think we wrap it up. Everyone, thank you so much. It's been a long one today. These keep getting a little longer, and I'm okay with that. As long as, you know, some people are here hanging out and you're enjoying it, everybody's learning something, and if you would, please take just a minute to like and subscribe to the channel. It means a lot. We've got so many cool interviews that we've done, and I hope you'll go check them out. Eve Bodnia dropped this week.

2:56:00Bryan Goode:She's the founder of Logical Intelligence, which Jan LeCun heads up. So we talked to her about energy-based models, and it's an amazing conversation. We've had a number of really strong, true innovators who are trying to break paradigms right now and lately. And we're pushing to get more of those. You know, you don't get one every week, but by golly, George, if they're out there, we want to know about them. RVT, Kat, you're here like every week. Like, shout out. But you need to engage more in the chat. Let us know you're here. I didn't even realize you were here. That's funny. Oh, my gosh. Well, hey, yeah, thanks.

2:56:39Bryan Goode:Please like, subscribe. Go check out the newsletter at theneuron.ai. Sign up for it. It's growing every day. I think we're creeping up on 650 ,000, Grant. I noticed it's been climbing heavy this last week or so. But that's it from us. And we'll see you back next time. Thanks for helping us plan the Neuron for tomorrow. That was a lot of fun. And always, you know, reach out. Do you have any ideas, thoughts, problems you want to talk through or ideas for future live streams, channels? We want to know about them and we appreciate your time. And I'm just keep talking because I'm watching Grant Giggle while he types and it's cracking me up a little.

2:57:22OK, I'm hearing you keep working in the other time.

2:57:25Bryan Goode:I feel at nine design. Yeah, we will put together a blog for this, too, so you can watch like the highlights. Yeah, I'm going to work on that after this because that'll be part of the inspiration for the main story. with some extra stuff put in there. We're also now, they're going to be several weeks late, but we're uploading these to Spotify and Apple as well. So, you know, if you're ever in the car, wherever, it's there. Yeah. Thanks again, everybody. And thanks to Microsoft for bringing Brian in. I appreciate that. And shout out to Microsoft. Definitely check out Copilot Studio and Agent 365.

2:57:59Bryan Goode:Absolutely. And on that note, farewell for now, humans. Bye-bye!