Replay, recorded Monday, September 14

Fable 5.1 vs GPT-6 Astra
20 AI Power Users Compare Notes

Which one should you actually be using? Two weeks after both launched, a room full of e-commerce operators, developers and AI power users compared notes. No benchmarks, just what is working in their businesses right now.

Recorded Monday, September 1477 minFull transcript below

Get invited to the next one Watch on YouTube

What the Room Figured Out

Fable vs Astra turned out to be the wrong question. The people getting results win on workflow: effort settings, harness design, clean context, and a person approving every change.

1
Pick the model by cost per task, not the leaderboard
The room’s read: Astra on High beat Fable on Ultra for security audits and code quality, and Fable still wins some jobs. Test both on the task you actually run and compare what each one costs to get it done.
2
Plan high, execute lower
One member ran Astra on Ultra and burned a week of tokens in a day. Max effort also overthinks, like writing a test for every tiny thing. Plan with the top model on the top effort setting, then do the work on High or below.
3
The harness matters more than the model
Shubhash walked through six steps: task, prompt and context, tools, model, retry loop, done. The model is the engine and the harness is the rest of the car. Changing only the harness can move results many times over.
4
Real workflows, not demos
Codex cut a raw six-minute talking-head video to five with branded motion graphics in about 17 minutes. A long-running Astra task cloned three paid Shopify apps. Codex plus Blender over MCP rebuilt a 3D site animation that cost around $30K to hand-build.
5
A person approves every change
The multi-marketplace repricing flow has the AI audit listings across Walmart, Best Buy and Shopify, write suggestions to a Google Sheet, and only push to each marketplace after a person signs off.
6
Keep context clean and cheap
voice.md files for a consistent brand voice, auditing past chats for wasted tokens, cheap models for cheap tasks (Hermes + OpenRouter), decision logs so agents stop repeating context, and loading MCPs only when a task needs them.

Chapters

Links Mentioned

Copy the Full Transcript

The whole call in one click, split by chapter with timestamps. Paste it into Claude or ChatGPT and ask it what applies to your business. It is a machine transcript with no speaker names, so some names and terms may be misheard.

Full transcript

Fable 5.1 vs GPT-6 Astra: AI Agent Hackathon, full transcript
Recorded Monday, September 14, 2026. Replay: https://youtu.be/z6qy0zQ_nqg
Machine transcription with no speaker labels, so some names and terms may be misheard. Timestamps match the YouTube video.

0:00 Welcome + Alibaba CoCreate recap

[0:00] Good, good, good to see you, Ed, Can, awesome. Can, tell us, while I get my notes prepped here, I think of a fun thing would be to, if you want to help us talk about the event we were just at, Co-Create, sponsored by Alibaba, we were at last week. So why don't you tell the group a little bit about that while I get my notes prepped here. Yeah, yeah, yeah, do you want me to talk about the topic I shared there? Or just have a general idea of what Co-Create looked like, yeah. Yeah, well, so maybe can I say an intro of what Co-Create by Alibaba was? And then maybe a little bit, some of the things you learn and what's the cool stuff that's happening in that space. Maybe five minutes, just to kind of tell people what's up. Yeah, yeah, great.

[0:52] Actually, the Co-Create is a very big conference hosted by Alibaba. They told me this time they have about 20,000 attendees. So well, it's a big room. You remember we attend the VIP party there. It doesn't look like a VIP party. Yeah, more than 300 attendees in the room in just a VIP party. So it seems like people are so passionate about all this AI and e-commerce thing. Yeah, because this is an event by Alibaba. They want to promote one about their AI agent actual work. And two, more about to group the people who are in the e-commerce area, yeah, to discuss the next things. I think this year, people pay a lot more attention to AI agent. Before, it's just like AI bot, right? How to use a lot more like a talking ball. How to use AI to give inspiration for our daily lives, to ask question.

[1:49] But this year, it seems that everyone want to use agent. Want to include agent in our life, in our business to make our business more in a automated way and enjoy more efficiency. I believe, although now actual work is not perfect right now, right? Like many agent, but it do open our mind that maybe next year, yeah, all this AI agent can be very around and they can handle a lot of our day life and business. And I really imagine, yeah, and I have a big expectation of all this AI, how they can create new business models. Yeah, like from internet era versus the previous era, the business model have been totally strange. And I believe in the AI era, it can be another level. Yeah, something really new.

[2:44] And I mean, you also bring something good ideas and good things along this journey I believe we will, yeah. See a lot of new things happening. Yeah. Love it. Yeah, it was a fun event, a little bit different than ones I usually go to. I'm usually very specific like operator. Here's how I SEO maximize my listing on TikTok and Amazon and this was a little bit more broad audience, but it was a good conference. Cool guys, we already have 20 people in here already, which is very exciting. It looks like we have a pretty good participation this week. It looks like I shared a link, unfortunately, that named half of you guys Andrew Erickson. There is really only the one real Andrew Erickson. And so if you don't mind changing your name makes conversations a little bit easier if you change your name to your preferred name. Otherwise, because it's easier for us all kind of to know and see each other.

[3:45] But okay, I put together a few slides, of course, Claude and Fable and Astra put together. Somebody put Andrew Erickson V2, nice. I put together a few slides, I thought it'd be kind of good little, like here's a quick little intro. It's the other Andrew, that's the V2 Andrew, awesome. Quick little intro just to kind of set the tone here so everyone kind of is on the same page because I know not everyone is obsessed with all the AI news like I am. So I'm going to share my screen here and kind of show you guys a few things that happened just in the last three weeks, last two weeks, in September. September has been a very busy month so far. So I am going to share my screen here. And then after that, we're going to have an open discussion.

[4:40] I see John Aspinall is here. He's always a fun lively chat, so I'm sure he'll have some opinions on these things. And I would love to just get, if anybody has some, the thing I love here is when all these smart people, I pull together lots and lots of smart people. If you're here, you're an awesome Ecom AI smart person. If you have some insight or something you've found that works really well or some new kind of insight into what is happening, I'd love for you guys to share that after the talk here. But let me get my share screen.

5:21 Two weeks that changed AI: Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra

[5:20] Okay, can you guys see my little slide deck here? You got two robots, Astra 6, Fable 5. Perfect, yes, perfect. Okay, cool, I just want, this is just like a five, ten minute thing just to kind of put you guys on the same page. So, yeah, of course, here's my fun graphic. Of course, this was created by Astra, by ChatGPT. Because their new image model is incredible. It is breaking all of the leaderboards now, so it's pretty cool to see. Okay, here's what's happened in the last two weeks. Claude Fable 5.1 came out, which of course is just kind of an upgrade from Fable 5.0, which came out August at some point, mid-August. It's a little bit more refinement of the 5.0, but the Fable model is incredible.

[6:20] I'm like, I love the Fable model. 5.1 is a little bit of just an update onto the basic Fable model. A Gemini, one thing, I'm not seeing a lot of press on this, but Google shipped their own Gemini 3.8 flash model. This is a really good model. It's very smart. It's not smarter than the other models, but it's incredibly fast and cheap. So, I think for a lot of tasks, I think it's actually something worth looking at, obviously, depending on what you want to do. And of course, here's our big one is that OpenAI shipped out GPT-6 model. And then, this was really cool. I have a math background, and so this is big news.

7:08 OpenAI's Navier-Stokes proof: 10,000 agents, 88 hours, $15M in tokens

[7:07] Last week, OpenAI published a proof to the Navier-Stokes equation. It doesn't really matter what that is. It's some crazy math theorem that we've been trying to prove for almost 100 years, and OpenAI published a proof to it. There's a little bit of controversy of how they got the answer to that proof, but we do have a proof, and it was pretty remarkable. This is kind of showing that we're ushering into a new realm where the AIs are, they were good, and now they're great. They seem to be superhuman in a lot of spaces now. And then, lastly, just a few days ago, OpenAI shipped out their image 2.5 model, which is pretty amazing. So, yeah, I'm gonna kind of go through these ones really quick. So yeah, the big thing that Azure's really good at, GPT-6 is really good at, is the driving your browser and your desktop.

[8:01] It's incredibly good at doing your desktop now. It is price points there, and it's really good at in-depth reasoning. Here it is. Okay, so the Navier-Stokes equation, it's like an equation of how fluids move. And when they move, if they create a singularity, cuz the way the equations work, it looks like there's little singularities everywhere. And so they had this out as a problem that was almost, yes, 200 years old. But it was a millennium prize question, which is a big deal. These are really important questions. At the turn of the millennium, 25 years ago, a bunch of mathematicians got together and they set out nine problems that they are really excited, they really want proved. And they set a million dollar prize on each one of those nine problems, and OpenAI just solved one of them. So the way they solved it, it's a part where I'm like, this is crazy, cuz people think, do these things actually scale?

[8:59] Can you actually get better? Are we gonna hit some ceiling? And I think the answer is, hell no, we're hitting a ceiling anytime soon. Because what they did is they took their model, it actually wasn't Astra, it was actually their secret model that's behind closed doors. And they ran 10,000 agents at once. They ran it for 88 hours straight, so four days basically. And then it took 17 hours to write the summary, and then it won the millennium prize. So of the nine equations, only one has ever been solved, and this is the second one. So it's just a big deal in the space. Now, does it really affect us? Not really, but it's just kind of a testament to how incredible these models are. GBT Image 2.5, it's awesome. It's, there's a, there is a, This is my favorite site. I like to use to compare models, cuz all these things are always changing constantly.

[10:00] I'll drop this in the chat. Where's the chat button? Okay. Shubhash asks, is OpenAI getting the million dollars? Yes, I think that's kind of the, it's not fully accepted, proven, fully whatever. They haven't officially designated a winner. But everyone's kind of like, okay, yeah, minus the actual legal stuff. Yes, it won, and I guess OpenAI is going to make the million dollars. Ironically, they spent $15 million in tokens. At least that's the retail price of the tokens. $15 million getting that million dollar prize, so not a great ROI. But of course, I think the press is well worth that $15 million.

10:48 Arena.ai head-to-head: Fable vs. Astra

[10:48] Here, I dropped in the chat here a, there's some good chat messages here. I dropped in the chat one of my favorite sites here, arena.ai. I love this site that helps, you can kind of just compare side by side the different models out there, and you can see that for agentic use, Claude and GPT Astra are both side by side. So they're technically, according to the way they measure this here, this is a human, this is a side by side human comparison, rather than some benchmark thing. I think it's a good way of measuring it, because you just give two answers. A person shows up on the site, and they get two answers, and they have to compare the two, which one they like better. In my mind, this is just a good way of just having a regular person using this. Which one do they like better? And this is the answer.

[11:40] So this is the agentic. We're technically fables beating Astra, but they're so incredibly close. And then also here, I like this one a lot. If you go to image, text to image, and then product and branding, commercial. This to me is a really important one that we all care about, since we're in the product category, and here's text to image. And so you can see what the images, who's winning here. I dropped the link in the chat. And then just GPT is just crushing, absolutely crushing this one. So if you, Grok is doing well, Gemini's okay, and there's some. Microsoft, I guess, is doing pretty well too. But I would definitely highly suggest going with GPT image 2.5. Again, that model just came out like four days ago. Okay. Ritu's coming in, awesome.

12:38 Where Fable wins, cost per task, and which model for what

[12:38] Okay, where is Fable really good at? Fable is still really good at coding. If you're doing long run coding projects, this works really well. There's the price point. The real cost is a little bit cheaper. The way the price points work, it kind of depends. You want to measure per task rather than the actual input output, because there's all sorts of weird little nuances around caching. And how long it takes, and how many tokens they use, and stuff like that. The important thing is just kind of look at how many, what the cost is per task. And we can get into the whole idea of how do we, if you care about cost, how we can maximize those things. There's all sorts of fun ways where you can have Fable do one thing and farm it out to Sonnet or to the GPT models. And they're a little bit faster, a little bit cheaper to do it that way. But that's it on price and then real agents. Those are the one I just showed you guys.

[13:31] But yeah, technically, Fable is beating Astra right now by a tick. But they're so close, it's just they're neck and neck. Okay, here you go. So the cost, yeah, so in general, if you're looking at cost per task, Fable is about double GPT right now, even a little bit more than double. Again, these tasks are kind of defined in a special way. The way you do your own task will be very different than how you, these things are measuring it. Let's see, and then again, last mention to Gemini 3.8. It is very smart, very fast, and very cheap. So it's also, if you're kind of looking for a model that isn't just the frontier, most expensive thing possible. Gemini, just kind of an auto mention there too. So there you go. There is the last slide. If you're doing images, I highly suggest you use images and listings, definitely GPT. If you're doing code clots, probably still winning, although they're both, they both complement each other really well.

[14:32] And then if you're doing volume, speed, and cheapness, you want a frontier model, Gemini is really good. That's what I got. Let's open it up. I'd love to hear from the rest of our smart, capable Ecom AI people here.

14:49 GPT-6 Astra on Ultra burned a week of tokens in a day

[14:49] What's everyone thinking of these models? Well, I'm really liking a GPT-6 Astra. I'm using it more than Claude, although I did make one mistake. I set it on a task on Ultra, and it killed the weekly tokens in the day. So I now just set it to high only, and it seems to last a lot longer. But the output has been really good, especially with the security audit, because obviously it doesn't drop the model down. Fable drops me down to 4.8, misses stuff. Astra picks up loads of stuff. It's great. I really like it. What were you using the Ultra code or the high end for in Astra that sucked up all the tokens? I was working on a CMS system, and it had about 15 different things it wanted to implement in a plan.

[15:48] I thought, "Oh, do you know what? I'm going to stick on Ultra and just let it go ahead." No, it massacred the token so quickly. I've now had it do a similar plan on high, and I've still got limit. Yeah, I typically would always put it on top model, top effort level for planning. And developing the plan, because nine times out of 10, from my experience, you don't need that ultra high effort and ultra high model to actually execute. So that's another way to think of it too. Yeah, it was an expensive lesson for a day, because I normally now get Fable to do the plan. I then have Fable and Astra chat together about the plan and work out the best way. And then one of them does the work, and the other one checks the work. But Astra, I find, catches quite a lot of bugs, and also is really good at improving the code of it, just the speed of it.

16:53 Max effort overthinks: plan high, execute lower

[16:53] I was going to say, there's also an issue if you use the max level effort, max level thinking, and they can sometimes overthink, especially if you're doing knowledge work or coding. For example, if you're writing test at the highest, they might literally test every single little thing when you really don't need that. So that's what I would caution against when you're using the max effort. I haven't specifically looked at Astra's benchmarks, but I believe that was true for Fable. And it also takes forever. So I don't know if anyone's ever put on like ultra or... It takes forever because it's a lot of effort. And sometimes I almost, I was reading something that talked about how GPT-6 Astra on lower medium is comparable to like Soul56 on high or ultra high, right? So I would say, you know, right person, right job, right tool, where it's like, unless you're doing, if you're doing it like what I do, like a lot of content, knowledge based kind of things, GPT-6 Astra on even low does laps around five, six soul.

[18:10] So I think it's really important to understand like what you're doing and what you need it for. And if it's doing like a lot of coding and a lot of checking, because nine times out of 10, I talk to people, they do not need to be on ultra code or X code or whatever it is, because it's just, it's going to take forever and it's just going to suck up a ton of usage. Yeah. I found the speed is, is frustrating. Cause I, I have like six sessions going at, or even more, but try to get six sessions moving. But at some point I'm like, Oh man, I'm just trying to update one little thing. And it spins for 14 minutes updating that one little thing. Like just, I don't need to super duper. I don't need the council of PhDs chime in on this, like updating MOQs on, on something, you know, Does anyone, does anyone actually make, um,

19:00 Codex edited a raw talking-head video with motion graphics

[18:59] uh, cause this is what I found and I literally just did it just now. Um, does anyone do like video content, like either demo videos or talking head videos or just marketing videos? Because, um, in the codex, I strongly believe that the codex app is the best app, um, compared to like Claude or anything else because the features inside of it are just insane. So I just shot a video talking about how you could help local businesses use AI just nonsense, right? Not nonsense is good, but it was like six minutes and I didn't want to edit it. So I just dragged and dropped it into a codex chat. And I said, here's my branding on my website, go and edit the video for me, remove out the ums, ahs, whatever. And then layer on like branded react components, animations, charts, and graphs. It did it. And it chopped it from a six minute video. It's like a five minute video with full blown editing, full blown animations. So it's like, if you think outside the box of, oh, it's just for coding and making an app or doing a tool, like you can put it on autopilot for your content, which is pretty cool. So I don't know if anyone's doing video content, John, it's a good shout out. So I I've used, uh, so like a week or two ago, I used Opus five for creating a couple of videos and I was actually taking one of these calls and chopping it up into shorts and it did a pretty good job. I was very impressed with it. And I'm, this is the kind of the fun thing where we're finding that jagged edge of like, what is, what are these things really good at that we can step up as, you know, as business owners, like, Hey, I never was able to make shorts, but now I can do this. And now I can do that. Like, where are these new, uh, capabilities enabling us? Can you give us a, like, pretend we all know nothing? Uh, give us a quick, quick basis. I'm like, Hey, I have codex installed. I want to, I want to take a couple of these pictures or a couple of these videos and make them into something cool. Yeah. So I do, you're gonna do something cool. Okay. Thank you, John.

[20:56] So I literally said, Hey, can you edit my video and my downloads folder? So I think it's important to know the tool you're in, right? So I don't have to drag and drop my video here and waste time. It's already in my downloads folder. So I'm telling codex last video, my downloads folder, really good cuts with visual, uh, uh, motion visual graphics for me. And in eight minutes, it gave me a video and the video was like, okay. So then I was like, okay, well, can you, I don't want my last name at the top of the video. It's not really, you know, I have a point, but make it more visually aesthetic on brand. And I also want you to create animated, um, react visuals. So react is like, if anyone doesn't know what, like the components, like, so Claude, for example, can't do visual, like actual images, but it can use react components to put together a visual, which is really basic rudimentary kind of like shapes and stuff. So it took nine more minutes. So if I gave it that prompt initially, it would only be nine minutes, but just to give you some perspective, this was the video that I shot here, where it's just me and a codex screen and all this kind of stuff. And it's, you know, going through it. And then if I go to the newly edited one, see this, I would never have the wherewithal to do motion graphics popping up from idea to AI service over here. And if I just scrub through it, you see how it took my video and like shrunk it down, but then put the positioning over here at room. It basically moved my head from being in the lower left hand side to being in the little dedicated box over here. And then even this part right here, I missed the trends of this, like, so it does these visual, I'm sorry, one second.

[22:49] It does all these transitions like this. So if I was to actually do this myself, I wouldn't, number one, but if I was to have someone else do it, they would charge me an arm and a leg just to do this at scale. So that even something like this, like it knew like the context I was talking about someone selling custom printed shirts on Facebook, right? In Facebook marketplace, it pulled out all the context and gave me this really well done edited video that I could just take and scale. So that's like one way that I'm using it. The other way that I'm using it over

23:23 A TikTok cover skill that reads the video

[23:23] here is I'm scaling my TikTok content because, you know, it's where all the kids are. So I made a skill over here that's a TikTok cover because what I found was that when you upload a TikTok, you can do a custom cover thumbnail for your feed. But what happens is you have to like pick a scene in the video and then put text on it. But whenever I picked one, it was always like a weird like motion of me doing something. So I was like, wait, let me just make this skill where I give it the URL. It's going to read the context of the video, like watch it, get the transcript and then build me out the thumbnail that I can go and add. So now at scale, I can then take it and let me spell TikTok, right? And if I look at my profile, now I'm able to have instead of like weird kind of things like this. Now they're a little bit more tailored and buttoned up. So when someone takes a look at it. So just some cool stuff that I'm doing because for me, Codex is just like it runs everything, right? Like I pay for $200 Claude code, $200 Codex, $300 Grok. Grok's a whole different story, right? But for Codex, fundamentally with GBTSixastra, it could do anything you need. I love it, man. That's good. Good share there. And I think a lesson here for the rest of us is that there are some really cool videos coming out, video capabilities. You don't necessarily need to have a really expensive model or some special software for it outside of just these AIs are already like good, already pretty good at it.

24:59 No excuse for a bad website in 2026

[24:59] But like you had said, sorry, you had said Opus, right? So I always look at it like, I don't want to pay for more tools. So how can I leverage the crap that I pay for now to replace that? So like, I never wanted to pay three or $4,000 for a website. I never wanted to. I didn't want to go and buy it. So I just have AI make it for me. And as it gets better, you could, so like there's no excuse why your brand shouldn't have a website in 2026. In September of 2026, there's no excuse why you shouldn't have a website and it shouldn't be awesome. Like your business, your agency, like if your website sucks, it's all on you because you have it all at your fingertips to do it awesome. I'm going to what you just said right there. I'm going to clip that out. I'm going to have I'm going to have Astrid clip that out and make it a short for you. We'll see how it works.

25:47 Cloning Opus Clip + rebuilding the Igloo site with Blender

[25:46] Yeah, go ahead. Dig. Yeah. And talk about video and website. I was working, doing a day last week for a friend to set up their AI. They were spending a fortune on Opus clip, you know, that takes their videos, cuts them down for shorts, for Instagram. In a couple of hours, we made pretty much a usable copy of Opus clip. And I'm going back in there on Wednesday to refine it as I think you can make virtually a 99% exact copy of Opus clip. And it will save them a fortune because they're just going to tie it into their codex account and use their monthly subs. It's going to cost them next to nothing. And on the website, I did get codex to connect to Blender. And someone did a video a few months back. It was like a website called Igloo. And it's got a lovely animation that breaks up of an igloo on the front. But that's not made with normal HTML.

[26:53] They would have done it in Cinema 4D. But I got it to make a wireframe copy in Blender. So I couldn't be bothered to add in all the textures for ice and stuff. But it made it nearly perfectly, perfect copy of it, and then imported that into the website. You got the same effect that it exploded as your mouse over it. So yeah, connected to something like Blender can even boost your websites even further, can make it look amazing. Yeah, what you're talking about is like parallax. So when a website runs parallax, when you scroll down and things are moving along with it. No, no, no, no, it wasn't that. It wasn't that. Because I asked the AI, how did they make this specific animation? I find the website in a minute and link it. And they said, no, this was made in Cinema 4D. It was more like a kind of particle based animation. It was a bit more complex. Because I see a lot of people doing a lot of people doing the parallax where things are exploding as you scroll down. And you could do that in Codex, you can have it generate an image, right as like the hero section of the image. And then you can say break this image into multiple layers, and apply parallax to it. So what happens is if it's like a scene of like nature and a bird flying, they're all independent layers that are broken up. So as you scroll down, it'll go to the next thing and the next thing and get more and more immersive. So that's a cool way to use it too, is to just tell Codex like, I mean, you can just find a website and give it like, just make this website like mine, but just apply my branding. That's what a lot of people do. And no one says where they just take a website that looks really good and has a really cool element. And you know, just apply it to my website and they'll just go and do it for you. - Love it. Cool. - I'll link the igloo site. If you have a look and you look at that website, that's Cinema 4D that they've used to do a separate animation for it's really, really fucking cool website. But yeah, if you connect through to something like Blender, you can start to produce websites like that as well. It takes a little bit more effort because you've got to understand about shading and stuff. But the AI will just take control of Blender and make that for you. - Here's the site that, not the real Andrew Erickson. I renamed you Jake. Not the real Andrew. Here it is. Let's see. Let's see if we can open this again in the incognito window. Let's see the loading screen. Yeah, just incredible kind of stuff we can do nowadays. I bet you this is like a one prompt or like a three prompt situation. So love it. Cool guys. I want to, if I can't, oh is there sound? - This would have been hand coded. This really would have been probably like a 20, 30 thousand dollar website or more for that company because someone's hand coded their animators in Cinema 4D. But AI can start to replicate this. If you just connect it with the MCPs to Blender, you can produce it and then have that brought in the effect and it will save you a ton of money. It can take your websites to even better level. - I was a web developer circa 2014 and this would have been, I mean, we were just, first of all, I don't think we would have done this at all, period. It just wouldn't work. But if we quoted something like this, this would be six figures to get something like this built out 10, 12 years ago. Ritu, I would love to hear from you. You always have such great insight and I always love your tips and tricks and everything. If I can get your, I wonder if you have a hot take on Astra versus Claude or Fable and also from your perspective as a PPC person, is there any like cool badass stuff that you're doing at PPC that we could all learn from?

30:51 Codex vs. Claude: verbosity, plan mode and a Rubik's cube website

[30:50] - Oh, thank you so much, Andrew. Yeah, appreciate it. So, you know, I've been using both, of course, and using hand-off documents or just having both of these read the same directory structure. So it's kind of seamless. You don't really need to copy and paste, et cetera, from Claude to Codex and vice versa. You know, I've found Claude to be pretty verbose recently. Like it just talks a lot. And honestly, reading between the lines is a little bit of an extra brain load. Like I find that Codex does a much better job with simplifying the message and it's not that verbose. It gets the job done. Definitely great at coding. Definitely I agree with John that it's so much better at the visual side of things. Connecting Blender, yes, we've done that as well. So video creation is through, you know, like connected with Blender. I can show you some cool motion graphics stuff.

[32:00] - Yeah, sure. - Can I add in a quick thing? - Yes. - It's okay with you, Rita? - Yeah. - So one of the things that I do just to standardize between this, Claude, I would just standardize what it reports back. So you're showing your screen right now, but I can share it later. But basically it's just like, hey, give me an abstract, give me like any action items that I need to do. And that way it always looks the same, whether I'm using, you know, Claude or ChatGBT or, you know, whatever the next frontier model is. If you just have a standardized output, then yeah, it doesn't matter how verbose it is, you know, just read the summary. - Yeah. - Exactly. - Yeah, no, that's a great tip. Actually the verbosity is mostly to do with when it's planning. Like I put it in plan mode and I ask it to explain how it's going to do its thing. 'Cause I don't ever, or at least I try not to go straight into coding. I always introduce the plan step because there's an extra bit of thinking that it does in order to present it to me. And then it actually does the implementation. And so that saves a little bit of the wasted kind of thinking and collaborating later on. So that's what I'm saying.

[33:20] It's like, it tends to be a little verbose. Okay, so my screen is showing, right? - Yes, we see codecs here. - You can see it, okay. Yeah, yeah, yeah. So basically I was remodeling my website entirely. And so this is, you know, I wanted some sort of cool elements. Like I definitely wanted some sort of visual kind of motion graphics that are interactive and, you know, it showcases all of our work very quickly. And so it did bring this. Now, honestly, this was an idea I picked up from a different website. So I just, I told it to make it a Rubik's cube and I wanted it to have all the faces have some of our work. But then it also did some of the other stuff pretty well. Like it did, it made all of our, you know, videos. It arranged it thoughtfully. It made it, you know, it added extra information that I didn't have to specify and stuff like that. So I do think that website creation with Astra has been quite amazing.

[34:33] And with Claude, actually did I share my screen at all or it's not showing, right? - We see your codecs in your, yeah. - You do see, okay. - We see your desktop, yeah, with codecs here, website. - Yeah, okay. - There you go. - Yeah, okay, yeah. So I mean, with Claude, I do think that, you know, it does a great job with orchestration and, you know, combining Claude and Astra. I think everybody would agree that that's the way to go because they're neck to neck as Andrew was showing the stats. It was like, yeah, they're neck to neck, but you want to take advantage of both. And one more thing to add, John, you're paying quite a bit. I'm impressed. Like I don't pay as much. I'm paying a hundred dollars for Claude. I'm a cheapskate. I'm paying a hundred dollars for Claude and for codecs, I'm only paying 20. And yes, I do run into those five-hour limits and I'm like, fine with that.

[35:38] I don't know. I just want to make sure. - Pay a little more. Never be limited by tech. - I know, I know. - You're an incredible brain.

35:47 "I'm the bottleneck, not the AI"

[35:47] - I know. You know, Andrew's given me that advice earlier also. But the thing is that at the end of the day, I'm the bottleneck. The AI is no longer the bottleneck. So I'm like, okay, if I'm going to be waiting on this, if I'm going to be the one that it needs to wait on, it's okay. Because I have so many tabs just like you, Andrew. Like I have like, I don't know, 50 different projects that are happening simultaneously. So I'm the bottleneck here. But you're right. I mean, maybe I should go up a little bit on the codec side. But anyway. - We're going to pass a tip jar around for you. That was a good tip. We're going to pass it around and get you the next subscription level. - Love it. Awesome. Ritu, I love it. I always love your insight. This is fantastic. Good share. Shubhash, if I can put you on the spot here. I saw in the chat, you were mentioning a little bit about how to connect Clod and ChatGBT together. I've found that this works really well to have them kind of sort of critique each other. It works very, very well. So I'm wondering if you

36:57 Running Codex and Claude Code side by side in VS Code

[36:56] can give us a little bit of insight. - Yeah, no, thank you. So I like to use, I'm a big VS Code fan. I haven't moved from VS Code for a while now. And what works well for me is that I have organized it in a way where I have a master folder where all my work is organized in subfolders within it. And so when I use any of the extensions, official extensions from Anthropic or OpenAI. So I can run codex and cloud code in parallel or copilot as well for that matter. And they have full context of it. So if I were to break a task in 10 parts and we know that certain model is good at certain part of the task. So that handoff happens. And when that handoff happens, they get full context along with it. So it becomes really easy like, and you can see both of them working in same time parallel next to each other. So yeah, that works really well for me.

38:06 The harness matters more than the model

[38:06] One other thing I wanted to share was like when we were talking about Astra versus Fable, Andrew, when we were discussing about that. When we discuss these things, I think we forget or we don't take into mind the importance of hardness in these models. So like when, like, let me share. I just quickly made a small presentation for that. How do I share? I'm excited because I got a lot of notes on both their harnesses as well. Both Raymond and Shubhash are both like IT experts, so they can nerd out pretty hard on some of the technical stuff. Go for it. Let's see it. So this is like what I was talking about in terms of hardness. Like models are okay. Like this is what happens if I were to break down it into steps. So task is broken down into steps. Then there's prompt and context. The tools it can call. Then there's the model. Then there's retry and loop. And then what can be counted as done. So for any task of these six steps, model is only here, the fourth one. It does play a good role. It does play a very sizable role, I would say.

[39:31] But the hardness becomes more important because that literally is stage one, two, three, five, and six. So I would say that when we argue about models being better than one or the other, it's just one stage of those six stages. And so it kind of limits the whole scope of argument, in my opinion, on that. And the model is the engine, obviously, yes. The hardness is the chassis of it. The tire is the pit crew and the race is the strategy. So two teams can bolt in identical engine and finish a lap apart. Nobody watching would conclude that the engine was a difference. So that is what is kind of misleading, I would say. I mean, misleading sounds a bit hard, but yeah, that's what a benchmark headline can exactly say. And if I were to go a step more on it, so the effect is not small. Same model, same benchmark, same week, only the hardness was changed and all these metrics change by many, many folds. So that's like, if you would just take one thing from it, it would be just understanding this part of it. When we say you use a model, so it's not exactly just using the model. Yes, that's the engine part of it. But the chassis, the tire, the strategy, the crew, everything else in it becomes equally important. So yeah, that's my two thoughts on that. - Awesome, man. That's fun. Yeah, that's good stuff. Raymond, I know you're keen on these things too. Anything to add or anything?

[41:09] - It sounds, Shubhash, it sounds like you were using a different harness from what was stock. I had notes on the stock harness. Did I understand that correctly? I didn't see what your two graphs were comparing. - I mean, so the graphs, they could be not exactly the right numbers in there. Don't quote me on the numbers in the graph, but that was just a quick pulling together. But the point of showing that was that the model becomes just one part of it. It's not the whole picture of it. And secondly, what I think as well, like reading from many thoughts on Twitter as well, the way Astra is designed or what it is actually very good at is all these shady videos like renderer and all of these stuff, which makes it very flashy on social media. So it kind of builds on that momentum itself and it undermines the capability of other models. So when we think about it, we have to weigh both the sides of the argument. That's what I would, I would think on

42:18 Why Claude understands intent better

[42:18] that. - Okay. So I would say like my short notes of the difference between the two harnesses, I would say that I prefer Claude because they take an extra few steps to understand what your intent is. So for example, you're saying like, oh, I want to save this in my documents folder, right? But let's say you misspelled documents or when you transcribe it, it was document. I would find that Chateau T would save a literal folder called document. Whereas Claude would be like, oh, you probably meant documents and we'll save it there. And I just kept running into that issue time and again, where that little extra step of understanding my intent that Claude did was all the difference in having less of my attention, less of my brainpower. So in that sense, I kind of prefer Claude to start the task, to do the planning for the task. I don't know if that's anyone else's

43:20 Reviewing work across multiple models

[43:19] experience, but. - The way I do it, I just have, so I'm using a terminal called Claude code in terminal, right? And then I have codex installed as a CLI or plugin, however it works. I told Claude to do it and I did it, right? And then I'm able to say like, hey, use codex to double check this plan, use codex to look at the visuals, use codex to do this, or have codex plan it, and then Claude can double check it. That's what I found. I found that they are really good at countering each other. So one of our processes on everything we do is we say, we do like a team review. And the team review is both like different avatars looking at the stuff, and then also

44:09 Agents reviewing each other's work

[44:08] different models. And that seems to work really well. - Does anyone use the buzz to have them chat together when they're in the lab? - Buzz? - Yeah. - I know John Aspinall is pretty keen on Buzz. I think I just saw him disappear. Oh, there he is. - Buzz, no, I'm here, sorry. I double book calls all the time 'cause I'm useless with that kind of stuff. But no, Buzz is, so Buzz, so Jack Dorsey also has Bird, B-E-R-D, which is basically Buzz for just one person and you don't have to be collaborative. It has very weird avatars inside of it. It's very odd looking. But Buzz is great because then you can set up separate agents that lay on top. So Buzz lays on top of whatever's on your computer. So he literally doesn't ask you for any logins for Claude or codex or anything like that. And then like in my use case, I'll have codex go out and do an image stack for a client.

[45:09] And then part of the workflow is that then once it's done, it doesn't give it to me, it gives it to another agent who is on Fable 5. And that's for QC and quality control. And then if Fable 5 passes it, then I get it. But if Fable 5 says, no, it's not good, it kicks it back to codex. So Buzz is an alum, I guess, kind of, but Buzz is a little bit better because Buzz is basically a Slack replacement. So essentially it's free and it could be more collaborative. So Drew, if you have like five people on your team, you can invite them into your Buzz and then you can have them on an allow list where they can talk to your agents. So I see a lot of business owners making value in this because they don't want to go and set up another team account or pay for a call. They'd rather just use, have their VA or EA, someone use their already dialed in account. So then someone that comes into your Buzz can be like, okay, at codex agent, do this for me. And then that workflow kicks off and all that kind of cool stuff.

[46:10] It is a RAM hog. I don't know why it's one of those apps that you load it and your RAM shoots through the roof. I thought I got a new Mac mini like eight months ago with 24 gigs of RAM. I thought I was sitting pretty, but it will top out really, really heavily. But no, it's good. But it's still, how can I say this? Like Andrew, you were talking about how you have codex in terminal. I almost use everything in terminal where it's cloud code, grok, anything like that. But codex is the one that I specifically use the app for because of certain features in it. Cause I do a lot with visuals. So I like when it gives me the preview of the image and I can comment on it on the X, Y axis coordinate about changing different things. So that's literally the only actual native app I will always use is the codex app. - Love it. Good shout out. I want to switch it

47:05 voice.md: give your AI a consistent brand voice

[47:10] a little bit here and I know we're getting a little technical, which is fun. This is why I bring super smart, amazing people in here. And I know some of this is probably going over some people's heads, which is totally fine because this is how we learn as we stretch our consumption a little bit, stretch our information diet. I would love to bring it down maybe a little bit more practical for all of us who are traditional private label brand sellers. What are people using these things that are really useful in your everyday thing? It doesn't have to be some crazy super duper triple hack thing. Could be just like listing images, analyzing your PPC, looking at your supply chain, looking at your different stuff like that. These processes that we had to do every single day back in 2024 before AI was super useful. Any use cases that people want to share? - Yeah, I've been working on something like, you know, you get the design MD file, the Google open sourced. I'm so bored of the AI slot with content when you're asking it to make stuff. There's so many skills out there to do with it. I got the AI really to have a look at just about all these AI slot skills, take the best, get rid of the worst. And I created a markdown folder I call voice.md. I then had it go and look at the top 100 kind of branded sites out there, Apple, everyone, et cetera, to learn their tone of voice and how they write their content.

[48:49] So where you would, if you're doing a website, load in the design MD file for the fonts and, you know, the colours, et cetera, and the layout. I can now load in something where it will load in a voice. So if I want Apple's voice or Ferrari's voice or Nike's voice, now I probably want Nike's voice. I can just tell the agent go and get that. And it actually makes much higher quality content now. And it wasn't that difficult to set up. I basically told it, this is design MD, these are Google's docs. This is what they expect from it. I want one that's a voice for written content only. And it went and produced it. And the content output is far superior. Get the odd bit of slop in there still. We can't get rid of every year of Dash. But it generally does a much better job. And I find it does a better job than all these ones. You just, you see a YouTuber, oh, I've got this. No, I've combined them all to take the best. That's, I think, a very good everyday use because you can have it, you got a product. You'd say I want it in, I don't know, Jim Bean's voice, you know, and I just got it to, and it's made the files automatic. I've got 100 of them there. It can pick from or I can send it to any website. And it will make a branded voice in that company's style, all their marketing material, the whole lot. Love it. Good share. Not through

50:18 Auditing token waste (and the $20 plan challenge)

[50:18] Landry Erickson. Good share. Raymond, go ahead.  So this is something that is kind of like a lower priority task, but I think all of us will eventually have to do it. And that's just, you know, keeping our tokens, our turn use in check. So one of the biggest tasks that I had codecs do was just, okay, here's all my claw transcripts, here's all my codecs transcripts. Just analyze things and figure out what used up more tokens than it should have. So that could be like, you know, you're making a billion subagents when you really didn't have to. It could be that you have a bunch of skills that you're not using. And one of the things that I challenged myself before Astra came out was I just downgraded to the $20 clawed plan just to really figure out what's going on with all my contacts. And, yeah, it did a really good job with that. I think I could probably go down to the $20 plan now. So, you know, it gives me that much more headroom

51:25 Hermes agents + OpenRouter: cheap models for cheap tasks

[51:25] when I go back to $100.  Yeah, that's a good one. You know, I have a couple of Hermes agents that I run, and they run through my -- I use them as like a proper executive assistant for each one of my companies. And they're -- I use Hermes because it's sitting in a VPS, a cloud computer, and it's throwing stuff, you know, every night and every afternoon and stuff like that, right? And it's all consistent. It's all -- I don't have to have my laptop open, you know, to like run something. So, that's why I use Hermes. But Hermes does -- I do have to use the non-subscription stuff for it. I know there's ways to set up subscription, but I set up open router.ai, which is a tool that makes it really easy to switch between models. And I have different models do different tasks, right? So, I found that email triage. So, I go through and it triages my email and adds tags and adds filters and everything. That is really easy for really cheap models to do.

[52:33] And so, we have an incredibly cheap model doing it. I was asking my bot to give me like a rundown of each one of the models that I'm using, and it's computing it right now. I'll be able to paste it in. But that triaging, you know, triaging email really easy. We also analyze everything and update my LLM Wiki, which is like a way of kind of doing a knowledge base, and it's shared on GitHub. So, that way it's in the cloud and always accessible between all the agents. That needs to be a little bit smarter because it's analyzing and reviewing things. So, I use a different model. It's a little bit smarter. And of course, anytime I am personally talking to it, I want it to be very smart. I don't need exactly Fable 5.1 smartness, but like I need OpenSonet 5, something that's like pretty smart if I'm actually talking directly. But I agree, Raymond, it drastically cuts down on the cost and optimizes tokens if you kind of triage like that.

53:35 Context hygiene, second brains and decision logs

[53:34] Well, I'm going to actually push back on what you're saying. So, it's not necessarily just switching out the model. And Jake did point out there is like a slash doctor skill, which will in Clod, which will like analyze your setup and everything. But it's not just that because it's also your context. So, it's potential like just in general with these agents, they would prefer to just add text rather than look through and see, "Hey, was this already there?" So, what can end up happening is that you're going to have an agent figure out like, let's say, "Oh, I want to do like a Q4 schedule." But maybe you had like your PPC numbers in two different places and they say two different things just because they were updated at different times. And yeah, so it might pull from that, but then you're like, "Oh, I've got this step as well." And it tries to like combine things and it can be very messy if you're not also monitoring how your context is looking. Like, is there duplicates? Is there separate information elsewhere? Because like not only will that expand the file, but it'll also take up like a ton of tokens. - I think that's also important to note that if you don't have a system in place, then yeah, we'll 100% do that. But like in my like second brain system, there's like persistent memory and decision logs. So, it'll do a sweep every seven days of what have I been working on that's relevant and all that kind of stuff. And then keep that up higher in the decision log. So, no matter if I'm putting it in Grok, Astra, Claude code, it has that context. I think Raymond, you're right. If you don't have the system in place where it's gonna like really guide the tool, if you just give it free range and just say, "Do this resource for me," or, "Do this outline for me," then it's gonna go wild and do whatever it wants and suck up as much tokens and usage. But if you have those parameters in place of like decision logs and persistent memory, it would definitely aid in not going so crazy. - Yeah, definitely. And it's not like the sexiest like, "Oh, here's this website that I made or here's this visual," but it's definitely something that's very important that we should all be doing and paying attention to. - Love it, guys. Well, we're at the top of the hour here.

[56:02] This is a fun one. I'm curious if someone's done... Oh, there we go. John's sharing something.

56:09 Using the Codex app for A+ content

[56:09] - All right. I just wanna share last minute. So, this is the reason why I use the Codex app because I cannot do this in Claude code. I cannot do this in Grok. It's not as good, but I can literally take a plus content for a client and say, "Hey, just spin this up, keep it the same, remix it for their orthopedic line instead." And in a matter of two minutes, it gives me hyper exactly what I need with the context. And if I needed to go in here and be like, "Okay, change this, change this, change this," it's able to do it. And I don't care if you're using this in terminal. Yeah, you could patch in the API. You can do all these kinds of things and you're never gonna get this interface that is so quick, number one, so accurate, number two, and so user-friendly, number three. It's a no-brainer for visuals having to use the app because the app is super, super strong. Just wanna share that. - I'll back Jon up on that. I did his cohort and it's amazing. My problem is I'm just spending time redoing it and playing with it and it's addictive.

[57:10] - Love it. I'm curious somebody who's, again, practical private label business owner here, who's kind of done a big sweep of their listing using AI images or who's just chomping down all their PPC. Ed, I know you, Ed Chan, I know you're doing a lot of really amazing hacker stuff with it. Curious if there's any practical, even more examples of fun practical use cases that really changed in the last three, six, 12 months. - I think not on the PPC or visual side, but

57:48 Cloning 3 paid Shopify apps with Astra

[57:51] what I recently did, I did a great job at it, is for our Shopify apps, like Shopify paid apps. So I've distilled three Shopify paid apps that we paid for and verbatim copied it with Astra and got them live and it's amazing. And I added more features that we were using before and now got rid of three paid subscriptions that we had to pay for Shopify apps. I did that right when Astra came out. I was like, "Yeah, let's give it a shot. Let's see how good you're at." And it nailed it. Yeah. It was long horizon tasks for sure. It took a while. I think they were running for like two, three days to get every single feature and function, but the end result was just phenomenal. - Even though Claude and that are meant to have some guard rails on them about going behind paint walls, I've cloned a few apps. I've sent Claude inside to go through a system to work out exactly how it works. And it's done that, made a report and then cloned it pretty much 99% of the way to that piece of software. And I've canceled a few of my paid kind of subscriptions. I've literally just copied it.

59:15 The 3x3 image grid trick

[59:15] - Yeah. We live in a brave new world. - I have one quick tip, not really Astra specific, but Astra did a really good job. Let me share my screen. There you go. Let me know when you guys can see my screen. 'Cause I needed some lifestyle. I just wanted to get some lifestyle images 'cause mine needed improvement. So what I do is instead of generating one image at a time, you can generate a grid. And because these models, they generally have like a super resolution type structure into their architecture. So I know like Gemini does that and obviously Chat GPT does that. So what you can do is like, you can do a three by three grid and you're saying like, "Oh, I like number three." So you can just have it expand that image. So now you're able to iterate through a whole bunch of these a lot quicker and get a full resolution image just by picking whichever numbers you want. But yeah, the reason I was saying that Astra did a lot better is it did a lot better in figuring out what the dimensions of my product were, like relatively what the scale is and just the images look better as well.

[1:00:33] There's no weird artifacts that I found with when I was using Gemini. It just knows the topic better. Is this a UV light? Yeah, it's my UV light. Nice. That's great, man. Love it. Chat, go ahead.

1:00:52 Upscaling AI images for 4-foot print displays

[1:00:52] I just had a quick question regarding images. Is any of the AI models able to create high quality images? I'm talking about for one of the four foot tall displays that you'd put inside of a store and I need a big image on the side. I create great images for the rendering, but I got to give the file to the people that are printing the product display. I create this nice AI image, but it's got to blow up to where it's like four feet by two feet, you know what I'm saying? I bet if you don't have just the basic output working well enough for you, because they can output to 4K or whatever, you could probably tell Codex, "Hey, take this image and generate it in blocks. Break it up if it's vertical. Break it up into eight blocks, four by two, and then generate a 2K image for each one of those blocks." I feel like that would work. If it wasn't already sharp enough, you could probably do that. They also can do vectors too. You can do SVG vectors, which just means that instead of being like pixels, it's like lines, circles, and stuff. Try that. That's just what I would try to get started with. I'm curious if anyone else has insight for Chuck, though.

[1:02:22] I've not tried it because it's not something I really thought about, but I, if you downloaded something like Topaz Desktop, which I do have because I do use it for upscaling, I've got a funny feeling you can actually set the size quite high. What's it called? Topaz, T-O-P-A-Z. I don't know if they do a free trial or not. There's a desktop version if you've got a beefy GPU, I would go with that because it'd be cheaper than a cloud version. Or even Magnific, their upscaler might do it. I would see if one of them will upscale to the size you want. How long it'll take. I've got no idea. Take a 4k image and make it like eight foot tall. I don't know. But maybe it would work, maybe it wouldn't. But I would certainly look at that as a possibility.

[1:03:19] Yeah, I ran it through just real quick. I ran it through like, you know, ask Codex, ask Fable, and I'm like, oh, we're not able to create that. I use cloud design. And I'm like, I gave up and I said I'll revisit it because I know somebody can make this shit. I mean, come on. You guys can build full websites in seconds, I mean. One thing I would say, if you do look up Topaz, don't use it on something like Higgsfield because they'll charge you through the ass for the fucking tokens. Just go straight to Topaz. I've got an annual plan with them, which I did get on the Black Friday sale so it was dirt cheap. But I would try one of their desktop models. They've got quite a few models. There's video and there's image models. But I get a funny feeling one of them might do what you want. I appreciate it.

1:04:11 Repricing across Walmart, Best Buy and Shopify with MCP + Composio

[1:04:10] Love it. Any more questions? I think, yeah, one other use case being a big unlock for us is definitely setting up all the APIs and MCPs for additional marketplaces like Walmart, Best Buy, also Shopify. So we can do like price re-pricing in one click across all the different marketplaces and also for listing updates. We don't have to go through every single marketplace now or have a third-party plugin. We can just do it in-house where we update all the inventory in one click. We can put caps on inventory reservations for different marketplaces. We did that for our Walmart marketplace and Best Buy and whatnot. So that's a big one that we've done in the last week. And also updating them. How we update the listings is, there's two ways we do it. And one way is through Clod, through the MCP. And we have it make a Google Sheet. And also I think I shared it in the group chat. If you use Google Sheets docs, not the native connector, but through Composio, you can update the Google Sheet versus replacing it. Because the native connector on Google Sheets always creates a new Google Sheet every time you want to update it or replace.

[1:05:39] But with Composio connector, it'll update the sheet instead of replacing it. So what we have, we have a pull-down of suggestions. So we get Clod or Codex to make all the suggestions of an audit of all the listings to see if there's any inconsistencies with the spelling or whatnot or with the title and the copy. And then we have a lot of listings, say we have 500 or 300 per marketplace. It's tedious to go through it all. And then so we make separate spreadsheets, go through the spreadsheet and then look at what it suggests, go through the pull-down of all the different suggestions on the pull-down. And then we have a notes section of what we, if it's not on that list, what we want it to say. And then we have it run an audit, then make all the edits and then push it to each separate marketplace. And it takes a fraction of the time to do everything. And it's like instant. That's killer. Tell me about Composio. Is it paid or how does that work? I'm on the Composio free. It actually gives you a lot of flexibility, even on the free version. And they have over a thousand probably plugins and MCPs that you can connect. And you can, the beauty of it is for Google, you can, for, I think most of their plugins, you can connect multiple accounts. So that was the biggest unlock of, if you just have the Google connector under cloud, you can only use that email that you're signing with. With Composio, you can attach unlimited emails to it or Google accounts. Yeah, I use Composio so I can log into multiple Gmail accounts because I have like 10. And it manages it well. In fact, I was going to see if it would do the same with GitHub because when I was mentioning earlier that on Ultra, Codex screwed up my weekly limit in a day, it also screwed up my usage. I mean, it's on GitHub in a day as well, max them out. So I was going to maybe get the AI to go and make me a couple of hundred free GitHub accounts. And basically each project has its own GitHub account and connect them via Composio so they just go out and use each one for nothing. You can, Jake, you can have a GitHub credential connected to each folder so that it uses it automatically because I have two of them.

[1:08:06] I have one for personal, one for business. And for the personal stuff, when it's working in that repo, it always commits to the personal and vice versa. So you don't need a connector app. It'll just do that natively if you just tell it to always use this one account for this repo. Yeah, because it went absolutely bad. I just thought, no, I'm getting all these push errors coming through. My email got flooded with like three or four hundred push errors from GitHub. It was Codex.

1:08:46 MCPs eating RAM: load skills on demand

[1:08:45] That's funny. One thing I have a question about, I've been getting or noticing that when you're using a lot of MCPs on Codex or Clod, it tends to leave them hanging like as nodes or and it just eats up my RAM like crazy. I got to come go in and turn off. Is there a way or have anybody has anybody found a way, maybe through a prompt or, you know, part of its memory to kind of turn it off after they're done with that type of thing? I don't think you can. Well, I mean, maybe someone will correct me, but I don't think you can turn it off per task. That was kind of my my issue. It's just kind of like all the MCPs are just always on is my understanding. So I've just kind of switched more to APIs. So since that's freer, you know, you can have Clod or Codex consume the APIs. And it's usually the exact same thing, but also like what were you going to say, Jake, for interrupting?

[1:09:48] Well, you can set the MCPs the same way as skills. I have like a master agent that has very few skills loaded in. I have a separate repository with all my skills. If I tell it I want to go and make some images, it already knows what skills I want for that. And it just goes and spins up an agent, pulls down those three or four skills. So it's not loading hundreds and hundreds of skills into every single task I'm doing. But if you do the same with MCPs, you say, I will need to do this and go, all right, I'll just go and quickly reinstall that in that MCP. That's it. And then when the job's done, it just kills it. I think, Jake, you might want to check because by default, all of these harnesses, they'll save like a short description of what the skill is. So when you're interacting with it, they'll know, oh, I invoke this skill. So I think you might you may actually end up having a bunch of skills that are like you didn't set to inactive is because you can't, like I said, I did have an absolute ton of skills, like hundreds of them in there. But obviously, your agents read every single skill that it's got regardless of your task. So you've got a massive use of context straight away.

[1:11:03] So I remove them. And so yeah, depending on what task I want, I say I want to do this, the agent knows what I want. And it spins up a new one and just pulls them down. It also checks weekly to see if those skills get updated, because obviously, they're not all from me, the ones I've pulled off of GitHub. So it'll have a look like Matt Pocock's, it'll have a look and so he's updated those and every update my repository where it pulls them from. I mean, I'm assuming it's saving me context, you know, because it's literally, instead of the agents spinning up with 200 skills to check straight away, it's got two in three, just because that's all it needs for that particular task. But not so much on the context. I'm running into issues with my RAM. Like, I've quite a bit around 64 gigs of RAM. But I'm always at like 50. This thing is like just running hot, my fans going I'm a MacBook Pro. And then my, my fans are keep going nuts. And I was like, looking into it, I got codex to look into it. I have a little I built a little monitor. And I was like, okay, it's just pull up. It's just rerunning these skills. Last time I had not skills, MCPs, I had 33 of the same MCP running through 33 instances. I was like, why is it doing this?

[1:12:23] Like, you should, you should not be using 50 gigs of RAM, there's definitely something wrong. I don't know. I have two dozen MCPs. And I'm not anywhere near that. I run Docker too. So Docker like pulls out. You might notice that but for everyone else, Docker is like a like a little tiny machine, just that I can put stuff on there so that it stays kind of contained inside that little tiny machine. And that's that's taking up five gigs for me. And then my my whole laptop was only 16. Right. So I have like, 10, 12 that I use for my, like, regular laptop, and then five for my little Docker session. And I'm running usually pretty okay on that I could always use more, but you shouldn't be using 50 gigs unless you're doing. I'm not sure you'd be doing that wouldn't use that much. But yeah, there's definitely a problem there. Okay, I'm gonna take a deeper dive into that. Yeah. Um, cool, guys, I'm gonna, uh, well, you know, one last one last year,

1:13:23 Tax prep with Claude in the browser

[1:13:23] I just cuz it's a fun one. That was really easy. That was a big win this last week for me was tax prep, man. These things are incredibly good at tax prep now. So I pulled down, you know, don't judge me. I'm trying to do my 2025 stuff. I always follow extensions and extension, you know, the, the S corp stuff extension is due tomorrow. FYI, if you didn't file back in March. And, yeah, pulling down, I have QuickBooks setup, but also I just go to my credit card statements and sometimes or PDF, sometimes or CSV, sometimes they're similar stupid junk that the banks give you because they're annoying. And I had Claude in browser open up and say, okay, like, I want you to click all 24 links and pull down 24 months of data and CSV or sorry, and PDFs. And I popped into a folder, told Claude, Hey, convert all these stupid PDFs into something into a CSV. So I can like, so it's like a, um, an easy text file instead of a big, stupid PDF to read and then go through and analyze it, help me grab my tax, my tax prep, help me look at the things. Let me reconcile my books.

[1:14:36] I still hand it to a professional accountant. I still believe in, you know, you still need a license. Like these AIs are really good, but I need, I need a human with experience and a license behind them to be able to stand behind my taxes still. Uh, but it turned this horrible, terrible task. I hate doing that. It takes me about a week to get all my stuff collected together, turned it into like a two or three hour task to do the whole thing. And so, and also on top of that, I have all this data connected and I can say, now help me analyze my spend. I don't usually pull down my personal checking account cause like just, it's just too much to look through, but I'm like, grab the, grab my emails, grab my personal checking account information, reconcile stuff, look at everything that's going on and suggest ways to improve it. Where can I cut expenses? Where can I, um, what are the things I missed? And I brought up a whole bunch of stuff. Um, so it's just like just a, just an easy win, a fun thing that it's really good at that kind of

1:15:30 Wrap-up + upcoming events

[1:15:29] stuff. So, um, anything, any last parting words before we end the call here? One last thing. Thank you, Andrew, for organizing this. I look forward to this hackathons. There's so much value that I am sure everyone else gets from it. So yeah, really, really thank you for the fact that for these thoughts, people usually charge and we get this for free when you organize it. So yeah, really thank you for that. Awesome. Thank you, Shubhash. Well on that note, that's a great note to end on. So thank you guys. Uh, we went a little past the hour about that. The last 20 minutes are a good one. Um, I'll send out a replay to everyone and then, uh, uh, of course the transcript too. So your AIs can chomp this and summarize it and pull out information for you. Um, and then, uh, we'll probably do another one of these things. Uh, oh, I'm gonna be at Accelerate next week. Um, I would love to meet up in person if anybody's at Accelerate. And then also Kanjan has a conference in LA, um, uh, um, top tier conference. It's October 7th. Can you tell us?

[1:16:38] October 6th to 9th, top tier. Yeah. We have 10,000 attendees. Yeah. All above, mostly TikTok shop creators and creators among area. Also like AI strategies like Andrew and also some big brand operators like from Anchor or from other big teams. So if you are interested, yeah, welcome to join us. Awesome. If you have a link, drop it for us, Can. I'm speaking there. So I'm excited. So, uh, hopefully I'll see some of you guys in Seattle next week and, or LA in two weeks or three, I guess three weeks. Um, and, uh, yeah, thanks guys for joining and we'll talk to you all later. Thank you. Thank you so much. Top tier league.com/summit. Bye guys. Thank you. Thanks.

Watch the earlier hackathon replays

Replay, recorded Monday, September 14

Fable 5.1 vs GPT-6 Astra
20 AI Power Users Compare Notes

Recorded Monday, September 1477 minFull transcript below

Get invited to the next one Watch on YouTube

What the Room Figured Out

Fable vs Astra turned out to be the wrong question. The people getting results win on workflow: effort settings, harness design, clean context, and a person approving every change.

1
Pick the model by cost per task, not the leaderboard
The room’s read: Astra on High beat Fable on Ultra for security audits and code quality, and Fable still wins some jobs. Test both on the task you actually run and compare what each one costs to get it done.
2
Plan high, execute lower
One member ran Astra on Ultra and burned a week of tokens in a day. Max effort also overthinks, like writing a test for every tiny thing. Plan with the top model on the top effort setting, then do the work on High or below.
3
The harness matters more than the model
Shubhash walked through six steps: task, prompt and context, tools, model, retry loop, done. The model is the engine and the harness is the rest of the car. Changing only the harness can move results many times over.
4
Real workflows, not demos
Codex cut a raw six-minute talking-head video to five with branded motion graphics in about 17 minutes. A long-running Astra task cloned three paid Shopify apps. Codex plus Blender over MCP rebuilt a 3D site animation that cost around $30K to hand-build.
5
A person approves every change
The multi-marketplace repricing flow has the AI audit listings across Walmart, Best Buy and Shopify, write suggestions to a Google Sheet, and only push to each marketplace after a person signs off.
6
Keep context clean and cheap
voice.md files for a consistent brand voice, auditing past chats for wasted tokens, cheap models for cheap tasks (Hermes + OpenRouter), decision logs so agents stop repeating context, and loading MCPs only when a task needs them.

Chapters

Links Mentioned

Copy the Full Transcript

Full transcript

Fable 5.1 vs GPT-6 Astra: AI Agent Hackathon, full transcript
Recorded Monday, September 14, 2026. Replay: https://youtu.be/z6qy0zQ_nqg
Machine transcription with no speaker labels, so some names and terms may be misheard. Timestamps match the YouTube video.

0:00 Welcome + Alibaba CoCreate recap

[0:00] Good, good, good to see you, Ed, Can, awesome. Can, tell us, while I get my notes prepped here, I think of a fun thing would be to, if you want to help us talk about the event we were just at, Co-Create, sponsored by Alibaba, we were at last week. So why don't you tell the group a little bit about that while I get my notes prepped here. Yeah, yeah, yeah, do you want me to talk about the topic I shared there? Or just have a general idea of what Co-Create looked like, yeah. Yeah, well, so maybe can I say an intro of what Co-Create by Alibaba was? And then maybe a little bit, some of the things you learn and what's the cool stuff that's happening in that space. Maybe five minutes, just to kind of tell people what's up. Yeah, yeah, great.

[0:52] Actually, the Co-Create is a very big conference hosted by Alibaba. They told me this time they have about 20,000 attendees. So well, it's a big room. You remember we attend the VIP party there. It doesn't look like a VIP party. Yeah, more than 300 attendees in the room in just a VIP party. So it seems like people are so passionate about all this AI and e-commerce thing. Yeah, because this is an event by Alibaba. They want to promote one about their AI agent actual work. And two, more about to group the people who are in the e-commerce area, yeah, to discuss the next things. I think this year, people pay a lot more attention to AI agent. Before, it's just like AI bot, right? How to use a lot more like a talking ball. How to use AI to give inspiration for our daily lives, to ask question.

[1:49] But this year, it seems that everyone want to use agent. Want to include agent in our life, in our business to make our business more in a automated way and enjoy more efficiency. I believe, although now actual work is not perfect right now, right? Like many agent, but it do open our mind that maybe next year, yeah, all this AI agent can be very around and they can handle a lot of our day life and business. And I really imagine, yeah, and I have a big expectation of all this AI, how they can create new business models. Yeah, like from internet era versus the previous era, the business model have been totally strange. And I believe in the AI era, it can be another level. Yeah, something really new.

[2:44] And I mean, you also bring something good ideas and good things along this journey I believe we will, yeah. See a lot of new things happening. Yeah. Love it. Yeah, it was a fun event, a little bit different than ones I usually go to. I'm usually very specific like operator. Here's how I SEO maximize my listing on TikTok and Amazon and this was a little bit more broad audience, but it was a good conference. Cool guys, we already have 20 people in here already, which is very exciting. It looks like we have a pretty good participation this week. It looks like I shared a link, unfortunately, that named half of you guys Andrew Erickson. There is really only the one real Andrew Erickson. And so if you don't mind changing your name makes conversations a little bit easier if you change your name to your preferred name. Otherwise, because it's easier for us all kind of to know and see each other.

[3:45] But okay, I put together a few slides, of course, Claude and Fable and Astra put together. Somebody put Andrew Erickson V2, nice. I put together a few slides, I thought it'd be kind of good little, like here's a quick little intro. It's the other Andrew, that's the V2 Andrew, awesome. Quick little intro just to kind of set the tone here so everyone kind of is on the same page because I know not everyone is obsessed with all the AI news like I am. So I'm going to share my screen here and kind of show you guys a few things that happened just in the last three weeks, last two weeks, in September. September has been a very busy month so far. So I am going to share my screen here. And then after that, we're going to have an open discussion.

[4:40] I see John Aspinall is here. He's always a fun lively chat, so I'm sure he'll have some opinions on these things. And I would love to just get, if anybody has some, the thing I love here is when all these smart people, I pull together lots and lots of smart people. If you're here, you're an awesome Ecom AI smart person. If you have some insight or something you've found that works really well or some new kind of insight into what is happening, I'd love for you guys to share that after the talk here. But let me get my share screen.

5:21 Two weeks that changed AI: Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra

[5:20] Okay, can you guys see my little slide deck here? You got two robots, Astra 6, Fable 5. Perfect, yes, perfect. Okay, cool, I just want, this is just like a five, ten minute thing just to kind of put you guys on the same page. So, yeah, of course, here's my fun graphic. Of course, this was created by Astra, by ChatGPT. Because their new image model is incredible. It is breaking all of the leaderboards now, so it's pretty cool to see. Okay, here's what's happened in the last two weeks. Claude Fable 5.1 came out, which of course is just kind of an upgrade from Fable 5.0, which came out August at some point, mid-August. It's a little bit more refinement of the 5.0, but the Fable model is incredible.

[6:20] I'm like, I love the Fable model. 5.1 is a little bit of just an update onto the basic Fable model. A Gemini, one thing, I'm not seeing a lot of press on this, but Google shipped their own Gemini 3.8 flash model. This is a really good model. It's very smart. It's not smarter than the other models, but it's incredibly fast and cheap. So, I think for a lot of tasks, I think it's actually something worth looking at, obviously, depending on what you want to do. And of course, here's our big one is that OpenAI shipped out GPT-6 model. And then, this was really cool. I have a math background, and so this is big news.

7:08 OpenAI's Navier-Stokes proof: 10,000 agents, 88 hours, $15M in tokens

[7:07] Last week, OpenAI published a proof to the Navier-Stokes equation. It doesn't really matter what that is. It's some crazy math theorem that we've been trying to prove for almost 100 years, and OpenAI published a proof to it. There's a little bit of controversy of how they got the answer to that proof, but we do have a proof, and it was pretty remarkable. This is kind of showing that we're ushering into a new realm where the AIs are, they were good, and now they're great. They seem to be superhuman in a lot of spaces now. And then, lastly, just a few days ago, OpenAI shipped out their image 2.5 model, which is pretty amazing. So, yeah, I'm gonna kind of go through these ones really quick. So yeah, the big thing that Azure's really good at, GPT-6 is really good at, is the driving your browser and your desktop.

[8:01] It's incredibly good at doing your desktop now. It is price points there, and it's really good at in-depth reasoning. Here it is. Okay, so the Navier-Stokes equation, it's like an equation of how fluids move. And when they move, if they create a singularity, cuz the way the equations work, it looks like there's little singularities everywhere. And so they had this out as a problem that was almost, yes, 200 years old. But it was a millennium prize question, which is a big deal. These are really important questions. At the turn of the millennium, 25 years ago, a bunch of mathematicians got together and they set out nine problems that they are really excited, they really want proved. And they set a million dollar prize on each one of those nine problems, and OpenAI just solved one of them. So the way they solved it, it's a part where I'm like, this is crazy, cuz people think, do these things actually scale?

[8:59] Can you actually get better? Are we gonna hit some ceiling? And I think the answer is, hell no, we're hitting a ceiling anytime soon. Because what they did is they took their model, it actually wasn't Astra, it was actually their secret model that's behind closed doors. And they ran 10,000 agents at once. They ran it for 88 hours straight, so four days basically. And then it took 17 hours to write the summary, and then it won the millennium prize. So of the nine equations, only one has ever been solved, and this is the second one. So it's just a big deal in the space. Now, does it really affect us? Not really, but it's just kind of a testament to how incredible these models are. GBT Image 2.5, it's awesome. It's, there's a, there is a, This is my favorite site. I like to use to compare models, cuz all these things are always changing constantly.

[10:00] I'll drop this in the chat. Where's the chat button? Okay. Shubhash asks, is OpenAI getting the million dollars? Yes, I think that's kind of the, it's not fully accepted, proven, fully whatever. They haven't officially designated a winner. But everyone's kind of like, okay, yeah, minus the actual legal stuff. Yes, it won, and I guess OpenAI is going to make the million dollars. Ironically, they spent $15 million in tokens. At least that's the retail price of the tokens. $15 million getting that million dollar prize, so not a great ROI. But of course, I think the press is well worth that $15 million.

10:48 Arena.ai head-to-head: Fable vs. Astra

[10:48] Here, I dropped in the chat here a, there's some good chat messages here. I dropped in the chat one of my favorite sites here, arena.ai. I love this site that helps, you can kind of just compare side by side the different models out there, and you can see that for agentic use, Claude and GPT Astra are both side by side. So they're technically, according to the way they measure this here, this is a human, this is a side by side human comparison, rather than some benchmark thing. I think it's a good way of measuring it, because you just give two answers. A person shows up on the site, and they get two answers, and they have to compare the two, which one they like better. In my mind, this is just a good way of just having a regular person using this. Which one do they like better? And this is the answer.

[11:40] So this is the agentic. We're technically fables beating Astra, but they're so incredibly close. And then also here, I like this one a lot. If you go to image, text to image, and then product and branding, commercial. This to me is a really important one that we all care about, since we're in the product category, and here's text to image. And so you can see what the images, who's winning here. I dropped the link in the chat. And then just GPT is just crushing, absolutely crushing this one. So if you, Grok is doing well, Gemini's okay, and there's some. Microsoft, I guess, is doing pretty well too. But I would definitely highly suggest going with GPT image 2.5. Again, that model just came out like four days ago. Okay. Ritu's coming in, awesome.

12:38 Where Fable wins, cost per task, and which model for what

[12:38] Okay, where is Fable really good at? Fable is still really good at coding. If you're doing long run coding projects, this works really well. There's the price point. The real cost is a little bit cheaper. The way the price points work, it kind of depends. You want to measure per task rather than the actual input output, because there's all sorts of weird little nuances around caching. And how long it takes, and how many tokens they use, and stuff like that. The important thing is just kind of look at how many, what the cost is per task. And we can get into the whole idea of how do we, if you care about cost, how we can maximize those things. There's all sorts of fun ways where you can have Fable do one thing and farm it out to Sonnet or to the GPT models. And they're a little bit faster, a little bit cheaper to do it that way. But that's it on price and then real agents. Those are the one I just showed you guys.

[13:31] But yeah, technically, Fable is beating Astra right now by a tick. But they're so close, it's just they're neck and neck. Okay, here you go. So the cost, yeah, so in general, if you're looking at cost per task, Fable is about double GPT right now, even a little bit more than double. Again, these tasks are kind of defined in a special way. The way you do your own task will be very different than how you, these things are measuring it. Let's see, and then again, last mention to Gemini 3.8. It is very smart, very fast, and very cheap. So it's also, if you're kind of looking for a model that isn't just the frontier, most expensive thing possible. Gemini, just kind of an auto mention there too. So there you go. There is the last slide. If you're doing images, I highly suggest you use images and listings, definitely GPT. If you're doing code clots, probably still winning, although they're both, they both complement each other really well.

[14:32] And then if you're doing volume, speed, and cheapness, you want a frontier model, Gemini is really good. That's what I got. Let's open it up. I'd love to hear from the rest of our smart, capable Ecom AI people here.

14:49 GPT-6 Astra on Ultra burned a week of tokens in a day

[14:49] What's everyone thinking of these models? Well, I'm really liking a GPT-6 Astra. I'm using it more than Claude, although I did make one mistake. I set it on a task on Ultra, and it killed the weekly tokens in the day. So I now just set it to high only, and it seems to last a lot longer. But the output has been really good, especially with the security audit, because obviously it doesn't drop the model down. Fable drops me down to 4.8, misses stuff. Astra picks up loads of stuff. It's great. I really like it. What were you using the Ultra code or the high end for in Astra that sucked up all the tokens? I was working on a CMS system, and it had about 15 different things it wanted to implement in a plan.

[15:48] I thought, "Oh, do you know what? I'm going to stick on Ultra and just let it go ahead." No, it massacred the token so quickly. I've now had it do a similar plan on high, and I've still got limit. Yeah, I typically would always put it on top model, top effort level for planning. And developing the plan, because nine times out of 10, from my experience, you don't need that ultra high effort and ultra high model to actually execute. So that's another way to think of it too. Yeah, it was an expensive lesson for a day, because I normally now get Fable to do the plan. I then have Fable and Astra chat together about the plan and work out the best way. And then one of them does the work, and the other one checks the work. But Astra, I find, catches quite a lot of bugs, and also is really good at improving the code of it, just the speed of it.

16:53 Max effort overthinks: plan high, execute lower

[16:53] I was going to say, there's also an issue if you use the max level effort, max level thinking, and they can sometimes overthink, especially if you're doing knowledge work or coding. For example, if you're writing test at the highest, they might literally test every single little thing when you really don't need that. So that's what I would caution against when you're using the max effort. I haven't specifically looked at Astra's benchmarks, but I believe that was true for Fable. And it also takes forever. So I don't know if anyone's ever put on like ultra or... It takes forever because it's a lot of effort. And sometimes I almost, I was reading something that talked about how GPT-6 Astra on lower medium is comparable to like Soul56 on high or ultra high, right? So I would say, you know, right person, right job, right tool, where it's like, unless you're doing, if you're doing it like what I do, like a lot of content, knowledge based kind of things, GPT-6 Astra on even low does laps around five, six soul.

[18:10] So I think it's really important to understand like what you're doing and what you need it for. And if it's doing like a lot of coding and a lot of checking, because nine times out of 10, I talk to people, they do not need to be on ultra code or X code or whatever it is, because it's just, it's going to take forever and it's just going to suck up a ton of usage. Yeah. I found the speed is, is frustrating. Cause I, I have like six sessions going at, or even more, but try to get six sessions moving. But at some point I'm like, Oh man, I'm just trying to update one little thing. And it spins for 14 minutes updating that one little thing. Like just, I don't need to super duper. I don't need the council of PhDs chime in on this, like updating MOQs on, on something, you know, Does anyone, does anyone actually make, um,

19:00 Codex edited a raw talking-head video with motion graphics

[18:59] uh, cause this is what I found and I literally just did it just now. Um, does anyone do like video content, like either demo videos or talking head videos or just marketing videos? Because, um, in the codex, I strongly believe that the codex app is the best app, um, compared to like Claude or anything else because the features inside of it are just insane. So I just shot a video talking about how you could help local businesses use AI just nonsense, right? Not nonsense is good, but it was like six minutes and I didn't want to edit it. So I just dragged and dropped it into a codex chat. And I said, here's my branding on my website, go and edit the video for me, remove out the ums, ahs, whatever. And then layer on like branded react components, animations, charts, and graphs. It did it. And it chopped it from a six minute video. It's like a five minute video with full blown editing, full blown animations. So it's like, if you think outside the box of, oh, it's just for coding and making an app or doing a tool, like you can put it on autopilot for your content, which is pretty cool. So I don't know if anyone's doing video content, John, it's a good shout out. So I I've used, uh, so like a week or two ago, I used Opus five for creating a couple of videos and I was actually taking one of these calls and chopping it up into shorts and it did a pretty good job. I was very impressed with it. And I'm, this is the kind of the fun thing where we're finding that jagged edge of like, what is, what are these things really good at that we can step up as, you know, as business owners, like, Hey, I never was able to make shorts, but now I can do this. And now I can do that. Like, where are these new, uh, capabilities enabling us? Can you give us a, like, pretend we all know nothing? Uh, give us a quick, quick basis. I'm like, Hey, I have codex installed. I want to, I want to take a couple of these pictures or a couple of these videos and make them into something cool. Yeah. So I do, you're gonna do something cool. Okay. Thank you, John.

[20:56] So I literally said, Hey, can you edit my video and my downloads folder? So I think it's important to know the tool you're in, right? So I don't have to drag and drop my video here and waste time. It's already in my downloads folder. So I'm telling codex last video, my downloads folder, really good cuts with visual, uh, uh, motion visual graphics for me. And in eight minutes, it gave me a video and the video was like, okay. So then I was like, okay, well, can you, I don't want my last name at the top of the video. It's not really, you know, I have a point, but make it more visually aesthetic on brand. And I also want you to create animated, um, react visuals. So react is like, if anyone doesn't know what, like the components, like, so Claude, for example, can't do visual, like actual images, but it can use react components to put together a visual, which is really basic rudimentary kind of like shapes and stuff. So it took nine more minutes. So if I gave it that prompt initially, it would only be nine minutes, but just to give you some perspective, this was the video that I shot here, where it's just me and a codex screen and all this kind of stuff. And it's, you know, going through it. And then if I go to the newly edited one, see this, I would never have the wherewithal to do motion graphics popping up from idea to AI service over here. And if I just scrub through it, you see how it took my video and like shrunk it down, but then put the positioning over here at room. It basically moved my head from being in the lower left hand side to being in the little dedicated box over here. And then even this part right here, I missed the trends of this, like, so it does these visual, I'm sorry, one second.

[22:49] It does all these transitions like this. So if I was to actually do this myself, I wouldn't, number one, but if I was to have someone else do it, they would charge me an arm and a leg just to do this at scale. So that even something like this, like it knew like the context I was talking about someone selling custom printed shirts on Facebook, right? In Facebook marketplace, it pulled out all the context and gave me this really well done edited video that I could just take and scale. So that's like one way that I'm using it. The other way that I'm using it over

23:23 A TikTok cover skill that reads the video

[23:23] here is I'm scaling my TikTok content because, you know, it's where all the kids are. So I made a skill over here that's a TikTok cover because what I found was that when you upload a TikTok, you can do a custom cover thumbnail for your feed. But what happens is you have to like pick a scene in the video and then put text on it. But whenever I picked one, it was always like a weird like motion of me doing something. So I was like, wait, let me just make this skill where I give it the URL. It's going to read the context of the video, like watch it, get the transcript and then build me out the thumbnail that I can go and add. So now at scale, I can then take it and let me spell TikTok, right? And if I look at my profile, now I'm able to have instead of like weird kind of things like this. Now they're a little bit more tailored and buttoned up. So when someone takes a look at it. So just some cool stuff that I'm doing because for me, Codex is just like it runs everything, right? Like I pay for $200 Claude code, $200 Codex, $300 Grok. Grok's a whole different story, right? But for Codex, fundamentally with GBTSixastra, it could do anything you need. I love it, man. That's good. Good share there. And I think a lesson here for the rest of us is that there are some really cool videos coming out, video capabilities. You don't necessarily need to have a really expensive model or some special software for it outside of just these AIs are already like good, already pretty good at it.

24:59 No excuse for a bad website in 2026

[24:59] But like you had said, sorry, you had said Opus, right? So I always look at it like, I don't want to pay for more tools. So how can I leverage the crap that I pay for now to replace that? So like, I never wanted to pay three or $4,000 for a website. I never wanted to. I didn't want to go and buy it. So I just have AI make it for me. And as it gets better, you could, so like there's no excuse why your brand shouldn't have a website in 2026. In September of 2026, there's no excuse why you shouldn't have a website and it shouldn't be awesome. Like your business, your agency, like if your website sucks, it's all on you because you have it all at your fingertips to do it awesome. I'm going to what you just said right there. I'm going to clip that out. I'm going to have I'm going to have Astrid clip that out and make it a short for you. We'll see how it works.

25:47 Cloning Opus Clip + rebuilding the Igloo site with Blender

[25:46] Yeah, go ahead. Dig. Yeah. And talk about video and website. I was working, doing a day last week for a friend to set up their AI. They were spending a fortune on Opus clip, you know, that takes their videos, cuts them down for shorts, for Instagram. In a couple of hours, we made pretty much a usable copy of Opus clip. And I'm going back in there on Wednesday to refine it as I think you can make virtually a 99% exact copy of Opus clip. And it will save them a fortune because they're just going to tie it into their codex account and use their monthly subs. It's going to cost them next to nothing. And on the website, I did get codex to connect to Blender. And someone did a video a few months back. It was like a website called Igloo. And it's got a lovely animation that breaks up of an igloo on the front. But that's not made with normal HTML.

[26:53] They would have done it in Cinema 4D. But I got it to make a wireframe copy in Blender. So I couldn't be bothered to add in all the textures for ice and stuff. But it made it nearly perfectly, perfect copy of it, and then imported that into the website. You got the same effect that it exploded as your mouse over it. So yeah, connected to something like Blender can even boost your websites even further, can make it look amazing. Yeah, what you're talking about is like parallax. So when a website runs parallax, when you scroll down and things are moving along with it. No, no, no, no, it wasn't that. It wasn't that. Because I asked the AI, how did they make this specific animation? I find the website in a minute and link it. And they said, no, this was made in Cinema 4D. It was more like a kind of particle based animation. It was a bit more complex. Because I see a lot of people doing a lot of people doing the parallax where things are exploding as you scroll down. And you could do that in Codex, you can have it generate an image, right as like the hero section of the image. And then you can say break this image into multiple layers, and apply parallax to it. So what happens is if it's like a scene of like nature and a bird flying, they're all independent layers that are broken up. So as you scroll down, it'll go to the next thing and the next thing and get more and more immersive. So that's a cool way to use it too, is to just tell Codex like, I mean, you can just find a website and give it like, just make this website like mine, but just apply my branding. That's what a lot of people do. And no one says where they just take a website that looks really good and has a really cool element. And you know, just apply it to my website and they'll just go and do it for you. - Love it. Cool. - I'll link the igloo site. If you have a look and you look at that website, that's Cinema 4D that they've used to do a separate animation for it's really, really fucking cool website. But yeah, if you connect through to something like Blender, you can start to produce websites like that as well. It takes a little bit more effort because you've got to understand about shading and stuff. But the AI will just take control of Blender and make that for you. - Here's the site that, not the real Andrew Erickson. I renamed you Jake. Not the real Andrew. Here it is. Let's see. Let's see if we can open this again in the incognito window. Let's see the loading screen. Yeah, just incredible kind of stuff we can do nowadays. I bet you this is like a one prompt or like a three prompt situation. So love it. Cool guys. I want to, if I can't, oh is there sound? - This would have been hand coded. This really would have been probably like a 20, 30 thousand dollar website or more for that company because someone's hand coded their animators in Cinema 4D. But AI can start to replicate this. If you just connect it with the MCPs to Blender, you can produce it and then have that brought in the effect and it will save you a ton of money. It can take your websites to even better level. - I was a web developer circa 2014 and this would have been, I mean, we were just, first of all, I don't think we would have done this at all, period. It just wouldn't work. But if we quoted something like this, this would be six figures to get something like this built out 10, 12 years ago. Ritu, I would love to hear from you. You always have such great insight and I always love your tips and tricks and everything. If I can get your, I wonder if you have a hot take on Astra versus Claude or Fable and also from your perspective as a PPC person, is there any like cool badass stuff that you're doing at PPC that we could all learn from?

30:51 Codex vs. Claude: verbosity, plan mode and a Rubik's cube website

[30:50] - Oh, thank you so much, Andrew. Yeah, appreciate it. So, you know, I've been using both, of course, and using hand-off documents or just having both of these read the same directory structure. So it's kind of seamless. You don't really need to copy and paste, et cetera, from Claude to Codex and vice versa. You know, I've found Claude to be pretty verbose recently. Like it just talks a lot. And honestly, reading between the lines is a little bit of an extra brain load. Like I find that Codex does a much better job with simplifying the message and it's not that verbose. It gets the job done. Definitely great at coding. Definitely I agree with John that it's so much better at the visual side of things. Connecting Blender, yes, we've done that as well. So video creation is through, you know, like connected with Blender. I can show you some cool motion graphics stuff.

[32:00] - Yeah, sure. - Can I add in a quick thing? - Yes. - It's okay with you, Rita? - Yeah. - So one of the things that I do just to standardize between this, Claude, I would just standardize what it reports back. So you're showing your screen right now, but I can share it later. But basically it's just like, hey, give me an abstract, give me like any action items that I need to do. And that way it always looks the same, whether I'm using, you know, Claude or ChatGBT or, you know, whatever the next frontier model is. If you just have a standardized output, then yeah, it doesn't matter how verbose it is, you know, just read the summary. - Yeah. - Exactly. - Yeah, no, that's a great tip. Actually the verbosity is mostly to do with when it's planning. Like I put it in plan mode and I ask it to explain how it's going to do its thing. 'Cause I don't ever, or at least I try not to go straight into coding. I always introduce the plan step because there's an extra bit of thinking that it does in order to present it to me. And then it actually does the implementation. And so that saves a little bit of the wasted kind of thinking and collaborating later on. So that's what I'm saying.

[33:20] It's like, it tends to be a little verbose. Okay, so my screen is showing, right? - Yes, we see codecs here. - You can see it, okay. Yeah, yeah, yeah. So basically I was remodeling my website entirely. And so this is, you know, I wanted some sort of cool elements. Like I definitely wanted some sort of visual kind of motion graphics that are interactive and, you know, it showcases all of our work very quickly. And so it did bring this. Now, honestly, this was an idea I picked up from a different website. So I just, I told it to make it a Rubik's cube and I wanted it to have all the faces have some of our work. But then it also did some of the other stuff pretty well. Like it did, it made all of our, you know, videos. It arranged it thoughtfully. It made it, you know, it added extra information that I didn't have to specify and stuff like that. So I do think that website creation with Astra has been quite amazing.

[34:33] And with Claude, actually did I share my screen at all or it's not showing, right? - We see your codecs in your, yeah. - You do see, okay. - We see your desktop, yeah, with codecs here, website. - Yeah, okay. - There you go. - Yeah, okay, yeah. So I mean, with Claude, I do think that, you know, it does a great job with orchestration and, you know, combining Claude and Astra. I think everybody would agree that that's the way to go because they're neck to neck as Andrew was showing the stats. It was like, yeah, they're neck to neck, but you want to take advantage of both. And one more thing to add, John, you're paying quite a bit. I'm impressed. Like I don't pay as much. I'm paying a hundred dollars for Claude. I'm a cheapskate. I'm paying a hundred dollars for Claude and for codecs, I'm only paying 20. And yes, I do run into those five-hour limits and I'm like, fine with that.

[35:38] I don't know. I just want to make sure. - Pay a little more. Never be limited by tech. - I know, I know. - You're an incredible brain.

35:47 "I'm the bottleneck, not the AI"

[35:47] - I know. You know, Andrew's given me that advice earlier also. But the thing is that at the end of the day, I'm the bottleneck. The AI is no longer the bottleneck. So I'm like, okay, if I'm going to be waiting on this, if I'm going to be the one that it needs to wait on, it's okay. Because I have so many tabs just like you, Andrew. Like I have like, I don't know, 50 different projects that are happening simultaneously. So I'm the bottleneck here. But you're right. I mean, maybe I should go up a little bit on the codec side. But anyway. - We're going to pass a tip jar around for you. That was a good tip. We're going to pass it around and get you the next subscription level. - Love it. Awesome. Ritu, I love it. I always love your insight. This is fantastic. Good share. Shubhash, if I can put you on the spot here. I saw in the chat, you were mentioning a little bit about how to connect Clod and ChatGBT together. I've found that this works really well to have them kind of sort of critique each other. It works very, very well. So I'm wondering if you

36:57 Running Codex and Claude Code side by side in VS Code

[36:56] can give us a little bit of insight. - Yeah, no, thank you. So I like to use, I'm a big VS Code fan. I haven't moved from VS Code for a while now. And what works well for me is that I have organized it in a way where I have a master folder where all my work is organized in subfolders within it. And so when I use any of the extensions, official extensions from Anthropic or OpenAI. So I can run codex and cloud code in parallel or copilot as well for that matter. And they have full context of it. So if I were to break a task in 10 parts and we know that certain model is good at certain part of the task. So that handoff happens. And when that handoff happens, they get full context along with it. So it becomes really easy like, and you can see both of them working in same time parallel next to each other. So yeah, that works really well for me.

38:06 The harness matters more than the model

[38:06] One other thing I wanted to share was like when we were talking about Astra versus Fable, Andrew, when we were discussing about that. When we discuss these things, I think we forget or we don't take into mind the importance of hardness in these models. So like when, like, let me share. I just quickly made a small presentation for that. How do I share? I'm excited because I got a lot of notes on both their harnesses as well. Both Raymond and Shubhash are both like IT experts, so they can nerd out pretty hard on some of the technical stuff. Go for it. Let's see it. So this is like what I was talking about in terms of hardness. Like models are okay. Like this is what happens if I were to break down it into steps. So task is broken down into steps. Then there's prompt and context. The tools it can call. Then there's the model. Then there's retry and loop. And then what can be counted as done. So for any task of these six steps, model is only here, the fourth one. It does play a good role. It does play a very sizable role, I would say.

[39:31] But the hardness becomes more important because that literally is stage one, two, three, five, and six. So I would say that when we argue about models being better than one or the other, it's just one stage of those six stages. And so it kind of limits the whole scope of argument, in my opinion, on that. And the model is the engine, obviously, yes. The hardness is the chassis of it. The tire is the pit crew and the race is the strategy. So two teams can bolt in identical engine and finish a lap apart. Nobody watching would conclude that the engine was a difference. So that is what is kind of misleading, I would say. I mean, misleading sounds a bit hard, but yeah, that's what a benchmark headline can exactly say. And if I were to go a step more on it, so the effect is not small. Same model, same benchmark, same week, only the hardness was changed and all these metrics change by many, many folds. So that's like, if you would just take one thing from it, it would be just understanding this part of it. When we say you use a model, so it's not exactly just using the model. Yes, that's the engine part of it. But the chassis, the tire, the strategy, the crew, everything else in it becomes equally important. So yeah, that's my two thoughts on that. - Awesome, man. That's fun. Yeah, that's good stuff. Raymond, I know you're keen on these things too. Anything to add or anything?

[41:09] - It sounds, Shubhash, it sounds like you were using a different harness from what was stock. I had notes on the stock harness. Did I understand that correctly? I didn't see what your two graphs were comparing. - I mean, so the graphs, they could be not exactly the right numbers in there. Don't quote me on the numbers in the graph, but that was just a quick pulling together. But the point of showing that was that the model becomes just one part of it. It's not the whole picture of it. And secondly, what I think as well, like reading from many thoughts on Twitter as well, the way Astra is designed or what it is actually very good at is all these shady videos like renderer and all of these stuff, which makes it very flashy on social media. So it kind of builds on that momentum itself and it undermines the capability of other models. So when we think about it, we have to weigh both the sides of the argument. That's what I would, I would think on

42:18 Why Claude understands intent better

[42:18] that. - Okay. So I would say like my short notes of the difference between the two harnesses, I would say that I prefer Claude because they take an extra few steps to understand what your intent is. So for example, you're saying like, oh, I want to save this in my documents folder, right? But let's say you misspelled documents or when you transcribe it, it was document. I would find that Chateau T would save a literal folder called document. Whereas Claude would be like, oh, you probably meant documents and we'll save it there. And I just kept running into that issue time and again, where that little extra step of understanding my intent that Claude did was all the difference in having less of my attention, less of my brainpower. So in that sense, I kind of prefer Claude to start the task, to do the planning for the task. I don't know if that's anyone else's

43:20 Reviewing work across multiple models

[43:19] experience, but. - The way I do it, I just have, so I'm using a terminal called Claude code in terminal, right? And then I have codex installed as a CLI or plugin, however it works. I told Claude to do it and I did it, right? And then I'm able to say like, hey, use codex to double check this plan, use codex to look at the visuals, use codex to do this, or have codex plan it, and then Claude can double check it. That's what I found. I found that they are really good at countering each other. So one of our processes on everything we do is we say, we do like a team review. And the team review is both like different avatars looking at the stuff, and then also

44:09 Agents reviewing each other's work

[44:08] different models. And that seems to work really well. - Does anyone use the buzz to have them chat together when they're in the lab? - Buzz? - Yeah. - I know John Aspinall is pretty keen on Buzz. I think I just saw him disappear. Oh, there he is. - Buzz, no, I'm here, sorry. I double book calls all the time 'cause I'm useless with that kind of stuff. But no, Buzz is, so Buzz, so Jack Dorsey also has Bird, B-E-R-D, which is basically Buzz for just one person and you don't have to be collaborative. It has very weird avatars inside of it. It's very odd looking. But Buzz is great because then you can set up separate agents that lay on top. So Buzz lays on top of whatever's on your computer. So he literally doesn't ask you for any logins for Claude or codex or anything like that. And then like in my use case, I'll have codex go out and do an image stack for a client.

[45:09] And then part of the workflow is that then once it's done, it doesn't give it to me, it gives it to another agent who is on Fable 5. And that's for QC and quality control. And then if Fable 5 passes it, then I get it. But if Fable 5 says, no, it's not good, it kicks it back to codex. So Buzz is an alum, I guess, kind of, but Buzz is a little bit better because Buzz is basically a Slack replacement. So essentially it's free and it could be more collaborative. So Drew, if you have like five people on your team, you can invite them into your Buzz and then you can have them on an allow list where they can talk to your agents. So I see a lot of business owners making value in this because they don't want to go and set up another team account or pay for a call. They'd rather just use, have their VA or EA, someone use their already dialed in account. So then someone that comes into your Buzz can be like, okay, at codex agent, do this for me. And then that workflow kicks off and all that kind of cool stuff.

[46:10] It is a RAM hog. I don't know why it's one of those apps that you load it and your RAM shoots through the roof. I thought I got a new Mac mini like eight months ago with 24 gigs of RAM. I thought I was sitting pretty, but it will top out really, really heavily. But no, it's good. But it's still, how can I say this? Like Andrew, you were talking about how you have codex in terminal. I almost use everything in terminal where it's cloud code, grok, anything like that. But codex is the one that I specifically use the app for because of certain features in it. Cause I do a lot with visuals. So I like when it gives me the preview of the image and I can comment on it on the X, Y axis coordinate about changing different things. So that's literally the only actual native app I will always use is the codex app. - Love it. Good shout out. I want to switch it

47:05 voice.md: give your AI a consistent brand voice

[47:10] a little bit here and I know we're getting a little technical, which is fun. This is why I bring super smart, amazing people in here. And I know some of this is probably going over some people's heads, which is totally fine because this is how we learn as we stretch our consumption a little bit, stretch our information diet. I would love to bring it down maybe a little bit more practical for all of us who are traditional private label brand sellers. What are people using these things that are really useful in your everyday thing? It doesn't have to be some crazy super duper triple hack thing. Could be just like listing images, analyzing your PPC, looking at your supply chain, looking at your different stuff like that. These processes that we had to do every single day back in 2024 before AI was super useful. Any use cases that people want to share? - Yeah, I've been working on something like, you know, you get the design MD file, the Google open sourced. I'm so bored of the AI slot with content when you're asking it to make stuff. There's so many skills out there to do with it. I got the AI really to have a look at just about all these AI slot skills, take the best, get rid of the worst. And I created a markdown folder I call voice.md. I then had it go and look at the top 100 kind of branded sites out there, Apple, everyone, et cetera, to learn their tone of voice and how they write their content.

[48:49] So where you would, if you're doing a website, load in the design MD file for the fonts and, you know, the colours, et cetera, and the layout. I can now load in something where it will load in a voice. So if I want Apple's voice or Ferrari's voice or Nike's voice, now I probably want Nike's voice. I can just tell the agent go and get that. And it actually makes much higher quality content now. And it wasn't that difficult to set up. I basically told it, this is design MD, these are Google's docs. This is what they expect from it. I want one that's a voice for written content only. And it went and produced it. And the content output is far superior. Get the odd bit of slop in there still. We can't get rid of every year of Dash. But it generally does a much better job. And I find it does a better job than all these ones. You just, you see a YouTuber, oh, I've got this. No, I've combined them all to take the best. That's, I think, a very good everyday use because you can have it, you got a product. You'd say I want it in, I don't know, Jim Bean's voice, you know, and I just got it to, and it's made the files automatic. I've got 100 of them there. It can pick from or I can send it to any website. And it will make a branded voice in that company's style, all their marketing material, the whole lot. Love it. Good share. Not through

50:18 Auditing token waste (and the $20 plan challenge)

[50:18] Landry Erickson. Good share. Raymond, go ahead.  So this is something that is kind of like a lower priority task, but I think all of us will eventually have to do it. And that's just, you know, keeping our tokens, our turn use in check. So one of the biggest tasks that I had codecs do was just, okay, here's all my claw transcripts, here's all my codecs transcripts. Just analyze things and figure out what used up more tokens than it should have. So that could be like, you know, you're making a billion subagents when you really didn't have to. It could be that you have a bunch of skills that you're not using. And one of the things that I challenged myself before Astra came out was I just downgraded to the $20 clawed plan just to really figure out what's going on with all my contacts. And, yeah, it did a really good job with that. I think I could probably go down to the $20 plan now. So, you know, it gives me that much more headroom

51:25 Hermes agents + OpenRouter: cheap models for cheap tasks

[51:25] when I go back to $100.  Yeah, that's a good one. You know, I have a couple of Hermes agents that I run, and they run through my -- I use them as like a proper executive assistant for each one of my companies. And they're -- I use Hermes because it's sitting in a VPS, a cloud computer, and it's throwing stuff, you know, every night and every afternoon and stuff like that, right? And it's all consistent. It's all -- I don't have to have my laptop open, you know, to like run something. So, that's why I use Hermes. But Hermes does -- I do have to use the non-subscription stuff for it. I know there's ways to set up subscription, but I set up open router.ai, which is a tool that makes it really easy to switch between models. And I have different models do different tasks, right? So, I found that email triage. So, I go through and it triages my email and adds tags and adds filters and everything. That is really easy for really cheap models to do.

[52:33] And so, we have an incredibly cheap model doing it. I was asking my bot to give me like a rundown of each one of the models that I'm using, and it's computing it right now. I'll be able to paste it in. But that triaging, you know, triaging email really easy. We also analyze everything and update my LLM Wiki, which is like a way of kind of doing a knowledge base, and it's shared on GitHub. So, that way it's in the cloud and always accessible between all the agents. That needs to be a little bit smarter because it's analyzing and reviewing things. So, I use a different model. It's a little bit smarter. And of course, anytime I am personally talking to it, I want it to be very smart. I don't need exactly Fable 5.1 smartness, but like I need OpenSonet 5, something that's like pretty smart if I'm actually talking directly. But I agree, Raymond, it drastically cuts down on the cost and optimizes tokens if you kind of triage like that.

53:35 Context hygiene, second brains and decision logs

[53:34] Well, I'm going to actually push back on what you're saying. So, it's not necessarily just switching out the model. And Jake did point out there is like a slash doctor skill, which will in Clod, which will like analyze your setup and everything. But it's not just that because it's also your context. So, it's potential like just in general with these agents, they would prefer to just add text rather than look through and see, "Hey, was this already there?" So, what can end up happening is that you're going to have an agent figure out like, let's say, "Oh, I want to do like a Q4 schedule." But maybe you had like your PPC numbers in two different places and they say two different things just because they were updated at different times. And yeah, so it might pull from that, but then you're like, "Oh, I've got this step as well." And it tries to like combine things and it can be very messy if you're not also monitoring how your context is looking. Like, is there duplicates? Is there separate information elsewhere? Because like not only will that expand the file, but it'll also take up like a ton of tokens. - I think that's also important to note that if you don't have a system in place, then yeah, we'll 100% do that. But like in my like second brain system, there's like persistent memory and decision logs. So, it'll do a sweep every seven days of what have I been working on that's relevant and all that kind of stuff. And then keep that up higher in the decision log. So, no matter if I'm putting it in Grok, Astra, Claude code, it has that context. I think Raymond, you're right. If you don't have the system in place where it's gonna like really guide the tool, if you just give it free range and just say, "Do this resource for me," or, "Do this outline for me," then it's gonna go wild and do whatever it wants and suck up as much tokens and usage. But if you have those parameters in place of like decision logs and persistent memory, it would definitely aid in not going so crazy. - Yeah, definitely. And it's not like the sexiest like, "Oh, here's this website that I made or here's this visual," but it's definitely something that's very important that we should all be doing and paying attention to. - Love it, guys. Well, we're at the top of the hour here.

[56:02] This is a fun one. I'm curious if someone's done... Oh, there we go. John's sharing something.

56:09 Using the Codex app for A+ content

[56:09] - All right. I just wanna share last minute. So, this is the reason why I use the Codex app because I cannot do this in Claude code. I cannot do this in Grok. It's not as good, but I can literally take a plus content for a client and say, "Hey, just spin this up, keep it the same, remix it for their orthopedic line instead." And in a matter of two minutes, it gives me hyper exactly what I need with the context. And if I needed to go in here and be like, "Okay, change this, change this, change this," it's able to do it. And I don't care if you're using this in terminal. Yeah, you could patch in the API. You can do all these kinds of things and you're never gonna get this interface that is so quick, number one, so accurate, number two, and so user-friendly, number three. It's a no-brainer for visuals having to use the app because the app is super, super strong. Just wanna share that. - I'll back Jon up on that. I did his cohort and it's amazing. My problem is I'm just spending time redoing it and playing with it and it's addictive.

[57:10] - Love it. I'm curious somebody who's, again, practical private label business owner here, who's kind of done a big sweep of their listing using AI images or who's just chomping down all their PPC. Ed, I know you, Ed Chan, I know you're doing a lot of really amazing hacker stuff with it. Curious if there's any practical, even more examples of fun practical use cases that really changed in the last three, six, 12 months. - I think not on the PPC or visual side, but

57:48 Cloning 3 paid Shopify apps with Astra

[57:51] what I recently did, I did a great job at it, is for our Shopify apps, like Shopify paid apps. So I've distilled three Shopify paid apps that we paid for and verbatim copied it with Astra and got them live and it's amazing. And I added more features that we were using before and now got rid of three paid subscriptions that we had to pay for Shopify apps. I did that right when Astra came out. I was like, "Yeah, let's give it a shot. Let's see how good you're at." And it nailed it. Yeah. It was long horizon tasks for sure. It took a while. I think they were running for like two, three days to get every single feature and function, but the end result was just phenomenal. - Even though Claude and that are meant to have some guard rails on them about going behind paint walls, I've cloned a few apps. I've sent Claude inside to go through a system to work out exactly how it works. And it's done that, made a report and then cloned it pretty much 99% of the way to that piece of software. And I've canceled a few of my paid kind of subscriptions. I've literally just copied it.

59:15 The 3x3 image grid trick

[59:15] - Yeah. We live in a brave new world. - I have one quick tip, not really Astra specific, but Astra did a really good job. Let me share my screen. There you go. Let me know when you guys can see my screen. 'Cause I needed some lifestyle. I just wanted to get some lifestyle images 'cause mine needed improvement. So what I do is instead of generating one image at a time, you can generate a grid. And because these models, they generally have like a super resolution type structure into their architecture. So I know like Gemini does that and obviously Chat GPT does that. So what you can do is like, you can do a three by three grid and you're saying like, "Oh, I like number three." So you can just have it expand that image. So now you're able to iterate through a whole bunch of these a lot quicker and get a full resolution image just by picking whichever numbers you want. But yeah, the reason I was saying that Astra did a lot better is it did a lot better in figuring out what the dimensions of my product were, like relatively what the scale is and just the images look better as well.

[1:00:33] There's no weird artifacts that I found with when I was using Gemini. It just knows the topic better. Is this a UV light? Yeah, it's my UV light. Nice. That's great, man. Love it. Chat, go ahead.

1:00:52 Upscaling AI images for 4-foot print displays

[1:00:52] I just had a quick question regarding images. Is any of the AI models able to create high quality images? I'm talking about for one of the four foot tall displays that you'd put inside of a store and I need a big image on the side. I create great images for the rendering, but I got to give the file to the people that are printing the product display. I create this nice AI image, but it's got to blow up to where it's like four feet by two feet, you know what I'm saying? I bet if you don't have just the basic output working well enough for you, because they can output to 4K or whatever, you could probably tell Codex, "Hey, take this image and generate it in blocks. Break it up if it's vertical. Break it up into eight blocks, four by two, and then generate a 2K image for each one of those blocks." I feel like that would work. If it wasn't already sharp enough, you could probably do that. They also can do vectors too. You can do SVG vectors, which just means that instead of being like pixels, it's like lines, circles, and stuff. Try that. That's just what I would try to get started with. I'm curious if anyone else has insight for Chuck, though.

[1:02:22] I've not tried it because it's not something I really thought about, but I, if you downloaded something like Topaz Desktop, which I do have because I do use it for upscaling, I've got a funny feeling you can actually set the size quite high. What's it called? Topaz, T-O-P-A-Z. I don't know if they do a free trial or not. There's a desktop version if you've got a beefy GPU, I would go with that because it'd be cheaper than a cloud version. Or even Magnific, their upscaler might do it. I would see if one of them will upscale to the size you want. How long it'll take. I've got no idea. Take a 4k image and make it like eight foot tall. I don't know. But maybe it would work, maybe it wouldn't. But I would certainly look at that as a possibility.

[1:03:19] Yeah, I ran it through just real quick. I ran it through like, you know, ask Codex, ask Fable, and I'm like, oh, we're not able to create that. I use cloud design. And I'm like, I gave up and I said I'll revisit it because I know somebody can make this shit. I mean, come on. You guys can build full websites in seconds, I mean. One thing I would say, if you do look up Topaz, don't use it on something like Higgsfield because they'll charge you through the ass for the fucking tokens. Just go straight to Topaz. I've got an annual plan with them, which I did get on the Black Friday sale so it was dirt cheap. But I would try one of their desktop models. They've got quite a few models. There's video and there's image models. But I get a funny feeling one of them might do what you want. I appreciate it.

1:04:11 Repricing across Walmart, Best Buy and Shopify with MCP + Composio

[1:04:10] Love it. Any more questions? I think, yeah, one other use case being a big unlock for us is definitely setting up all the APIs and MCPs for additional marketplaces like Walmart, Best Buy, also Shopify. So we can do like price re-pricing in one click across all the different marketplaces and also for listing updates. We don't have to go through every single marketplace now or have a third-party plugin. We can just do it in-house where we update all the inventory in one click. We can put caps on inventory reservations for different marketplaces. We did that for our Walmart marketplace and Best Buy and whatnot. So that's a big one that we've done in the last week. And also updating them. How we update the listings is, there's two ways we do it. And one way is through Clod, through the MCP. And we have it make a Google Sheet. And also I think I shared it in the group chat. If you use Google Sheets docs, not the native connector, but through Composio, you can update the Google Sheet versus replacing it. Because the native connector on Google Sheets always creates a new Google Sheet every time you want to update it or replace.

[1:05:39] But with Composio connector, it'll update the sheet instead of replacing it. So what we have, we have a pull-down of suggestions. So we get Clod or Codex to make all the suggestions of an audit of all the listings to see if there's any inconsistencies with the spelling or whatnot or with the title and the copy. And then we have a lot of listings, say we have 500 or 300 per marketplace. It's tedious to go through it all. And then so we make separate spreadsheets, go through the spreadsheet and then look at what it suggests, go through the pull-down of all the different suggestions on the pull-down. And then we have a notes section of what we, if it's not on that list, what we want it to say. And then we have it run an audit, then make all the edits and then push it to each separate marketplace. And it takes a fraction of the time to do everything. And it's like instant. That's killer. Tell me about Composio. Is it paid or how does that work? I'm on the Composio free. It actually gives you a lot of flexibility, even on the free version. And they have over a thousand probably plugins and MCPs that you can connect. And you can, the beauty of it is for Google, you can, for, I think most of their plugins, you can connect multiple accounts. So that was the biggest unlock of, if you just have the Google connector under cloud, you can only use that email that you're signing with. With Composio, you can attach unlimited emails to it or Google accounts. Yeah, I use Composio so I can log into multiple Gmail accounts because I have like 10. And it manages it well. In fact, I was going to see if it would do the same with GitHub because when I was mentioning earlier that on Ultra, Codex screwed up my weekly limit in a day, it also screwed up my usage. I mean, it's on GitHub in a day as well, max them out. So I was going to maybe get the AI to go and make me a couple of hundred free GitHub accounts. And basically each project has its own GitHub account and connect them via Composio so they just go out and use each one for nothing. You can, Jake, you can have a GitHub credential connected to each folder so that it uses it automatically because I have two of them.

[1:08:06] I have one for personal, one for business. And for the personal stuff, when it's working in that repo, it always commits to the personal and vice versa. So you don't need a connector app. It'll just do that natively if you just tell it to always use this one account for this repo. Yeah, because it went absolutely bad. I just thought, no, I'm getting all these push errors coming through. My email got flooded with like three or four hundred push errors from GitHub. It was Codex.

1:08:46 MCPs eating RAM: load skills on demand

[1:08:45] That's funny. One thing I have a question about, I've been getting or noticing that when you're using a lot of MCPs on Codex or Clod, it tends to leave them hanging like as nodes or and it just eats up my RAM like crazy. I got to come go in and turn off. Is there a way or have anybody has anybody found a way, maybe through a prompt or, you know, part of its memory to kind of turn it off after they're done with that type of thing? I don't think you can. Well, I mean, maybe someone will correct me, but I don't think you can turn it off per task. That was kind of my my issue. It's just kind of like all the MCPs are just always on is my understanding. So I've just kind of switched more to APIs. So since that's freer, you know, you can have Clod or Codex consume the APIs. And it's usually the exact same thing, but also like what were you going to say, Jake, for interrupting?

[1:09:48] Well, you can set the MCPs the same way as skills. I have like a master agent that has very few skills loaded in. I have a separate repository with all my skills. If I tell it I want to go and make some images, it already knows what skills I want for that. And it just goes and spins up an agent, pulls down those three or four skills. So it's not loading hundreds and hundreds of skills into every single task I'm doing. But if you do the same with MCPs, you say, I will need to do this and go, all right, I'll just go and quickly reinstall that in that MCP. That's it. And then when the job's done, it just kills it. I think, Jake, you might want to check because by default, all of these harnesses, they'll save like a short description of what the skill is. So when you're interacting with it, they'll know, oh, I invoke this skill. So I think you might you may actually end up having a bunch of skills that are like you didn't set to inactive is because you can't, like I said, I did have an absolute ton of skills, like hundreds of them in there. But obviously, your agents read every single skill that it's got regardless of your task. So you've got a massive use of context straight away.

[1:11:03] So I remove them. And so yeah, depending on what task I want, I say I want to do this, the agent knows what I want. And it spins up a new one and just pulls them down. It also checks weekly to see if those skills get updated, because obviously, they're not all from me, the ones I've pulled off of GitHub. So it'll have a look like Matt Pocock's, it'll have a look and so he's updated those and every update my repository where it pulls them from. I mean, I'm assuming it's saving me context, you know, because it's literally, instead of the agents spinning up with 200 skills to check straight away, it's got two in three, just because that's all it needs for that particular task. But not so much on the context. I'm running into issues with my RAM. Like, I've quite a bit around 64 gigs of RAM. But I'm always at like 50. This thing is like just running hot, my fans going I'm a MacBook Pro. And then my, my fans are keep going nuts. And I was like, looking into it, I got codex to look into it. I have a little I built a little monitor. And I was like, okay, it's just pull up. It's just rerunning these skills. Last time I had not skills, MCPs, I had 33 of the same MCP running through 33 instances. I was like, why is it doing this?

[1:12:23] Like, you should, you should not be using 50 gigs of RAM, there's definitely something wrong. I don't know. I have two dozen MCPs. And I'm not anywhere near that. I run Docker too. So Docker like pulls out. You might notice that but for everyone else, Docker is like a like a little tiny machine, just that I can put stuff on there so that it stays kind of contained inside that little tiny machine. And that's that's taking up five gigs for me. And then my my whole laptop was only 16. Right. So I have like, 10, 12 that I use for my, like, regular laptop, and then five for my little Docker session. And I'm running usually pretty okay on that I could always use more, but you shouldn't be using 50 gigs unless you're doing. I'm not sure you'd be doing that wouldn't use that much. But yeah, there's definitely a problem there. Okay, I'm gonna take a deeper dive into that. Yeah. Um, cool, guys, I'm gonna, uh, well, you know, one last one last year,

1:13:23 Tax prep with Claude in the browser

[1:13:23] I just cuz it's a fun one. That was really easy. That was a big win this last week for me was tax prep, man. These things are incredibly good at tax prep now. So I pulled down, you know, don't judge me. I'm trying to do my 2025 stuff. I always follow extensions and extension, you know, the, the S corp stuff extension is due tomorrow. FYI, if you didn't file back in March. And, yeah, pulling down, I have QuickBooks setup, but also I just go to my credit card statements and sometimes or PDF, sometimes or CSV, sometimes they're similar stupid junk that the banks give you because they're annoying. And I had Claude in browser open up and say, okay, like, I want you to click all 24 links and pull down 24 months of data and CSV or sorry, and PDFs. And I popped into a folder, told Claude, Hey, convert all these stupid PDFs into something into a CSV. So I can like, so it's like a, um, an easy text file instead of a big, stupid PDF to read and then go through and analyze it, help me grab my tax, my tax prep, help me look at the things. Let me reconcile my books.

[1:14:36] I still hand it to a professional accountant. I still believe in, you know, you still need a license. Like these AIs are really good, but I need, I need a human with experience and a license behind them to be able to stand behind my taxes still. Uh, but it turned this horrible, terrible task. I hate doing that. It takes me about a week to get all my stuff collected together, turned it into like a two or three hour task to do the whole thing. And so, and also on top of that, I have all this data connected and I can say, now help me analyze my spend. I don't usually pull down my personal checking account cause like just, it's just too much to look through, but I'm like, grab the, grab my emails, grab my personal checking account information, reconcile stuff, look at everything that's going on and suggest ways to improve it. Where can I cut expenses? Where can I, um, what are the things I missed? And I brought up a whole bunch of stuff. Um, so it's just like just a, just an easy win, a fun thing that it's really good at that kind of

1:15:30 Wrap-up + upcoming events

[1:15:29] stuff. So, um, anything, any last parting words before we end the call here? One last thing. Thank you, Andrew, for organizing this. I look forward to this hackathons. There's so much value that I am sure everyone else gets from it. So yeah, really, really thank you for the fact that for these thoughts, people usually charge and we get this for free when you organize it. So yeah, really thank you for that. Awesome. Thank you, Shubhash. Well on that note, that's a great note to end on. So thank you guys. Uh, we went a little past the hour about that. The last 20 minutes are a good one. Um, I'll send out a replay to everyone and then, uh, uh, of course the transcript too. So your AIs can chomp this and summarize it and pull out information for you. Um, and then, uh, we'll probably do another one of these things. Uh, oh, I'm gonna be at Accelerate next week. Um, I would love to meet up in person if anybody's at Accelerate. And then also Kanjan has a conference in LA, um, uh, um, top tier conference. It's October 7th. Can you tell us?

[1:16:38] October 6th to 9th, top tier. Yeah. We have 10,000 attendees. Yeah. All above, mostly TikTok shop creators and creators among area. Also like AI strategies like Andrew and also some big brand operators like from Anchor or from other big teams. So if you are interested, yeah, welcome to join us. Awesome. If you have a link, drop it for us, Can. I'm speaking there. So I'm excited. So, uh, hopefully I'll see some of you guys in Seattle next week and, or LA in two weeks or three, I guess three weeks. Um, and, uh, yeah, thanks guys for joining and we'll talk to you all later. Thank you. Thank you so much. Top tier league.com/summit. Bye guys. Thank you. Thanks.

Watch the earlier hackathon replays

Fable 5.1 vs GPT-6 Astra20 AI Power Users Compare Notes

What the Room Figured Out

Pick the model by cost per task, not the leaderboard

Plan high, execute lower

The harness matters more than the model

Real workflows, not demos

A person approves every change

Keep context clean and cheap

Chapters

Links Mentioned

Copy the Full Transcript

Get invited to the next session

Fable 5.1 vs GPT-6 Astra20 AI Power Users Compare Notes

What the Room Figured Out

Pick the model by cost per task, not the leaderboard

Plan high, execute lower

The harness matters more than the model

Real workflows, not demos

A person approves every change

Keep context clean and cheap

Chapters

Links Mentioned

Copy the Full Transcript

Get invited to the next session

Fable 5.1 vs GPT-6 Astra
20 AI Power Users Compare Notes

Fable 5.1 vs GPT-6 Astra
20 AI Power Users Compare Notes