§ 01 · The Hook
The bait, then the rug-pull.
Riley Brown opens cold on a solo shot with one number on screen — 2026 — and a thesis: the agent wars have spun out so fast that nobody, not even people in the bubble, can keep up. He pulls in Ras Mic for an 80-minute audit of who's actually winning the super-app race, why Anthropic's lead is slipping, and what knowledge workers should brace for next.
Who's talking.
00:00hostRiley Brown
00:32guestRas Mic
Where the time goes.
00:00 – 01:35
01 · Cold open + thesis
Riley solo-frames the four-month AI-agent firehose: OpenClaw's Mac-mini shortage, Anthropic's 50+ feature blitz, OpenAI's Codex surge, SpaceX semi-acquiring Cursor. Names today's guest Ras Mic and previews the full tour.
01:35 – 03:02
02 · Q1 shipping frenzy
They name Q1 'Anthropic's quarter' — January through March of nonstop feature drops. Cowork, OpenClaw craze, sold-out Mac minis, two-to-three-month wait lists. Ras Mic flags that Opus 4.5 was the real inflection point.
03:02 – 08:42
03 · Opus 4.5 inflection + Anthropic going too broad
Andrej Karpathy's 'never felt this much behind as a programmer' tweet converted the skeptics. But Anthropic then sprawled — Claude Code vs Cowork vs Schedule vs Routine vs Dispatch vs Remote Control — five names for the same primitive. Riley argues OpenAI got focused exactly when Anthropic lost focus.
08:42 – 12:53
04 · Codex vs Claude UX war
Codex is winning the bubble. Free usage, raised limits, smoother in-app browser, design mode, terminal-to-GUI continuity. Claude bans heavy users; Codex courts them. 'A great coding model is a great general-purpose knowledge model' — so whoever wraps the best GUI around it wins.
12:53 – 14:27
05 · Cursor's bet
Cursor was the innovator — Composer, the agents tab, in-app browser, sandbox preview videos — but their $200 Ultra plan can't compete with subsidized Codex/Claude compute. They are not a model provider. Hence the SpaceX deal.
14:27 – 21:45
06 · SpaceX–Cursor acquisition deal
$10B now, $60B opt-in acquisition. xAI has the H100s and Colossus in-house; Cursor has the GUI and distribution. The bet: xAI focuses purely on the model, Cursor becomes its consumer surface.
21:45 – 28:57
07 · OpenClaw and the agentic personal computer
Ras Mic walks through his sponsor-vetting workflow on OpenClaw: every inbound email gets researched (scam check, company profile, funding) and the agent only escalates the vetted ones. He frames the 15-minute heartbeat as the real unlock — agents that ping themselves get something close to surprise agency.
28:57 – 30:30
08 · Reactive vs proactive work
Riley splits the world in two: 'super app' tools you talk to (Codex, Claude, Cursor — reactive) and OpenClaw-style agents that talk to you (proactive). He says he no longer opens ChatGPT or Claude directly — he iMessages his OpenClaw.
30:30 – 33:12
09 · Keep agents narrow
The dominant failure mode is giving one agent 40 skills and 20 connectors. Treat agents like new hires — would you give a new employee 40 jobs on day one? Use one orchestrator agent that delegates to narrow sub-agents.
33:12 – 36:23
10 · Memory layer matters
Markdown won. OpenClaw's native memory is just .md files (agents.md, user.md, memory.md). Ras Mic uses Super Memory as a portable graph so memory survives nuking his instance. Obsidian's bet on plain markdown is paying off.
36:23 – 39:28
11 · Computer use gets fast
Opus 4.7's image resolution bumped to roughly Mac-screen dimensions specifically for computer use. Browser-use harnesses are appearing; Vercept got scooped by Meta then ended up at Anthropic. The takeaway: 'enjoy this brief period where you can still watch the browser move.'
39:28 – 41:06
12 · Who wins the super app
Manus was first; Cursor was first on the agents tab. Both proved you don't actually win by being first — you win by being best. Genspark stalled around $300M. The real prize is post-launch growth velocity, and Codex/Claude have it.
41:06 – 45:17
13 · The problem with Google
Google has GDP-level money, infinite talent, and the smartest model on knowledge — but tool-calling is embarrassing and team silos are vicious. Gemini, Notebook LM, AI Studio, Antigravity, Stitch — five product surfaces where there should be one. They're one model train away from being a real contender.
45:17 – 56:56
14 · Prompting and skills that scale
LLMs predict tokens — they don't understand. The English you use is the lever. WhisperFlow > typing because speaking is faster than thinking-while-typing. Skills should be domain expertise specific to YOU, not a downloaded pack. Build one skill by reverse-engineering a successful agent run — recursive onboarding.
56:56 – 59:39
15 · Integrations beat prompts
Context management > prompt engineering. The best 'prompt' for OpenClaw might be a Linear connector + a ticket title. Codex's Chronicle screenshots your screen every few seconds to build context passively. Aravind: the user is never wrong — everything that goes wrong is on the tool.
59:39 – 62:04
16 · All the muxes (cmux, tmux, dmux)
A small detour into terminal multiplexers. CMUX is built on libghosty, gives you a sidebar of terminal tabs plus a browser. The companies all push GUIs, but focused work still feels better in a terminal — 'I don't even know why, but a terminal makes me work on one thing.'
62:04 – 66:38
17 · Bold predictions
Coding got better but not the rapture everyone predicted — good engineers still vastly outrun vibe coders. The real shock is knowledge work: Ras Mic ran a 27-page contract through Claude instead of paying a $1000-an-hour lawyer; his friend did the company books with Clark Max and the CPA assumed he was certified. Knowledge work is about to get rocked.
66:38 – 78:19
18 · Agent commerce + image-gen scams
Stripe just launched issuable agent cards with addresses, spend limits, and human-in-the-loop approvals. x402 (Coinbase) and Stripe's machine-payment protocol race for the agent-to-agent rail. Image gen got so good it's invisible to Ras Mic now — he NBA-photoshopped himself in one prompt and his family believed it. Voice-cloning + image gen = a new scam wave coming for Facebook-aged relatives. He set up a paraphrase code with family.
78:19 – 83:01
19 · Permission to build now
Closing arc. Information is so cheap that what matters now is agency — actually doing the thing. They both reject the doomer line: even if anyone can build, most people won't. You don't need permission, a $4K camera, or a $20 subscription. Just start. Sign-off + handshake.
Lines you could clip.
03:04
"Opus 4.5 was the inflection for me. It didn't happen right away — it took a week or two when we all collectively realized, like, woah, this is good."
03:39
"I've never felt this much behind as a programmer — that Karpathy tweet was the final moment where the skeptics said fine, we'll use it."
06:55
"A great coding model is a great general-purpose knowledge model. They're all just files in a file system."
11:30
"Anthropic doesn't care about people in the bubble — they only care about perception outside the bubble. And they won that."
14:40
"The way I frame it is it's a $10 billion acquisition with a $60 billion opt-in clause."
24:53
"I'm more excited about this knowledge-work category than I am about the coding improvements. The models will get more intelligent — but this is where my time goes back to my family."
30:58
"If you came in on day one and listed 40 things you could do, I'd be like — wait, what are you actually good at? Just say coding."
34:50
"Markdown just won. The Obsidian founder bet his company on 'own your files' and now he's winning."
40:40
"You don't have to be first — you have to be best. Cursor was first on the agents tab. Manus was first on the super app. Neither of them won."
48:40
"The English you use is way more powerful than you think. The model doesn't think — it predicts the next token. So if you give it garbage, you get garbage."
63:57
"I had a 27-page contract. The lawyer's quote was a thousand dollars. I ran it through Claude and asked 'where can they screw me?' — it broke down every page. That's the unlock."
76:55
"I have a paraphrase with my family now — because my X account got hacked. If someone sounds like me and asks for money, ask them for the paraphrase."
81:20
"Information is so cheap. What matters now is actual agency — searching for the information, trying to do something, figuring it out. It is the best time to be alive to build."
Things they pointed at.
00:25toolOpenClaw
00:35toolCodex
03:38linkAndrej Karpathy 'feeling behind as a programmer' tweet
08:30toolClaude Cowork
08:42toolExcalidraw
14:13toolCursor Composer
14:30toolSpaceX/xAI
20:40toolConductor
21:20toolWhisperFlow
21:30toolClaude design
32:40toolSuper Memory
34:30toolObsidian
36:40toolVercept / Vi Computer
37:20toolManus
38:00toolGenspark
41:27toolNotebook LM
41:35toolGoogle AI Studio
41:40toolAntigravity (Windsurf)
41:43toolStitch
50:50toolPolja
56:20toolSERP API + SupaData (YouTube research skill)
57:55toolOpenClaw 'Clawhub' skills marketplace
58:50toolChronicle (Codex screen-recording feature)
61:20toolCMUX / TMUX / DMUX terminal multiplexers
61:32toollibghosty
62:40toolReplit clone via Daytona sandboxes
63:55productHarvey (legal AI)
64:30productClark Max (accounting)
71:20productStripe agent cards
72:00linkx402 protocol (Coinbase)
72:30linkStripe machine-payment protocol
72:40toolCrossmint
72:45toolNatural Pay
72:50linkagentcard.sh
74:40toolGPT Image 2
75:00toolSedance (video gen)
Word for word.
HOOK opening / re-engagementCTA the pitch metaphor analogy story
00:00HOOK2026 has been the year of AI agents. All of these agent companies have been releasing update after update, and it's getting very hard to follow. And today, we're gonna cover it all. In the past four months, OpenClaw has become the fastest growing open source software in history causing a Mac mini shortage. Anthropic has launched over 50 new features trying to build out their new claw desktop app. OpenAI has launched Codex, which is quickly becoming my favorite platform for both vibe coding and knowledge work. And now SpaceX is kind of acquiring Cursor to compete with all of them. Today, I'm joined by my friend and creator slash software engineer Ross Mike to talk about all of these advancements
00:42HOOKand what it all means. Ross Mike is one of the best educators on YouTube and has an inside perspective on what's going on in AI and software engineering in general. In this video, we first talk about what a super app is and why all the big labs are building the same thing. We talk about how Anthropic is starting to lose the lead on OpenAI. We talk about ClaudeCode,
01:04HOOKCodex, and OpenGLAW. We talk about computer use and browser use, and we also talk about agentic payments. And we end by talking about AI image models and the future of knowledge work. If you like videos like these, make sure to hit that subscribe button. We make videos just like this every single week. It helps me out a ton. Let's not waste any more time. Let's dive into the conversation.
01:30HOOKSo today, we're gonna be doing a little bit of a review. Right? It is we are four months into 2026. It feels like ten years has gone by. Yep. You know, q one, January, February, March. Now we're in April. Now we're in April. I feel like q one belonged to Anthropic. If you think about it, literally, it felt like every day from when they I think it was, like, January 10 when they launched Cowork
01:54Up until, like, the March, they were launching a new feature every single day, and it's just been so much to keep up with. During that whole period, there was the whole Open Claw Mhmm. Craze. Mhmm. Everyone on my Instagram feed, they were buying Mac minis. You know, I can't get one. I still can't get one. Yeah. And, honestly, there was people saying that, like, Mac minis are gonna sell out, and I was like, there's no way they're gonna sell out. They're sold out. It's really hard to get them. It it's hard to get them. You know, I've heard there's, like, two month wait list for them. In Canada, Toronto, it's three months. Damn. So, like, I don't even wanna buy one. Like, what's the point if I have to wait three months? Yeah. Might as well wait for the next version.
02:28Yeah. And so, yeah, we had that whole open claw craze. And then, obviously, you have, like, you know, some other companies coming in, like Perplexity Computer. And then now we have the tool that I've been having the most fun with, which is Codex released. And I think the common theme is AI agents are starting to get really good at using a computer. Yeah. You know, whether that is, you know, through computer use on codecs or they're just really good at manipulating files and controlling a computer. Mhmm. And so,
02:56yeah, I guess, what are your thoughts on twenty twenty six so far? What has been, like, the highlight? Talk to me. Yeah. So I think like, we were just saying this off pod. Like, opus four five was the inflection for me. I think that model and it didn't happen right away as soon as it dropped. Like, I feel like it took, like, a week or two when we all collectively realized, like, woah. Like,
03:17this is good. See, this this is my hot take. So if you, uh, if you look up at the board over here or I guess we can look here. Yeah. I think the inflection point was when Andrei Carpathi had that tweet. I've never felt this much behind as a programmer. You know, all the developers who were kind of skeptical about AI tried it. I think that was, like, kind of, like, the final moment where all of the skeptics are like, fine. We'll use it. When did Opus four five come out? I'm pretty sure, like, early, like, December. Like, maybe December. November or December, I think. Yeah. So, like, there was a time where, like, again, I you know, Opus four five dropped, and I think because we're both content creators, like, I messed with it, but, like, I didn't really get it. This tweet did help a lot, but I I don't know. Like, at some point, I think I was using it with cursor,
04:00and I told it to do a bunch of things. I'm like, yo, this model gets it, and it's fast, and it feels great. And ever since then, to your point, Anthropic, like, q one is theirs. They just back to back to back to back, like, nonstop, like, Cowork dropped. You know, I knew they were winning when people on Instagram were like, oh, everyone's using ChadGBT,
04:21but the secret tool that's gonna 10 x your business is Claude. Like, when I started to see those videos Yeah. And then, like, friends of mine who are teachers and stuff started to talk about, oh, I used Claude Cowork. It felt like, yeah, they had just crushed OpenAI, but I think the last month, OpenAI has Has has caught up. Yeah. We'll we'll get to that in just a second. I think it's useful to, like, make this tangible and and kinda show people because, in my opinion, they went
04:46too broad. Around late February, early March, there was that announcement where I I forget her name. She, like, runs OpenAI's operations. Yes. I know. You know who I'm talking about. We'll show it on the screen. She came out and said, like, we're refocusing the company, and we're putting all of our efforts into
05:02Codex, the ORA super app. That was, the news headline, and they're basically gonna combine Atlas Yeah. And all their other efforts, which is GBT Image, and they cut Sora, which I thought was a great decision to just 100%. And so so they were the unfocused company, and Claude, that you had this perception that Anthropic was super focused. Because they're all about coding models. They're about video coding
05:24models. Nothing. Right. And so they had this terminal. They had, like, the terminal interface that people loved. Right? And I would use it in cursor, and then they just started launching a bunch of stuff. So, like, while Claude or Anthropic was kind of unfocusing, OpenAI was Started to focus. Yeah. So, like, now, like, look at this. So, like, there's Claude co work, which is like Claude code. Mini Claude code with a bunch of restrictions. Basically. So it's like a similar technology, but you have to learn them separately. Yeah. And then with that, you can use what's called dispatch.
05:54But if it's Claude code, you can use the remote. So you and you it's not a it's not a thing you can use here. You have to type in slash remote control. Yeah. And then you have to, like, set it up, and it doesn't work in the desktop app. You have to do it in the terminal. And,
06:08like, from my perspective and then so, like, you have, like, routines, but in co work, you also have scheduled. Like, there's so many
06:18it's just like It's the same thing. Just push them all together. And so I told you that, you know and people are saying, like, oh, look, you're getting paid by OpenAI. I'm not. Like, OpenAI I wish my boys getting paid for OpenAI. Like, I was I told you. I was like, I want a tool like Claude Code that can build artifacts. Yeah. And I wanna be able to switch what I'm doing here. I wanna be able to create a new task. On one of them, I wanna be able to create a doc. I wanna be able to create a presentation. I wanna be able to create Excalidraw. I know you're you love Excalidraw. Excalidraw. Codex is so good at Excalidraw. I saw your I saw your team. Yeah. And so I just wanna be able do everything in one platform. And then lo and behold, like, basically,
06:57OpenAI releases that exact thing. Right? And then now so there's not automations and schedules And or dispatch and remote control. They're just gonna release a Codex app pretty soon. So, like, you agree that they're kind I 100% agree. I think, like, there like, I've noticed, especially with the AI labs, there's the announcements you make for people, and then there's the announcements you make for the headlines,
07:18which end up allowing you to raise a lot of money. Mhmm. I think Claude, like Anthropic, showing the iteration speed of how much they were shipping brought them a lot of attention. I think PR wise, it was a net positive. Right? Sure. And arguably, maybe that helped them raise a lot a bunch more money. Right? Like, if you're not in our space, which is a very small bubble,
07:37genuinely outside looking in, it looks like Anthropic is winning. Like That's probably true. By a magnitude of a thousand, not even like a small margin. By a magnitude of a thousand, every single person that I would say is quasi technical that maybe, like, used to jailbreak iPhones or is, like, hip with these type of things, but it's not, like, in our space. I am telling you they think Anthropic is winning. So in terms of, like, the outside bubble PR,
08:01uh, stint, Anthropic is winning, and the shipping speed helps a lot. Now to your point, they generally are using Cloud Code. I believe it because why the heck do you have routines, but then in co work, you have schedule? They're the same thing. Right. Right? Like, it just feels like different teams are spinning up Cloud Code instances, letting it rip, and they're shipping it into production. Sure. But the but their argument is that, like, oh, so but you have marketers and stuff, and you have business people in your company, and you don't know, they don't know what a terminal is. You don't wanna give them full access. You don't have to, though. Like, literally, like, everything I see on Cloud Code and Cowork can be one app. There's no reason why there should be, you know, separate things. Maybe you have, like, a toggle advanced and, like, you know, something like that. That's exactly what OpenAI has. So if you go to settings here and you go settings,
08:45they what they do is they have it for coding, and this is the meme that Sam often posted where it's just code maxing and, like, Excel mocking or something. But, like, you just switch it, and all it is sense. Is a different view on the same underlying, like, abilities Yeah. Which I think is great. Like Makes sense. For the people who do everyday work, you don't need to see what a git work tree is. Like,
09:07what even is that? And so, like, it feels like and Codex now, like OpenAI, I feel like they're trying to win people in our bubble. Mhmm. Right? They're increasing limits. Like, every time and let's be honest, Anthropic is being a little mean. Right? Like, they're banning people. They're like, oh, you can't use our things. And I understand it's costing them a lot of money. I don't think they're handling it the best way. OpenAI seems to be doing the exact opposite. Free usage. Free this. You don't like this? We'll fix this. Right? So it it it feels like OpenAI is listening to the people in the bubble, which I e is making the product better, and then Anthropic doesn't care about people in the bubble and only cares about perception outside the bubble. But but they won the perception outside the bubble from the bubble. Percent. Okay. And that's the thing because they started off as the focus. Like, think about it. If you remember the keynote, dev day keynote of OpenAI, they had, like, four other things that they had launched aside from a new model. The agents, the Zapier killer, the the dragon drop, the who
10:01uses that now? No. No. Literally zero No. Actually, zero users. I believe it. Yeah. I believe it. So, like, they were very wide and, like, there was even this clip of Sam. He was like, you know, like, don't try to come against us. We will steamroll you. Sure. And then Anthropic came and steamrolled OpenAI. But it's funny now they're doing the opposite. So Anthropic was doing what OpenAI was doing,
10:23and OpenAI is doing what the Anthropix is doing, but there's a difference. Both are close to IPO. Sure. So, like, does it even matter now? Like, from a, like, financial success, we've made a perspective. You can never see behind the curtain and see what their incentives are. Yeah. You can never see how how low they actually are running in money, like, badly they need to raise the next round. So it's hard to see. I just look at it from a pure utility standpoint. Yeah. And, like, for me, one of the companies is, like, shipping really useful stuff. OpenAI is is cooking. It definitely feels like, uh, OpenAI is cleaning up the product. They're cleaning up house. They're centralizing
10:57focus. And I think the one understanding everyone's come to, which makes sense, is a great coding model is a great general purpose model. Because I think there was this theory of, like, oh, we're gonna have models in specific niches. And, of course, I think for, like, medicine and biology, like, you have to have a separately trained model, etcetera, etcetera. But, like, for general knowledge work. Right? Create this Excel sheet, scrape this data, do this, do this math. It just so happens that a great coding model is a great general purpose knowledge model. So Anthropic one in a sense early on because they were building great coding models, and then they realized, oh,
11:30this thing can create spreadsheets and can analyze data, and it can do a lot of stuff. Right? But just like code files, they're all just files on a computer. Literally. Right? Yes. Like, it's that's literally it's it's just files in a file system. So, like, now everyone's building trying to build the best coding model, and now they're wrapping it around GUIs,
11:47for anyone who doesn't know GUI graphical user interface so that you can do different things with the models. And it just so happens where I feel like this matters a lot more now because the models are all great, but how you use the model, the tools you use with the model matter, and Codex is is 100% winning this arena. Yeah. And so the okay. So this is a good kind of segue into, like, what I've been talking about, which is super app. And I got that from the OpenAI headlines. They're releasing a super app, and I think it's I think of a super app as
12:14a coding model wrapped in a nice GUI that can do coding work and knowledge work. Mhmm. Because, like, what else is there in a in a business? Right? You just have coding work, knowledge work, and then the stuff you need to do in real life. Yeah. And that's what we're seeing. And one of the advantages I think OpenAI had is they had a big team working on Atlas. Yeah. And in that, they were working on browser use. And so, like, this here is a browser. By the way, if you try to use the browser in Cloud Code, doesn't No. That's okay. You can't type into it. Like, here, you can actually go to like, if I wanted to go to, like, twitter.com,
12:46like, you can actually go to it. Now it doesn't keep me signed in, but, like, I've read between the lines and they It's probably They're it's coming. But you know who's ahead of all this, though? Cursor had this early on. Yes. Cursor Cursor see, this is this is a talk Cursor was an innovator. Okay. Let's let's talk about Cursor. So the reason why people wouldn't wanna use Cursor over these companies, right, like Claude or OpenAI is because these companies are heavily subsidized with the best models in the world. 100%. Cursor I'm a huge Cursor fan. Like, before,
13:16like, when Claude code was on its come up and co didn't Yeah. Codex didn't exist to my knowledge. No. They they didn't. You know, when cursor was the first tool that I was using, you know, I I mean, I started off a year and a half ago You blew up on cursor. Right? No. Well, I blew up first because I was copying and pasting code from Claude desktop app into Replit. That's what got me into vibe coding. I was one of the first people to do that. But, like, immediately after that, found Cursor, and Cursor was the first Cursor Composer
13:43was still my favorite feature, not Composer, their models that they're doing now. Yeah. The Cursor Composer where you would type the app you want and it would make the edits. In in line edits, like, it would pop up on the screen and it was like Well, now that's trivial. Right? Like, everyone's doing that now, but, like, at the time, it was That was innovative. Yeah. Because, like, all they had at the beginning, which was still good, was, like, they had the tab feature, which doesn't really benefit me because Yeah. What am I gonna autocomplete? I don't even know what to type. And
14:07then then they had the chat, which would allow you to chat with the agent about the code, but it didn't actually make the changes. Right? Yeah. Because you have to go, like, command was it command z or command k? There was a many. Too many. There was one yeah. There was three commands you had to learn how to use. Yeah. Yeah. Yeah. Sure. And but that tool,
14:22it honestly felt like it went cursor, Claude code, codec. Like, it's kind of in that third iteration. The cursor dominated that first round. I mean, that's why they just got semi acquired by x AI Yeah. SpaceX for $60,000,000,000. So, yeah, let's talk about Cursor. So what do you what do you think of the state of Cursor right now? So to your point, Cursor
14:42like, the acquisition is a w. It is a big win because Let's let's talk about let's let's unpack this. Can you talk a little bit about the the So they basically got a deal with XAI that says, like, we're gonna we're gonna give you 10,000,000,000 now, and at some point, we can buy you for 60,000,000,000. We're gonna partner up. X AI has I think they have the, like, most compute out of all the labs in house. Right? OpenAI has to partner. Anthropic has to partner. X AI has the h one hundreds and h 2 Hundreds ready in house. So they're gonna train the model. Cursor's going to be the GUI that they're going to use. Cursor's biggest,
15:16uh, weakness is that they are not a model provider, and they don't have the benefit of subsidizing the price. For example, with a $200 Cloudmax subscription, you're actually getting $5,000 worth of compute. Sure. Unhealthy. For now. Yeah. They could and they're slowly rug pulling as well. Right? Oh, you can't use it with Oh, I I don't know. Rug pull let it's not rug pulling. They they're they're they're tearing the cost out. Yeah. But it I will say it it is, like, especially if you're an OpenClaw user, it was definitely a shock. Like, I think there's two sides to the problem. First and foremost, I think we got spoiled with a large amount of compute for a very, you know, decent price.
15:52But I guess it was hard for them to grow. Like, who's paying $5,000 a month? Right? Like, that's that's insane. Like, that's more than almost most people's rents. Right? So it makes sense. But going back to Cursor, the issue with Cursor is Cursor is probably the most expensive subscription. Right? The $200 ultra plan and the $200, whether it be Codex or Clog, it it doesn't compare. Doesn't compare. Right? Unless you use their composer model, which the models don't compare. Yeah. Like, it's a decent
16:19feat. Like, they basically built on top of Kimi 2.5, Kimi k 2.5, which was smart, but they're not going to win in the model arena. And I think they know that, and that's where the SpaceX partnership slash Aqua The way that I frame it is it's like they it is a
16:37dollar opt out clause. Basically. Yeah. Yeah. Basically. And to me, like, it makes sense. Right? Like, Grok is should just focus, like, XAS should just focus on the model. I don't think they have time between model, and then they have x, and then they have this, and then they have that. Like, it's good that they're focusing on the model. Cursor, like, for example, like, all the GUIs, like, Codex's GUI. I'll be honest. Cursor did it first. If you remember the agents tab Yeah. Like, the OG agents tab that everyone sort of hated in the beginning, that's basically what this was. Yes. And I I even feel like they were too ahead of their time. Right. Right? So, like, these guys have proved in terms of product. Like, they've they've made the right decisions. They've sort of guessed where the industry wanted to go, and they've led that path. They just can't win in the model front. So, you know, kudos to them. I mean, their in app browser is amazing. It's fantastic. It has And it persists. And, yeah, it persists. And then also,
17:28Anj, my cofounder, he uses their that has, like, some sandbox feature where they, like, test your app, and it'll, like, send you a video. Yeah. That's insane. Yeah. And they're betting big on that as well. Yeah. And so and so you you think, like, they might they're probably closer to creating this super app than, um, than Claude at this point. It would be difficult for them because, right, at the end of the day, what powers the super apps is a coding model. Right. So if we already have the coding model, it's just a matter of who's going to create a good enough GUI. Sure. And, unfortunately, Anthropic is not great at product.
18:00You think yeah. Yeah. You think so. Like, right now, at least, like, I don't again, we don't know the internals. We don't know whether people are spread too thin or they're focused or they're this. Like, but from my I mean, from your original usage and my personal usage, like, you can start a thread with Claude, and if you refresh the app, it breaks sometimes. Sure. Right? Where with Codex, even though, like, yes, it's not like I've had some bugs here and there, it is the smoother app. And with Cursor, like, I've never had Cursor. Like, there was a period where Cursor had issues, and they fixed a lot of it. I've never had Cursor go. Yeah. Cursor is a great software, and I've used a lot. I've used their new interface, and I think we're we're seeing, like, an identical use. Like, we're
18:39all of these companies are converging on I think Conductor did it first. I don't know if you ever used Conductor. Yes. Yes. They were the first person Yes. First company to put this side panel where you basically have, like, your directory and then your threads. You can do multiple threads at the same time. It's very easy to switch. That's what I love about Codex Yeah. Because I'm literally just using WhisperFlow,
18:58typing in my prompt, and I'm like, oh, I have another idea. Like, let's go explore it. New new chat. WhisperFlow. I just speak into it. It's like a little Jarvis, and then whatever you want to show up on the right will just show up when you need it. And that's why, also, on top of all the features we talked about, design mode or Yeah. Cloud design. Yeah. So I've I've I've messed with Cloud design. I will say it feels like the most refreshing one because it was different
19:22out of, like, all the things we've seen the last couple weeks. And also with design, what's interesting is I don't think it was made for designers. I think it was made for people who have taste but don't know how to design, but you can, in a way, communicate it. Right? Sure. So you can upload images. You can, like, share, like, um, Figma inspo. Like, you go to, like, the Figma I think it's a marketplace so you could see people's designs.
19:44And, like, it will, like, tell you, okay. Like, what kind of design are you looking for? Uh, and you could be, give me some suggestions. I was do using it yesterday. It was like, oh, aquatic greenery with a moss. This like, it's very, like, particular, and the output was pretty great. But this goes back to my earlier thesis. A great coding model is a great general model. Sure. You just need to wrap the tool around. Design tool is literally a coding tool. Mean, it's code. Yeah. It's and so the way that I see it is, like, all of these design mode should be like, your chat should enter design mode. It should be the The user should not go to Cloud Design. By the way, it's only on their web. It's not enough. How how do you use Cowork? Oh, you can only use Cowork on the on the desktop app, but you can use Cloud Code on the web, but you can only do it with the git By the way, do you know how you export your cloud.ai/design
20:31like stuff? No. I You copy you copy something, and then you have to drag and drop it into claw code. So it's like, they're not even connected. Why would it not be like, why wouldn't you just open up a new project, create this, and say, I want you to design this? Right? And then the the the artifact that pops up on the right is design mode. Like like, the the browser enters design mode, and and maybe it's like a fixed layout where, like, you have the phone frames. It's a mobile app or, like you know what I'm saying? I hope they're I hope for our for their sake that they're rebuilding the entire app and all this stuff is gonna happen because it's it is such a fumble. Right? Like, the fact that I have to go on Claude design
21:07on the browser, and then I have to copy that on Claude code. Right? But then if I wanted to use co work, I have to open up the app. What if you wanted to turn whatever you coded into a investor deck? Yeah. Like, I wanna no. But have export and put in a co work because the Cloud Code it's not even a shared session. Right? Like, it's, like, it's different context. Have to explain again what exactly it is. Sure. And and, like, it's hard to to get the
21:31co work session to look at other file. Like, you can't it can't interact with any So other it's it's a weird situation. One thing I wanted to talk about before we get into, like like, I wanna talk about, like, what your favorite tools are right now and, like, some more tactical things. I do wanna talk about Open Claw because I love Open this Open Claw was
21:51something I was skeptical about at first. I thought they had a very the early movers to Open Claw were people that talked about other things that aren't I thought there was just a lot of, like, scams going on in in Open Claw, so I was skeptical at first. And then I don't look at Open Claw as just, how good is Open Claw? I look at it as an idea, which is basically an AI agent that runs on a computer, full control over a computer Mhmm. That has channels.
22:16It has what I file. So it basically has, like, these files that it keeps updated as you use it more. Yeah. So in theory, if it worked right, the more you use it, the more it knows about you, the more it can do for you. And then they made it so you could create skills with natural language. Yeah. So, like, there's a lot of things in there that has never been done before. As a team, we've spent a lot of time thinking about this, like,
22:38it's a whole new paradigm. And so I'm curious, like, are you using Open Claw or Hermes? Or Very much. I'm very much a Open Claw. Like, I have, like, a forked version, and I have So how do use it? Let's talk about that for a sec. Before I I get into how I use it, I wanna agree with you on one point where it's like, in theory, it wasn't anything, like, new,
22:57but, like, different new things got put together, and it just, like, you you kinda felt the spark. To me, the biggest thing was the heartbeat. Right? The fact that every fifteen minutes, the agent gets pinged. It has this prompt where it looks at all your stuff, and then it can message you. Because when when you think of humans, like, you and I have a heartbeat as well. Right? It's every second or microsecond, whatever it is, and it's like our heart's pumping and, like, you and I get to think and do things. Right? That fifteen minute heartbeat,
23:24like, I don't wanna say almost, but kinda simulates that. Right? Where, like, based on all the stuff you said, it has the user information, all these markdowns, and all this context about you. And if, again, the coding model underneath gets better, if the general if the model underneath gets better, then and it has memory and has context, it should be a net positive whether it's your business, your personal life. So how do I use OpenClaw?
23:45Um, you know, I'm sure as a creator, your channel being bigger than mine, you probably get hit up every single day. Sponsor this, sponsor that. Sure. Sure. I don't look at any of that. Right? So I have, uh, an email, a sponsor email that my agent has access to. Every time an email hits, it researches three things. First of all, are these guys a scam? Meaning, is the email a scam? If I get an @gmail.com,
24:06I'm sorry. Maybe you're legit, but I'm not even gonna consider it. And then if it gets a company email, we research the company. I actually give it an exact prompt. Research the company name on Google, add, Reddit scam. If anything pops up, I don't even consider it. And then let's say it's legitimate, then I'll it'll do some preemptive research like how much money have they raised, how many followers do they have, and all this every single morning, I get a report of Sure. Top of the money. So I'm looking at this report. I'm not even looking Do you have it respond to the emails directly? So Okay. Once once like, and that's, like so that's workflow number one. Workflow number two,
24:41Notion is, like, my database. Sure. Once I've approved, like, which ones we wanna talk to, the first email that gets sent is just pricing. Right? This price, that price, whatever. And then, again, the agent checks whether they've responded to it or they're willing to negotiate. At this point is when I take over. So when I take over, not only have they been vetted, but they're most likely going to be a sponsor of the channel. Sure. Right? And this is, like,
25:06maybe per day, like, an hour and a half, two hours work, but imagine it aggregated. Right? Like, aggregated over a week, over a month, over a year, like, how much time I'm saving, how much time I could spend with my family. Like, these things like, I'm more excited nowadays on this type of stuff than the coding improvements because this the models will get better, more intelligent, will write more code. But, like, this knowledge work category where you and I spent time on tasks that can be done by the model
25:31greatly, perfectly, um, that's the, like, sort of spark that OpenClaw ignited in me and excited me with. Sure. Yeah. Okay. So first of all, that's a great use case, and I use it in a similar way. I will say, when I was a creator, like, most of my time was spent just, like, doing content creation. Hiring a manager was the best thing I ever did. I was always like, oh, I don't wanna give up, you know, like, 10 to 20% of the brand deals. But, like, when you get someone to negotiate the deals, you end up get making more. Yeah. Because the you know, especially if you get a good manager.
26:02But it is pretty interesting because AI will never turn off. It can research enough, and it it can negotiate in however you want it to negotiate prices and etcetera. So I think that's such a good use case for it for individual creators who don't wanna deal with it because you're right. It can it can take up all your time as a creator. Like, yeah, if you're, like, a mid sized creator, like, on whatever channel, you're getting hit up a lot. And I think a lot of companies are starting to realize, like, organic social, like like, working with content creators is probably a great spend ad spend for them. So it's like a lot of creators get hit up a lot, and it's just it's too much to cycle through. And I and me, I got a full time job. I run a product studio. Like, there's too much going on. I don't have time to read. You know? I got a minimax email the other day, and it was a scam email. It was
26:47dash minimax dot l o whatever. Yeah. And if you're just manually going through, you're gonna click on that. You're gonna respond. Yeah. For sure. And the link, I clicked on the link, and their form to fill out asked me to authenticate with x. Right? So So you're about to get fished. Yeah. Yeah. I was basically. Right? Yeah. And this was, the one time I decided to go through my emails. I don't wanna wait till, like, the next day for, like, OpenCloud to catch it for me. But, like, the literally, the next day, OpenCloud's like, yeah. Scam email. Right? So, like, there's just these miniscule things that might not seem like, oh, it's only thirty minutes a day now already, but on aggregate, when you add all those things up and, like, the speed at which it does it, like, to your point, it doesn't get tired. It doesn't have a bad day. It doesn't get moody.
27:24That was, like, the spark when I'm like, okay. Like, these the open clause, the Hermes, like, these agents are And and what's interesting about sorry to cut you off there at the end, but it's interesting watching Claude trying to add the same features that open clock had. Schedules. Routine scheduled dispatch,
27:44and, like, they had, like, similar animations, the remote control, telegram that you can do with slash remote control slash telegram or however you do it. Yeah. It's been interesting watching them do that, and now they're, like, they're gonna have to start copying codecs Yeah. Because it's kind of what people want. And it's
28:02seeing them kind of get caught up in that and then also OpenAI semi acquired OpenCLOTS has been, like, a super interesting Yeah. So the the like, basically, like, it's so they hired Peter. Sure. And OpenClaw is run by a foundation,
28:18but I'm again, I don't have insider info, but I'm like, I would bet on OpenClaw versus a Hermes. Like, a lot of people, like, OpenClaw Hermes. Uh, for context, I haven't used Hermes. Right? So I I neither of them could win, like, the end of the day. So the the like, it's interesting OpenClaw OpenAI doing that is sort of confirmation that everyone's gonna go build that now. Right? Because OpenOpenAI
28:39is gonna build it and users are excited. Of course, Anthropic's gonna build it. Right? So I'm glad OpenClaw exists. I do use OpenClaw. Will a super app will codec super app destroy all of it? Maybe. We'll see. Yeah. But just this idea of, like, these mundane tasks in your life being taken care of by AI, I'm so bullish on. Yeah. Yeah. Yeah. And I think I put it into two categories. I think there's the super app, which is a it's a very react it's more of a reactive tool. Like, it's not a proactive tool like OpenClaw. Mhmm. As you there there's something about this interface, and I I I'll use the the Codex one, but,
29:13like, there's something very reactive about it. I know they have automations built in, and you can proactively create automations. But that's not what the first thing you're going for. Right. So so I think I think for me, there's two types of tools that I use. I either talk to a claw or a super app. Right? I never talk to, like, chat GBT or Claude anymore. Like, I'm gonna open up iMessage, and I'm gonna message my open claw Yeah. Because it has the context of me. It has heartbeat, so it's checking in with me. When there's an important email, it just sends it over. It's like, yo, Riley. Like, you have to do this, bro. Like, oh, rent's due tomorrow. Like, you gotta get this done.
29:45And so, like, I I'm either using a claw or an agent running on a computer or sandbox Mhmm. And then these tools on the computer. Right? This is Interesting. So that's fair because, like, I don't go to Codex for proactive work. I go for react. Like, meaning I'm going to be working with it. Yes. But my claw,
30:03I just wanna tell it once, and I just want it taken care of. Yeah. And, you know, there's certain things. Yeah. It but, like, the cool thing about a really good employee that if you were to hire human employees, they will surprise you with valuable stuff. Agency. Right? Surprise agency, but, like, you'll wake up one day, and they'll be like, hey, Riley. Like, I know you're doing this. Like, I made this report for you. Like, I if if this is useful Yeah. Like, let me know. I'll keep doing it. Yeah. And you're like, oh, that's really useful. It's like OpenClaw actually has that potential.
30:30Now, you you do have to kind of be proactive with OpenClaw at this stage. Right? Like, it's, um, and what I always tell people when they're setting up Open Claw is to, like, keep it narrow. Like, the more goals you give it, the more broad you are. People put, like, fifteen, thirty skills, like, 20 different connectors and all that stuff, and it's like, oh, none of it works. Yeah. Because, like, even imagine if I if I got hired here and I joined the Vibe Code team and, like, off rip, you give me 40 things to do. I'm quitting the next day. Like, issue is reverse. Right? If you came in and told me you had 40 skills and you owe you I'm like, hey. So what are you good at? I would much rather you say, I'm great at coding. Yeah. And then then versus you just listing, oh, I can do the dishes. I can go and do list off all these things, and it's just like, it's overwhelming. What do you get it to do? Yeah. And so it's like,
31:15I wonder if there's a world where there's many agents. They each have their own, like, niche. Like, I don't know if it's is it in a group chat? Like, I'm trying to figure out, yeah, what the interface for agents are in the future. What do think? Like, so, uh, from my, like, experimentation, I will say, like, the
31:33main agent with the ability to deploy sub agents for specific tasks seems to be the best one, especially with OpenCloud. Because if you give the main agent a super big task, um, the way it works is, especially the gateway, is, like, it's a queue system. So if you tell it to do one thing, it's basically stuck doing that one big thing. You can't tell you can't continue to communicate with it. But but this idea of, like, you have one orchestrator agent,
31:58and you're not even communicating with the sub agents. You tell the main agent what's going on. He tells he or she or whatever you wanna call it, tells the sub agents what to do, and they report back to it. That seems to be, like, the great sweet spot. The only issue I have with betting on this architecture is we don't know where the models are growing. Right? We don't know if, like, there's a new training method, if there's a new style. Like, there's kind of the models the model labs have the benefit
32:22of building great tooling because they see where the models are growing. It's funny though. Codex is the only one doing that right now. OpenAI is the only one doing that right now. But as of yet, like, the way I'm using OpenCon, other tools is, like, I'll have a main agent that sort of orchestrates everything that deploys sub agents. I don't communicate with the sub agents. I communicate with the main agent because I want one agent to at least have context That's like the where everything have Sub agent is an agent that is delegated by an orchestrator. Yeah. Right? And I I've seen, like, people, like, have, like, multiple
32:50main agents and, like, multiple teams. Again, maybe I haven't tried it enough. I I don't like, I haven't found great success with it, but it feels like having this one buddy, this friend, this agent that I could communicate with. And that person, that agent I can't believe I called it a person. But that agent is responsible for, like, delegating and making sure everything is done. That seems to have had the great, um, I've had a great experience with that. Memory is a big issue. Um, I don't know if you're familiar with super memory. I've tried building my own memory, um, thing. It kinda worked. It kinda didn't I don't really care to maintain it. There's certain things, um, I'm just willing to pay for. Super memory has been really good to me.
33:27Super memory is a tool that you can add to your OpenCLOS setup. Yes. So they have a simple plug in. It's basically, like It's like a database? A memory service. Right? Like, they, like, have graphs and, like, all this stuff. The cool thing I like about super memory, and I don't know if this is a product vertical that that's going to win. Uh, the founder's pretty cool. What I like about super memory is, like, my memory exists in that cloud, but then I can connect it to another app. So, like, I had my open cloud running on Hostinger,
33:53uh, because they were running the crazy deal. Sure. Sure. Then I hated running it on Hostinger. Why? So it broke. Like, it was slow. Like, I built, like, my own little custom wrapper to deploy it, and it took me, like, a minute and a half, whereas theirs will, like, take forever. And I'm sure, like, there's a lot of constraints and all this stuff they have to deal with, not crapping So on this is interesting. So when you create an Open Claw, right, and and it's just like a file system. Right? You you have
34:17just like code files, you have you have literally code files, and then one of the files is like you have like your agents dot m d file and your your User dot m d. Yeah. And it it and by default, it stores the memory in those somewhere in those files. Yeah. But what you're saying is you have this service that
34:35super memory, and there's others I I remember zero Yeah. And there's a couple others. Yeah. So this is kind of stored in this location where you could, in theory, have other agents using the same memory. And exactly. Right? Like, there's a part of me that's, like, I want, like like, you can imagine a world where agents are the product, and, like, every company has a specific agent or maybe one agent rules it all. I don't know.
34:57But, like, I own my memory, like or I can take my memory anywhere with me. The issue with the OpenClaw native memory system, it's basically just markdown files. You have a memory dot m d, and then there's a folder where every single day, all the things that you said with the agent are stored Yeah. And then it'll contextualize that. So when I nuked my hosting your instance, the memory died with it. Right? So but if I have super memory, like the plug in setup, I took my memory with it. I have it running in my local machine. I can deploy it on a separate service. Right? So that was another thing as well where, like, memory is very important. I'm seeing people use Obsidian. It's you used Obsidian?
35:34I I I've made a video, uh, using Obsidian. I was super big on it, like, before like, Claude Claude was still in I was using Claude in a terminal with Obsidian. Yeah. Because Obsidian is just Markdown files. Files. Yeah. . Yeah. Markdown files just won. And, yeah, the the Capano guy, the founder of Obsidian on Twitter, he's he's been saying it for a long time. Yeah. He's like, own your files. He bet he bet he bet his company on that, and now he's winning. Oh, yeah. So, like, OpenClaw, like, to just summarize, is very interesting.
36:02I don't think it's the only one to exist. I definitely think OpenAI for the will of good pre, uh, PR will probably continue to support it, which is great if you're an OpenCloud user. The last couple of updates have been great. Um, but I think that's where the world is going, like, these purposeful agents that can do these mundane tasks, and I love it. Yeah. Yeah. Yeah. It's just agents are starting to use computers.
36:27And Yorun, he tweeted the other day. He's like he's just like, enjoy this brief period of time where you can watch the browser use. Yeah. In a way, like because soon it's just gonna be so fast that, like, it's gonna be too fast for us to even follow. Yeah. And so it's it's this weird time where, you know, it's the they're they're speeding up. I don't know if you've used browser use. Have you used browser use? Yeah. Yeah. I had it build a chess. So in one prompt, this is in my Yeah. Long hour and forty minute Codex video, I had it build a chess board,
36:55like a chess game, and then play itself. Yeah. And it was just, like, moving fast. The first time it did it, it's like it for move for move, checkmated itself, and I was like, no, play for longer. And then it was like, okay. It just clicked on the restart button and then just played itself again really fast, like, faster than I would ever play anymore. And I was like, wow. Have you seen their repo? They released a repo where they have, like, a browser use harness. Yes. I did. Healing.
37:16Oh, what what are you talking about the thing that the guy made a video on the other day where you can, like, oh, that's real time. Yeah. That yeah. Yeah. Yeah. Yeah. Yeah. So, like and that's an another interesting tool, like, with agents. Now we're entering an arena where, like, yes, we know the agents can write code, but what if they can write code
37:31around their harness? Like, what if they can build themselves tools that persist? And, like like, it's like, imagine, like, a robot that can build extensions of itself. Right? So, again, the best coding model is the best model for everything. Yeah. Yeah. Yeah. And that's what everyone's betting on. Computer use, I don't know if you've looked at, like, the recent model drops. That's been a big focus for a lot of them. Even Opus, like, Opus four seven had a huge computer use boost because to your point, we've realized, oh, the agents are capable of doing all these things. Let's give them a computer. Let's give them the ability to do x, y, and z. Well, they also need training data on,
38:06like, computer use 100%. Like, they need this because, basically, the way it works is, like, it takes a screenshot. It feeds it to the LLM. The LLM is like, oh, click on dawn. Four. Yeah. Like, literally. Yeah. And then it'll click, and it'll take a screenshot. So, like, it's pretty, like, slow, but, like, it's going to get faster and faster. And you notice, Opus, if you looked at their new model,
38:25the resolution at which they can analyze images increased to, like I'm pretty sure it was, like, the exact dimensions of a normal macro. And that's because they're betting computer use. Yeah. And so it's very interesting. So then you have precision. And then so, like, now, you know, sometimes you're waiting ten seconds for the mouse to go from the the menu button to the submit button. But it's like, what if it's a tenth of a second? Yeah. And they acquired another company called Vercept.
38:48I didn't even know that. Yeah. Which is they created something called Vi Computer. It never took off because it wasn't it's not there yet, but it's like one of the best computer use guys. And they probably just hired they just wanted the team probably. Yeah. Well, Meta so this guy, this is like a side story, but it's really interesting. So Vercepta is this company. They're out of Seattle,
39:06maybe, and their computer use, the guy, his name's Matt, who I got on a podcast, like, a year ago, but then we ended up not releasing it because Zuck came in and bought him for 200 mil a year. You know, you remember when he was, like, buying the scientists? Yes. Yes. Yes. Just scooped him up out of the company, and then the company ended up going to Anthropic. Yeah. So clearly, meta is quite interested in that game too. They're just And they they tried to buy Manus. Yeah. And Manus it down. Manus was the first super app. We didn't even talk about Manus. And that's the thing. What's quite unfortunate is the first ones didn't really capture the huge upset. Again, Manus Manus captured. Dude, do Jen's spark is 100%.
39:45Actually, I don't wanna say Jen's spark is not good, but they're at, like, 300,000,000. Right? Yeah. Manus won. But Sure. It's like, to me, like, the level of how big the codec super app and Claude when they get stuff together, like, how high of a growth it's gonna be, I don't think it's going to be com it's I don't think it's comparable to where Manus is at right now. Same with Cursor. Like, Cursor was the first one with the agents tab. And I guess this is me sort of, like, inspiring myself because I don't think you have to necessarily be first. Sure. You just have to be the best. Yeah. I mean, you have to to a degree. You have to be early. Yeah. You can't simulate. Cursor you say, like, cursor hasn't won because it feels in, like, in the bubble. But there's so many people you talk to who are in software engineering, but, like, aren't paying attention. They're not even in use. But here's the thing. They're still on top complete. Like, a lot of, like, people are still, like, oh, Cursor. I have a buddy. But the you know you know, I just wanna finish by saying, like, Cursor in a very meaningful way has won. 100 tell me. Okay. Okay. Yeah. Yeah. Yeah. Yeah. Don't get it. Sure. Sure. Sure. Sure. I say hasn't won, I'm talking about number one. Number one. I'm talking about number one. 10 I take $10,000,000,000 No. No. At the end of the day, every like, at the end of the day, number one is gonna be the best coding model. Yeah. I think that's kind of the thing. And that that's the thing. Like, unfortunately, you can't be number one unless
40:59Okay. Okay. And I wanna talk about someone who's been hiding for the it feels like they've been hiding. Shout out Logan Kilpatrick out there. Where's Google right now? Like, doesn't it feel like they're they're, like So I ramping up to something maybe? Or I have a couple of homies who work at Google. And the one thing I think Google has Google has infinite money. Right? They have GDP level money. They have GDP level power, and they have they have, like, uh, like, great talent. The problem with Google from, again, insider sources for me is the teams are not really allowed to fully collaborate together. Like, information is very sparse. Right? Like so there will be people in Google who will find out the new model dropped the same time you and me find out. Sure. Right? And and it makes sense. Right? That organization at that level that deals with that much data. Right? So, like,
41:45it would be nice if the Gemini team was allowed to operate in their own box silo where they're a startup, and they can mess up and they can move. Because, like, Gemini, in terms of knowledge, is the smartest model. The tool calling is not that good. Sure. Like, I tried it for computer use. It couldn't even click right. Like, it is just it it's it's embarrassing to say it's a Google model. But
42:08if you remember how bad Bard was Sure. And then how good Gemini three or 2.5, I forgot what class model was, That leap was huge. Sure. So I have point one is their latest. Yeah. I have no doubt they'll be able to leap again. I genuinely think and, like, again, this is from people who I know work in Google or close and part of these teams. It's like they have a huge organizational problem. Sure. Like and they're not allowed to really collaborate. Information is sparse.
42:32Like So so, no, think about this. Like, for if you if you have, like, a chat or a question that involves, like, a Google search, you're gonna go to Gemini. If you if you wanna study for class, you're gonna use Notebook LM. If you want to make a quick vibe coding app, you're gonna use Google AI Studio. If you're if you want to do serious coding work, you're gonna use anti gravity with which is like the new version of Windsurf. Windsurf.
42:51And then and then if you wanna do a design, you're gonna use The Stitch. Stitch. Stitch. Stitch. Yes. Yeah. So, like, they have all these product surfaces. It's like, hey. What if you did what OpenAI did? One super app. You axed all of these teams. Yeah. What if you just you you have the Elon Musk level, like, come in, you interview everyone. What are you doing today? Which he is great at. I'm not gonna lie. He is great at getting stuff together. Imagine if if someone but the problem is is, like, that person's going to be you have to be the leader of the company to come in and just do it. You're not gonna get some consultant to come in and do it. No. You know, like, but, like, honestly, if I was Google right now, I would take the best people from all these teams, put them into one, and either make Google AI Studio or anti gravity the super. Like, like, literally, like, Gemini should be its own startup. Like, it should, like like, they should have, like, maybe their own leadership. Like, obviously, I'm not saying they don't report to the core leadership, but, like, they should be given freedom to, like, move fast, break things, and do all this stuff because Google doesn't have the problem the labs have, which I don't think the labs have a problem, but, like, Google doesn't have to worry about money. They make money. A lot of it too. They have great data source. Own 12% of Anthropic or something. Crazy. Yeah. Google's not losing financially. Yeah. Sure. Sure. It's just a matter of, like, getting it together, and I find that it's a lot harder or it's a lot bigger ass than we think. Because to me, you, it's like, okay. Let's just get everybody together, let's hunker down. I think at an organization, that level, it it takes a little bit of shaking. Sure. Sure. Sure. Sure. So so okay. So, like, I guess I guess the,
44:17you know, the way that I think about it is they're one model train away from, like, being in them in the running. Yeah. You know, you know, they're gonna release a model that's good at tool calling. Yeah. I mean, you would have think Like, it's in like, there's no way they can't do that. Right? It's, to me, it's just internal stuff. Right? Because, again, Bard was terrible. I don't I don't think people remember how bad Bard was. It was extremely bad. What's a bad name? A terrible name as well. Sure. Kinda like what was it? Multibot?
44:42Yeah. I don't know. Dude, we don't need to get into all the names of Openclaw. They did settle on the best name, though. Yeah. Open Claw is nice. Open worked out. Like, them being bought by Open AI worked out. But Gemini like, the Gemini team, I just think, like, needs to be allowed to operate as a startup. They are one model that way of catching up to everybody. I don't think they're, like, years behind.
45:03And, like, all the techniques that are being used, it's out in the open. It's not like Anthropic or Codex is doing anything novel. I think open like, jump Google should win, and I want them to win because I own Google stock. So please I love it. I love it. Okay. Okay. I wanna turn this to to kind of get into some more tactical things and, like, strategic things. Right? Yeah. So if you think about all of this, right, you said it yourself. You said coding models
45:27the best coding model is gonna be the best at everything. Right? So given the fact that everyone's gonna be interfacing with coding models soon, alright, that's gonna be part of most people's work, what are the things that are kind of universal across any of these coding tools? Like, I'm I heard you mention on Greg's podcast, I I think you said something along the lines of,
45:45you said, like, what your your inputs really matter a lot. Like, it's really important, uh, like, how you prompt. And so I'm curious, like yeah. Talk about prompting and, like, what do you think are the most important things to do right now that kind of stretch across all use cases? Yeah. Like, uh, the one main thing, like, I I want a lot of people to understand is, like, even though, like, we label these models, it's like, well, this model thinks or this model does.
46:09Underlying, the LLM is just a predictor of text. Right? Somehow, mathematically, when you map, uh, words
46:17to a vector database, you map them on, like, uh, a vector graph. The like, words that correlate together are close by. I I don't even know how the math works, but if you ask a model what is the capital of France, on that graph, Paris is close by. So it's gonna pick Paris. So what the models are great at is they're a great predictor of the next token. It doesn't actually think. It just predicts what it thinks is the not even thinks, what it knows to be the next answer. Sure. Right? So the reason why I say that is the English that you use, the words that you use are way more powerful than you think. It's not a like, you and me can understand like, for example, I I could say, oh, we watched a basketball game, and those guys killed them. Right? Sure. Sure. Based on language and history and stuff, you know I don't literally mean that they picked up an axe and killed all of them. You know that I mean, like, oh, they just beat them in the sub. The models do not have the ability to understand the way you and I do. It will it has the ability to predict the next token. That being said, when you understand this, then you understand how powerful, how important your prompts are. This is why I think something like WhisperFlow,
47:23um, is great because I think more people can speak better than they can type. Right? Typing just feels a little too long or you wanna shorten things. Right? But being able to speak and to articulate exactly what you want the model to do is one of the biggest plus EV things you can do. Because as the model gets better at predicting the next text, it's going to give you exactly what you ask for. On top of this, it's like, it it's really hard to articulate what it is that you want if you don't have any industry knowledge. 100%. Which is why I think we're seeing the marketing types are getting obsessed with vibe coding tools or easier than the other people, and then, like, we're kind of starting to see these vibe there's this tool called Polja
47:59that came out. Uh, it was an AI marketing agent or whatever. Okay. And it just spat out slop. And so, like so this is this will make sense in a second, but, like, my point is is, like, you need to have some industry knowledge in order to articulate it. And so, like, right now, we're seeing a lot of gimmicky tools get launched where, like, it just spits out slop, but, like, nonindustry experts, a, don't know how to articulate what it is that they want, and b, they don't really know how to verify whether it's So it's good or not. Yeah. In the same way for me when I, you know, vibe code it at the beginning, like, you know, if for me, if it ran, I was like, good. And then people would be like, oh, but the code's,
48:34like, not good. And it's just like, no. I mean, look at it. It works. It's working. Yeah. It's working. And like, know, and that that's kind of the difference. And so what I've noticed is people who have like a lot of range, you know, if they're good at marketing, they're good at coding, or they understand the basic concepts are able to just do so much Yeah. Because they they know a good enough amount. You don't need to go as deep anymore. No. You can be a great six, seven out of 10. You can go really, really far. You just have to be a six or a seven out of 10. Yeah. Yeah. So generalists are just killing it right now. I I think across the board, especially especially if you have, like, high agency and, like, you don't really believe in the walls between stuff, like like, if you if you just have some good taste in design, like, you're a designer now because you can just you know how to use the tool for knowledge work. You can you just have to ask your team, be like, hey, how do I create a branch on this repo? Like, let me let me hop in. I'll edit it. And so, like, now our whole team, when we wanna make design changes, can like, even even the nontechnical people can just hop in and make the changes directly to the code base, and they can push it to their own branch, and then the team can review it, make sure the code's good. But, like, it's sped up our velocity so much is, yeah, sorry. This is a No. No. You're right. Like, seven out of 10, like, generalists can go far because, like, if you have the ability to articulate what you want exactly how you want, we're getting to the point where the models are getting good at giving you that. Now it's not perfect. There's still room for improvement, but it's much better than it was last year. Another thing that I've noticed that a lot of people do is a lot of people are giving it all this unnecessary context that it doesn't need, especially in the coding arena. Like, one thing people will do, for example, I've seen agent MD files where they'll be like, this is a React code base. You don't need to tell it that. Like, it when it reads the file, it's gonna realize, oh, this is a React code base. Right? Sure. Like like,
50:08skills, agent MD files, all these things need to be perfectly articulated, and they need to be necessary. For example, with skills, in my opinion, the best skills are, to your point, domain expertise and workflows
50:23that are specific to you and your company. Right? For example, um, and this was, an early, uh, test I did of OpenClub. I connected OpenClub to my Notion to dub.sh, which is my analytics platform. I connected it to my YouTube analytics and to Stripe. And I told it I told the agent, generate me a report of how my channel did and how my business did this month.
50:46It generated a report. It was complete garbage. Complete garbage. It took these numbers, that numbers. It didn't even do the math right, and that's because I didn't explain to it what a report looks like. So what I did was I created a skill as to, like, okay. For CTR, this is how you calculate. For the sponsor videos, this how you calculate. This is I explained it cleanly,
51:06thoroughly in a skill, and in one shot, it got it the next time. Right? So keeping your agent stack minimal, keeping the skills minimal, right, being articulate with what you wanted to do. Like, I've almost seen a night and day. And, like, there are times I'll experiment where I'll talk to it like a friend, like, vaguely slang and all this stuff, and I'll talk to it, like, articulately.
51:27The results are night and day. Sure. Right? So if you give it garbage, it'll spit out garbage. If you give it good, maybe might give you garbage, but the likelihood of it being good is a lot higher. 100%. One thing I found for, like, knowledge work type stuff, I actually don't even think about prompting as much as
51:45getting to one good output Yeah. And then reverse engineering the skill out of And then tell yeah. Yeah. Yeah. Yeah. Percent. So, like, what I'll do is, like, I I used to be, like, you know, and I have a very structured way I make my videos. Like, you've seen my my videos, I have a there's a lot of This is crazy, by the way. Anyone watching, it's like boom boom boom, like, it's a very Dude, we got all we got all we can change the can't change the camera angle. Yeah. This is a this is There you go. There's a lot of effort that goes into, like, preparing for a video. And so my my number one thing for YouTube
52:13is my YouTube skill. So, like, if I were to go to Codex, right, and I type in slash YouTube, I have this YouTube researcher skill which uses what's called the, uh, it uses the SERP API for Searching. Right? For, like, it could search Ross Mike on Utah. On YouTube, and it will list all your videos. And then it uses another API called SupaData,
52:33which is, like, a one second grab full transcript. And so, like, oftentimes, I will when I want a good when I'm I'm, like, I need this hook to be absolutely amazing. You know, I'll spend three hours on a on just the hook if it's, like, a big long video. Like, if I'm gonna spend all this time on the video, like, we have to get people to the good part of the video, and so you gotta spend a lot of time on the hook. And what I'll do is I'll just write down kind of what I wanna include in the hook, and I'll send my agent off to analyze all the educational YouTubers,
52:59find a similar video that performed really, like, outside of their range of views. You know, like, if they if they're averaging 30,000 views, it'll look for videos over 200,000. Yeah. Find the best hooks that they did, and then apply the same style to my video. And that's kinda how I'll do it. And so, like, that's one thing. It's like you don't have to rely on typing. That's why I like the term context management. Yeah. It's like you can send them to the right API,
53:23or you can give them a bunch of really good examples. And, like, there's a there's a bunch of different ways to, like, engineer these workflows beyond just typing or speaking. It almost feels like employee onboarding. Like, uh, one example I gave Greg, and, like, it's exactly what you said. I will, like, go back and forth with the agent, show it exactly how it gets done. And once we've had a successful run, we're gonna be like, okay. Look at everything that you did, how it went right. Let's turn that to a skill. Sure. And I've even, like, done this recursively where, like, I'll keep telling you to do it again and again, and there'll be times, like, you'll have a hip hiccup. And because, like, the agents can, like you could tell it, oh, you had a hiccup. Fix this. And they'll fix it. Then you'd be like, okay. Now make sure the skill gets updated so this doesn't happen again. Right? When, um, like, I like to think I coined this term, but, like, you wanna scale for productivity with your agents. Right? Like, this exact workflow might not work for somebody else because this works exactly for you. Sure. Sure. And but a lot of people are just installing
54:1850 different, 100 different skills from different people. Like, they might install yours, mine, and they're like, oh, why doesn't it work? Because the agents are supposed to be personal. Yes. Right? Like, it's personal computer made sense back then. I think personal agent makes a lot more sense now. I mean, it's like hiring an employee. Like, when you hire an employee, there's a process by which you hire. Yeah. And you either get good or not get good at that process,
54:41you know, and and that's why I think a lot of people are falling prey on Instagram and other social platforms where they are getting sold skills. Yeah. But, like, what you don't understand is I've seen a skill course. Like Like So many. So much money in it, by the way. Like Like, I was selling you for a $150 and, like, the video had a 150 k likes it. I was just doing simple math. I'm like, man. Dude, hey. You'll
55:02yes. All of that stuff is happening, and I think, yeah, people are not understanding the point. Think Yeah. Especially as we move more into general agent use cases, like, you just need to get good at agents. Like, you can get good at it, and it just takes reps. And so, like, learning how to create one skill is way more valuable than someone giving you an Excel
55:21file of a 100 skills, you know, that that you know, where you can download them. And that's why, like, the Open Claw Skill Hub Claw Hub. Yeah. Claw Hub was such a nightmare. Like, every skill I downloaded, it was just doing random things and, like And some of were viruses too. Some of them were malware. Yeah. And so I think I think yeah. Learning that as a skill is really important. And as you were saying, the agent doesn't fully know everything that you're doing, but the model companies,
55:45they want to solve for that. Yeah. You heard of Chronicle? No. Chronicle's the new feature that on on Codex. So if you go to personalization
55:56and we go to this Chronicle research preview, it screenshots my screen, like, every few seconds, and it stores it. And so it's starting to, like, pick up on additional context. So if you're watching a YouTube video, it'll know you're watching a YouTube video when you type. And and now I'm seeing more and more that once you have this and by the way, I'm not. There's obviously a lot of, like Security and, like, privacy stuff that you're I'm not endorsing this as a feature. I'm just letting people know, like, this is the direction it's going. It's gonna have more and more context. And so I listened to Aravind,
56:29the CEO of Perplexity. Yeah. He he gave this speech once about how something flipped in his brain when he he used to get mad at users or blame the users for typing in really bad queries on Perplexity. And then one day, some guy told them, he's like, no. You need to have a 100% like, everything that goes wrong is on you. You need to make it as easy as possible for people. And, like, I feel like that's the direction it's going. It's like, they would rather this just work. And 100%. And so there's a lot of ways that they're trying to make it work for people, which is, like, recording your screen, eventually, maybe audio, and, like and you see all of these connections. Right? If you connect it to linear, right, maybe, like, the best prompt you could do is implement this feature at linear, and then you'd, like, type the name of the ticket. Yeah. And then that's the best way to do it. And so, like, it's fun to see how all these companies are doing integrations. Yeah. There's a lot of, like, new cool stuff. Like, this is interesting. Another thing that I I built, and I I didn't know if, like, it'd go far, but, like, now with all the stuff that's happening is interesting. What I thought of is, like, let's say you had, like, a web app, like like, a simple product that people love, and everyone wants to customize it to their own degree. And, like, you make it, like, module, you add drag and drop, and all this stuff, but you can only take it so far. But then when I thought of it's like, imagine you had, like, a chat bubble over here, and people can prompt
57:43the features they want on your app. And, like, this, like, spins up a sandbox, agent builds it, pushes it to their instance, and, like, at first, like Are these technical people or completely untenable? My thought like, I thought of this, like, four months ago. My goal was, like, for untechnical people nontechnical people to do this. But, like, for simple web apps and at the time, the models like, I at least the models didn't allow me to think this visionary, like, early December. I didn't have the Opus 4.5
58:09moment. But now it feels like with all the context these models will start to have of people, it might get to the point where, like, you might not be able to communicate it exactly what you want, but based on your behavior, it knows what you want. Right? Like and we already add this with, like, you know, ad platforms. Right? Like, you look at one thing and they based on whatever algorithm they have, know exactly
58:30what the next thing it is you want. Right? So I think personalization is, like, a big thing that everyone's going for. There's a couple different angles, but at the end of the day, the best coding model is the model that can do all of it. Yeah. I know. Honestly, like, that's actually what it just comes down to. It sounds like what you're describing is kind of like a GitHub for vibe coding. But, basically, like, imagine, like, there's this like like, let's say I'm on Twitch,
58:52and, like, I, like, want this specific like, I want the chat feature here, and I want this here. And, like, if I could, like, say, like, oh, can you drag this here, drag that there, drag this here? Like, the models can do it. Sure. And it's at a point where, like, you know, back then, like, one shotting was a miracle. Right? Like, remember, like, when that was, like, the benchmark? Like, why one shot this and it worked.
59:13Right? Now, like, once like, you expect the first shot to work. You expect the app to run. Right? So I think, like, stuff like that, like, personal customization, having the app specifically tailored to how you want, API as a service, like, these things are starting to become possible because, again, the best coding model is Yeah. Yeah. Yeah. And it it just connects to things, and it whatever you would want an app for, it just does the thing. Mhmm. It's it's it's a really, really weird world out there. I do wanna hop to
59:41you've been talking about c mux, t mux, v mux, g mux, all the muxes. Yeah. Yeah. Yeah. Let's let's dive into some tools that people aren't talking about as much right now, like non super apps. So you, my cofounder, and a couple other people on our team are using c mux, I believe, which from my perspective, it just looks like a chaotic terminal mess. Yeah. What what is the so, basically, c mux is built on top of wait. Is it c mux or t mux? I might be it's one of the muxes. One of the muxes. I I have all three. Like, I have and there's a demux as well. Like, if Elon you have mux. Yeah. There's all these muxes. But, basically, it's,
60:16like, a giant terminal. It uses libghosty, the same library that powers ghosty terminal. But then I have a tab, and I can have multiple terminals in one app. And I have these hot keys, and I can switch between different tabs. And I can have multiple terminals
60:32Okay. On the same screen, but I could also have a browser. So you have the you have the left sidebar like like this. Right? Yes. And that's, like, all the tabs for the different terminals, for the different, like, projects. So it's like it's like this? Basically, yes. Yeah. Yeah. Yeah. But, like, all these can be terminals. So it's just terminal, terminal, terminal, browser.
60:51It also like, CMUX has a browser, like, like, browser. Like, the hotkeys are great. And so this is the big thing. There are a lot of people and we'd see it. The companies are pushing for GUIs. Right? I don't know if the codec CLI
61:06has gotten an update recently, but it feels like they're pushing for the GUIs. There's still a part of me, especially for focused work, that likes working on a terminal. Right? Like, if I'm working on one project and one project alone, I'm solely focused on this, I don't know why. I feel like this incentivized me to work on multiple things at a time. And that's not bad. Like, there are times where I wanna work on multiple different things at a time.
61:28But it feels like for the terminal, I'm sure, like, the people on the Open it in a mini window. See, the super app could do everything. You just you just full screen it. It's easy as that. Super app. But, like and I think your team would agree with me. There there's just something about, like, this, like, focus, like, although, yes, I have multiple terminals of multiple threading, there is this, like, uh, thing about working at terminal, especially as a developer. I don't know if it's, like, you know, we've used it for so long and we can't let go or
61:55like, I can't even tell you there's productivity gains, but, like, I I love c mux Yeah. Yeah. Yeah. Or d mux. Cool. Yeah. D mux, whatever the mux is. Yeah. Have any bold predictions for this year? Let's let's think about this for a second together. Like, where are we going? I'd like this this is something that I like to talk about with the team. It's like it's like, let's think about this last year for a second. Right? Or or, you know, the last year, you know, we had really smart models. We had the the terminal interface came out. People started coding.
62:22AI agents are starting to think for longer and longer and longer. We're to the point where, like, I have done an hour and a half run on on I told you, I cloned Replit. Yeah. Or or level just released. It had sandboxes. I used Daytona. Yeah. And it just worked. I mean, hour and a half. Yeah. But still. But it was it was good. Yeah. And I could make edits to the app. It's actually insane. I was actually, like, blown away. But, like, yeah, it's like, okay. So these agents are thinking for longer. They're getting better and better at tool calls. There's more and more integration. We're giving them computers. We're giving them computers. They're gonna be able to spin up these computers, do things, and then yeah. And that begs the question, though, like, one thing about the negative downside of the sandbox are that they all get blocked by servers. Like, they see that they're in some AWS warehouse. They get blocked. Yeah. And so that's why I think Perplexity
63:12is doing their personal computers because it gives them an IP address. On their computer. Yeah. Yeah. Yeah. I I think a lot is I I think I a lot of companies are gonna realize, okay. Agents are gonna be customers. Like, for example, Stripe had Stripe sessions today. A good chunk of their announcements is like, oh, you can use your agents on Stripe. You can use your agents to issue cards, to do this, to do that, to do that. So I think from a business perspective,
63:34like, you used to think about developer experience, but now people are thinking about agent experience. Uh, like, how how can you have agents, like, um, integrate to where like, I work at Convex. Convex is a back end as a service. The one thing now we think about is how do we make the model so good at Convex? Like, what how can we change our APIs? What documentation can we give so that it's just so simple? You guys plugged in yet? You guys got a plug in? It's coming, I think. It's coming. I think. It's coming. Yes. Um, we have one on cursor. So, like, when you think about all these things, I I see a lot of company think about these things. But in terms of predictions, first and foremost, I've come to realize it's almost so difficult to be on the spot because one thing I don't think a lot of people saw because we were Dario and all these people were betting on coding being the one that it was going to absolutely kill at. And although it's got better, we still and I'm sure you've realized, uh, you know, with the company you're running, good engineers are still necessary. Like, the gap between someone who doesn't know and a good engineer both using the tools is very, very far. But
64:32knowledge work, the improvements in knowledge work is astronomical. I had a 27 page contract sent to me the other day, and I asked a lawyer for a quote, and the price had a comment. I'm like, there's no way. Right? There's no way I'm paying this. You know if you you paid him and you send it to him, he's running it through Harvey or whatever. Yeah. I ran it through Claude,
64:51and it broke every single page down. It gave me the gotchas. I literally asked it, where can they screw me? And it broke it down. When you take this clause and that clause, they can basically do whatever they want. That was something that I was going to pay a thousand dollars for. When it comes to knowledge work Which is a little amount for, like like, a thousand dollar an hour or some of these lawyers. Yeah. Like and and listen. Yeah. Like, I definitely got a good quote, but, like, some like, yeah. Some of these knowledge work tasks accounting, like, my my friend Muso, like, he did our company taxes,
65:20like, the bookkeeping, like, and gave it to an accountant. The accountant was like, oh, do you have CPA experience? No. He used Clark. He used Clark Max. Right? So there's like, we've made such an improvement with knowledge work, and I think, especially with, like, these computer use things and all that stuff, like, knowledge work is just gonna be rocked. Sure. And, you know, is that job loss? Is that people just using the tools more? I don't know that part because, you know, you're talking about people and, like, layoffs and all these things. You know, it's sad when you hear these things, but I feel like knowledge work is just going to exponentially
65:52take off like nothing before. Like, we're just seeing the beginning Yeah. Of, like, the computer taking the agent taking a screenshot, feeding it to its model, and then getting word to click. So so okay. So if we think about this in layers, right, you have this, like, super smart like, LLMs are getting really good, and then you have this agent layer, which allows it to use tools, and then it one of the tools that it can or can basically use,
66:13I guess, what makes an agent really useful is it can use all the tools Mhmm. On a computer. Mhmm. Right? And then one layer on that is it can actually use the interface of a computer, which is like mouse, keyboard, etcetera. And then on top of this, like, layer on top would probably be, like, to create, like, really good quality documents. Mhmm. Like, maybe it can start to learn how to use Canva, that type of thing, etcetera, etcetera. And then probably
66:35the next big unlock, which I'm actually unaware of, is payments. Right? Yeah. There's a lot Stripe announced a lot of that today. Right. Like, a really useful employee has the agency to go in and make payments, and and yeah. I guess do you know anything about that? I I actually, like, built a mini version of this for myself. I wanted to see if I could handle if my agent can handle Uber Eats. Right? So there are many, like, virtual card providers. They have APIs. You could do that programmatically. The issue with virtual cards is some, like, sites don't accept virtual cards. Like, your thing needs to have an address.
67:04Stripe today launched the ability for you to issue cards that are for agents. And the way this works is it's your Stripe credit card or debit card or whatever they call it. So it's a legitimate card tied to your business that has an address. The agent I don't know how the API works, but, like, no sensitive information is fed to the agent. Sure. It can make the payment. It can get to the stage where it's about to make the payment, and then it messages you. It says, should I make this payment? It doesn't It it so it can't so all way. If you can turn it off. I I'm pretty sure you should be allowed to do that. Because if you can preload the account like, I picture startups with millions of dollars. Like, if you give it $50 and it goes rogue, like, if you raise $10,000,000 That's what I'm saying. Like like You know? Like, obviously, don't give us access to, like, the Amex that has 100 k limit. Sure. Sure. Sure. But, I'm thinking, like, you know, I wanna be able to give it $2,000
67:53weekly limit. Yeah. And it can order rogue. You know, it's a expensive l, but, like, it's a good story. A 100%. I mean, we'd make it money back in the in just making content. But you know what? Industry is starting to change and on all this? It feels like all the crypto bros who are legitimately building, like, legit things that didn't get caught up with the pump and dump stuff. Now
68:13everyone's focused on agentic commerce because, like, now, like, you can issue, like, tokens and all this stuff, you don't need a bank or a subsidiary. So, uh, have you are you familiar with x four zero two? So it's a protocol Sure. That Coinbase is backing that basically it's like an HTTP request that an agent can fire to send payments using crypto. Right? So there's a lot of these like a buyer and a seller. You there's a way to connect. Yeah. And you can exchange funds. The only thing reversible? Like, is it backed by it? Like So Coinbase is backing you, which is, like, again, it's a pretty big company. And I know I think, uh, Stripe has their own implementation called machine,
68:50uh, something protocol. The only issue currently at this state is you have these fragmented protocols. There isn't one united thing. And I think, like, that's a big thing in the next year is, like, having these agents make payments. Right? Like, for example, if, like, know, at your guys' office, if there's, like, an Uber Eats stipend where, like, people can order lunch, heck, like, why doesn't the agent handle that? Well, I mean, we actually already Anj, my cofounder, he he has a grocery bot that he uses. He orders we like, we get blueberries and yogurt and stuff. He show up at the office unannounced. It's hilarious. Like or he'll message it on text and be like, hey, we were running out. But I think he's using the
69:27he's using, like, some Amazon API. Yeah. Yeah. It's not like he gave it a credit card. Give it a card. Like yeah. And I think Stripe is probably the one that's going to make it happen. Yeah. Again, they literally launched the though that are only doing agentic payments. Yeah. There's Cross Mint. There's Natural Pay. Agent yeah. Natural Pay. There's agent card dot s h.
69:46There's, like, a lot. Like, I've I've looked at It's gonna be I mean, think about how it's a big prize. 100%. So yeah. It's a lot of Get a lot money. Get a lot of players. Yes. Yes. Yeah. Like, pre revenue raising Yeah. Yeah. Yeah. Yeah. Just off the idea. It's just it's just easy. Yeah. Okay. So so that's a prediction is AI agents get Commerce. Yeah. Commerce is gonna be big. Commerce is gonna be big. Knowledge work documents will get 10 times better, I think. Like like, just the People are gonna get scared at how good they're gonna get at creating, like, presentations and depth. No. It's gonna be really good. Yeah. The next scary thing though is and I'm low key not a fan of, but I am at the same time, is how good image models and video models are getting. Yeah. Like, chat like, GBT image two
70:26is really good. Like It's really good. Again, in our bubble, you can be like, oh, I can pick d fix. Look, I can fool my mother. Like, I could take I could get GBT Image two, get a picture of me and LeBron James, and I know for a fact there are people in my family who believe it. I have I don't know if you do. I have a paraphrase now, um, because my ex account got hacked not too long ago. So now I have a paraphrase. If anyone sounds like me and is asking you for money, ask them for the paraphrase. A secret code. Yeah. Yeah. Yeah. Yeah. And that's because, like, again, the
70:55image gen models are getting good. The video models exceed dance is getting good. Like, again, there's quirks, but, like, don't look at it where it's at now. Where is it gonna be in six months? And then, like, voice cloning has been a thing. So there's, like, that part of models that scares me, and I feel like there's gonna be bigger scams, bigger hacks,
71:13um, especially on Facebook with older people. Yeah. That that that's going to be big and that's scary. It is scary for yeah. For whatever reason for me, you know, I I mid journey in mid twenty twenty two was my portal into AI. Yeah. That's when I got I dropped everything to to dive into AI when I was using Dolly two and Yeah. Mid journey in the early days. And I was obsessed with image generation. It's like, if you plotted my interest in image generation with how good they got Yeah. It they're inversely correlated. Yeah. Like, for whatever reason, now that they're so good, I don't even really think about them much.
71:46For whatever, there's something, like, magical in the early days about, like, the imperfections. Possible. Like, it's just, like It's gotten so good. It's, like, I'm numb to it. Like, I use it, and it's it's one of those things, like, Photoshop. For years, they they invented all of these different types of tools and toggles that you could do to achieve these effects. By the way, engineering wise, some of the most immaculate software because, like, anything to do with image, audio, like, being able to, like, change those date like, that data, it's insane. I'll never forget it. Like, one time I had an internship, and I talked to this guy, and he would spend, like, two days on, like, getting the outline for one of his images. And, like Yeah. His weekly deliverable was the cover an image on page six of this thing. Yeah. You know? And and now a model replaced Photoshop.
72:32Yeah. Like, that's insane. If you can articulate the change you want, right, which gets back to, like Yeah. Having industry knowledge and have being able to articulate what you want, you can get that effect. Can you go on Twitter real quick? I just wanna show you how I one shotted this one image, and that's when I was like, it's it's basically over. Here. So if you click on media
72:51Oh. Click on media, and then there should be scroll like, that image right there. That was one prompt. Yeah. What was the like, was just like I I said, like, make it look like I signed in the NBA for the Raptors, and I put my image in. And look at that. Look at that level of detail. Yeah. Yeah. Yeah. Like, this is insane. Yeah. And this was someone's job. So, yeah, I think this is a good theme to think about. Like, this is coming for all knowledge. Like, whatever you you're the feelings that you have right now that this is insane, like, it's coming to presentations.
73:23It's coming. And this is not, you know, and I and, know, which makes you think about tools like Gamma. You know, Gamma The presentation. Deck maker. It's like it's like a interface didn't do this. The model did it. Right? And that's where it's going for, you know, knowledge work. It's 100%. And, you know, like, that's the scary thing for a lot of rappers. It's like, there are no you know, it's it's not the it's not the risk of OpenAI releasing an interface for design. The model itself being It's just yeah. Being able to do it and then make all the changes, and it won't even need an interface, which is the crazy thing. And, like, the fact that this was one prompt and,
73:59like, just, like, the sweat, the glare, like, this is, like Yeah. That's insane. Like, it's just it's incredible. And, like, now imagine product images and all these stuff, like like, all these companies and services that existed because, like, this skill like, this like, someone got paid a lot of money ten years ago to create something like this. It was just incredible. Right? Like, now the model's been able to switch. So, like, anything in the knowledge work, creative work arena,
74:24like, the unless you yourself are the the differentiator, if you've just been basic the whole time, average skills the whole time Yeah. You know what? And and and I think one theme for me about content creation, and this applies to you, I think, when I think about content creation, is like all of the repetitive
74:44process based things in video too, because this is all coming for video too. Like, any like cause videos can be informational, and, like, AI is gonna be able to create demo videos. Yeah. The more AI gets better, the more I'm just tempted to literally stream for three hours a day. Do this three hours a day. Yeah. And clip it, make it human, wear funny, so, like, literally just be weird, like Yeah. Like Something that agents can't do. Like, for example, in the coding space, I'm sure you're aware, tutorials are dead. Yeah. Nobody watches coding tutorials anymore. Yeah. Oh, yeah. Coding tutorials. Yeah. Not like tool tutorials. Like Yeah. Sitting down, like, okay, guys, you're gonna learn Python today. So import this, this, this, like, line by line typing. Like, all those channels except the one guy, KodoThantonien. That's because he's a g. All those channels have went down. Well, well, the people who the the the tutorials that win are the people with great personalities. You look at, like, Theo, Primogen,
75:32those guys, they're fun to be around. You're hanging out with them, but they're also doing the thing that you like to do. Yeah. They're talking smack. They're, like, you know, having these type of discussions. Oh, Anthropic sucks. Open eyes this. Like like like, that's what people, like, the relationship like. So that, thank god, won't be taken from us. Yeah. At least I hope. No. No. No. They dude, I'm telling you right now, the biggest thing, like, right now, I tell people who are developers who are getting afraid and they're, like, trying to find their path out of Yeah. Doing software engineering. They're like, dude, I love them, man. Like, you make content. Like, how do I get into it? I'm like, dude, just turn on your computer, figure out the basics of a of a of a of a camera. You don't need to buy a $4,000 camera that I have after like, I had a crappiest
76:09camera for a year and a half. Yeah. Yeah. Yeah. It didn't matter. I had a mic. This is, like, road mic that was partly broken. I got to, like, 500 k follow. You don't need all the nice equipment, but, like, just learn the basics, and then just, like, turn on the computer and just have fun Yeah. And stream yourself, film it, you know, and you'll figure out, like, the intros if you're posting long form videos, but I think there's so much room in the market right now Yeah. For people to just
76:32because people aren't looking for tutorials per se. They're looking They want they want people and they want to, like, understand what you can do. Like, just wanna turn their TV sit, get some inspiration, and then they wanna go mess around with it. Yeah. Because AI is it's not as rigid as coding. It's and it's made it easier now. Like, there are apps that I built that I use, like, a technology stack that I would have never picked up. I would have never used because, like, I just didn't have the domain expertise nor did I care enough to know. Sure. Like, I built a mobile app the other day using Swift. Never written a line of Swift in my life. I and I don't care to learn to. Right? So it's to your point, like, it's made it AI has made it easier to start things, but, like, the there's still a need for community and and and and chatting and information, and there's also a lot of grift online. So, like, if you're, like, just honestly, like, an honest person and you just share space It's a lot harder to grift when you're in a streaming because, like, yeah, I mean, the long right there. The longer you're on camera, you know, most grifters there are on Twitter or they're on The Twitter grift is insane, though. I don't know if you noticed. Like Oh, I do. My Product product launch,
77:33and then all of a sudden, the same accounts are running Well, I mean I mean, you you would be surprised at how high that goes up, like, in terms of what companies are doing that. So, like, everyone is doing that right now. Yeah. For some of these launch videos, it should be like the Dude, I I don't I don't wanna call them out. I but there's a lot it's standard practice to do the botting, but that's not even the grilgriff that I'm
77:56what what I'm saying is, like, if you want to stand out, you want to build trust and build an audience, build a brand in this landscape, even if you're a company, like, go long, do stuff like this, or just, like, create, like, super high quality content. And, yeah, I think those people are gonna win, because AI is coming for your lunch. It's gonna be crazy.
78:16CTAMaybe an NBA player for a day. Wow. So, like, I guess, like, any final thoughts, anything that jogged your memory through this episode that you wanna get off your chest? Yeah. Like, I guess, like, like, couple things I'll say. Like like, as scary as some of these things is, it really is the best time to start. Like like, whatever it is. Right? Like, it's a service, it's a business, it's a hobby, it's a thing. Like, you wanna learn. I don't know. Whatever it is, like, it's the best time. Like, information is so cheap. Right? There are people who spend 10 years of their life to acquire information that you and I now have access to for $20 a month. Sure. And if you don't wanna pay, you get the chat GBT ads. Right? Like, information is so cheap. Like, really what matters now is and I hate the term because everyone uses it, but it's actual agency. Like like, you know, searching for the information, trying to do something. It doesn't work. Figuring it out. Like, keeping like, you know, you you keep on going. Like, it is such a great time to be alive where, like, you and me, like, met online a year ago. Now I'm here at your fucking studio. Companies hired from Twitter. Like Yeah. You know what I mean? They're, like, there's so much, like, power in, like, putting yourself out there. And, like, you don't have to be loud on YouTube. You don't have to like, you just be you. Yeah. Put yourself in some sort of meeting where other can see. Like, you know, learn, be honest, share. Like, you know, I think there's so much opportunity in that. And although there's so much gloom and doom with all these things, I generally think it is the best time to, like, build or do or It's the best time to build or do. I mean, if the world if the world is in the state that the doomers say that it's in, like, might as well build doing it. Yeah. Yeah.
79:45And, you know, whether or not it's the best time to get started, it's definitely better than, uh, one month from now or five months from now. A 100%. Yeah. I'll be honest. I would take this time now over any other time. Like, the, like, the amount of things that, like, you can learn so quickly, like, the amount of technologies you can learn, like, you can build apps. You don't even have to be a programmer. Sure. But, like, sure, it might not be the best the counterargument to that is, like, it's just, like, oh, but if anyone could build it, do they have value? But but but here's the thing. Most people won't. Right? Like, they and and and that's the one more people could have But it's agents won't. What if their agents just do? What if the a what if it's just some guy who spins up a 100 agents? Then you know what? Do nothing and live the life of your life. You know I'm saying? I get a devil's advocate here. Yeah. Yeah. Yeah. And I had someone say that to me, like, a a recent grad. He's like, oh, but, like, what if this happens? What if that? And I'm like, okay. Then be a bum. Like, do nothing. Like Do nothing. Yeah. Yeah.
80:31But you can. You can do stuff. It's really accessible. You can learn. You can make friends. You can partner up. There's so much, like, for example, get my friend there's a lot of things he does that I don't know how to do. Sure. I'm more of a technical, so, like, there's so much that can be done. It the the saddest thing that one person can do in this time is just be sad, doom, and gloom. Like, there there there's just so much you could do. Yeah. And also, like, a different framing of it that I've I've because I've talked to a lot of people that I used you know, went to college with, um, who'd, like, reach out to me because they see that I became a content creator, startup founder, etcetera.
81:04And, like, they just don't feel like they have permission. Like, they're waiting for someone's permission. And, like, the one thing that I just say is, like, you can do it. Like, whatever it is, the thing that you wanna do, like, if you think there's a 10 step plan Yeah. Just do just take the direct path. What is it the minimum viable thing that you could do to get you there and just you're allowed to do it and go do it? If you need to move cities, go move cities. If you need to
81:27CTAbuy a computer, buy a computer, figure it out. You have permission. You can do whatever you want. Like, it's just it's all right there. Like, the information's all right there. Like, you wanna learn how to do this. You wanna do like, I wanted to, like, learn networking. Right? I'm not really struggling networking. I wanna learn networking. You know, before you'd have to buy a book, I I literally went back and forth with Claude. And he was telling me, oh, you could do this. You could do that. I still don't know what networking is. So know what I mean? And what even is that? But but there's so you could do so much now. It's just a matter of, like, you know, saying, like, I'm gonna do it. I'm gonna try, like, being okay with, like, sucking at first because that's a lot of thing. People wanna be specialists on the first go. I wanna be Sure. Sure. A professor on the and that's just never gonna happen. Um, but,
82:07CTAyou know, if you have agency, if you have that dog in you, I don't think there's a better time than now. I'm I would I would take now versus any time in history. Yep. Well, in this episode, we talked about all of the tools. We talked about all the tools. We talked about all the advancements. We even laid out some predictions. AI is gonna get really good at everything you could possibly do on a computer. That's basically the theme. Mhmm. And we also gave you permission to go out and test anything you want. So that's pretty good. Please. You don't even need the $20 subscription. You can use the free one. Come on. Come on. Buy the subscription. But you should. Yeah. Because that's Don't watch ads. Yeah. Yeah. Don't use Yeah. Yeah. Yeah. Don't watch ads. That's not cool, man. But, anyway, thank you guys for watching. It's been a pleasure. Appreciate it. My brother. Handshake at the end. Alright. Peace, guys. Thanks. Subscribe. Like. Bye.
— full transcript
§ 05 · For Joe
Steal the 'two operators recap the quarter' format.
Riley + Ras Mic playbook
Two practitioners in one room beats one expert with a slide deck — and the whole thing is a hand-rolled show-bible written from their feeds.
- Pull a hard time-window — 'Q1 2026, the past four months' — and frame everything inside it. Time-fenced retrospectives travel.
- Use the YouTube description as a real show-bible: 17 timestamped chapters, 5 part-headers, a guest plug. The description IS the trailer.
- Open cold on the solo host with a giant on-screen number (2026), then cut to the two-shot. The cold-open-solo / sit-down-two-shot rhythm reads as 'recap show', not 'interview'.
- Frame the table-of-contents in your hook by name-dropping the entities you'll cover (Codex, Claude, Cursor, OpenClaw, SpaceX). Don't tease — list.
- Cut to B-roll screen-grabs of the actual tweets and product UIs you reference. Anchors abstract talk to real artifacts.
- Steal Ras Mic's 'agent as new hire' analogy verbatim — would-you-give-a-new-employee-40-jobs-on-day-one is the cleanest framing of narrow-agent design out there.
- End every long episode with a 'permission to build' close. Riley does this every time. Repeat watchers internalize the brand promise.
- When showing dollar figures (a $10B / $60B deal, a $1000 lawyer quote, a $5000 worth of compute), say the number out loud AND show it. Numbers travel.
§ 05 · For You
What this could mean for you.
If you're not a developer
The smartest people in AI are now spending most of their time on knowledge work, not code — and the tools to copy them are already cheap or free.
- Before paying a lawyer to read a contract, run it through Claude or ChatGPT and ask 'where can they screw me?' — you'll spot 80% of the gotchas before the meter starts.
- Set up a 'paraphrase' or family code-word right now. Voice cloning is good enough that someone can call your mom sounding exactly like you. The code-word stops it cold.
- If you keep notes, keep them in plain markdown files (e.g. in Obsidian or just a notes folder) — every AI assistant on Earth now reads markdown natively, so your notes become AI-readable without any extra work.
- Pick ONE narrow task in your week — sponsor emails, calendar sorting, expense receipts — and try to automate just that one. Don't try to automate 'your life' on day one. Narrow wins.
- Use voice-to-text (WhisperFlow, MacWhisper, even your phone's dictation) for everything you type into AI. You can articulate twice as much in half the time, and the model gives you back exactly what you described.
- If you've been waiting for permission to start something — a side project, a business, a skill — the cost of trying has never been lower. Twenty dollars a month, or zero with the free tiers, buys you everything the 2018 enterprise stack cost.
