AI Can Improve Itself Now. We're Sure That's Fine.

Summary

---

Podcast Summary: AI For Humans – "AI Can Improve Itself Now. We're Sure That's Fine."

Hosts: Kevin Pereira & Gavin Purcell
Date: March 13, 2026

---

Episode Overview

In this episode, Kevin and Gavin dive into the accelerating world of recursive self-learning AI, discuss the societal implications as AI systems begin to improve themselves autonomously, and review recent major news and launches from Anthropic, OpenAI, Meta, Cloudflare, and more. The conversation underlines not just the technical breakthroughs but the shifting dynamic as regular people, not just developers, start leveraging AI to build practical, personalized software. Expect their signature humor juxtaposed with insightful, sometimes skeptical, takes on the state of AI.

---

Major Discussion Points & Insights

1. Recursive Self-Learning AI Is Here

• Key Point: AI is not just a tool for humans—it is now improving itself, a reality in major AI labs and in the open-source community.
• Notable Example: Andrej Karpathy’s experiment: Used a GPT-2-level model to self-improve via recursive learning, achieving an 11% success rate over attempts—small, but significant and compounding.
• Industry Implication: Anthropic, OpenAI, Google, and smaller labs are incorporating self-improvement in their latest models.
• Quote/Insight:
  • “[Recursive self improvement] is really here… recursive self improvement in the broadest sense is not a future phenomenon, it is a present phenomenon.” – Evan Hubinger (Anthropic), cited by Gavin [09:25]

2. How Humans and Agents Collaborate (or Don’t)

• Key Point: While AI can autonomously execute complex tasks, the challenge for humans is articulating what they want and how to guide these agents efficiently—otherwise, outputs can miss the mark.
• Personal Anecdotes: Kevin describes using agentic tools to run a "paperclip company" and customize workflows on the fly, demonstrating near real-time product evolution.
• Practical Advice: Use built-in AI interview tools to fully specify what you want—“don’t just say ‘give me X’, ask for an in-depth interview about every aspect.”
• Quote:
  • “Recursive self learning doesn’t mean that much if you can’t, as the human, elucidate your own idea in a way that makes sense.” – Gavin [18:14]

3. Ordinary Users Becoming AI Software Builders

• Key Point: AI tools are reaching the point where non-technical users can write and deploy custom software, not just consume it.
• Examples:
  • Gavin’s brother-in-law built three business applications himself using OpenClaw.
  • Tools like Replit v4 aim to become the "Squarespace of software," letting anyone build not just websites, but the applications behind them effortlessly.

4. Agent Ecosystems and the Meta Acquisition

• Key News: Meta acquires Molt Book, a social network for AI agents.
• Perspective: Meta may be betting on agent-to-agent communication as the future of the web (vs. direct human interaction).
• Quote: “There’s still going to be a portion of the internet for humans… but a lot of it is going to be agent to agent communication. They don’t need half of the tools, half of the standards, even the graphics.” – Kevin [27:48]

5. A Wave of New AI Tools, Agents, and Features

• Anthropic: Launches rapid new features—Cloud now builds interactive chats and diagrams directly.
• Replit v4: Multiplayer, real-time collaborative software creation.
• Perplexity Computer: Another shot at personal cloud computing with agentic capabilities, but with mixed reviews.
• Runway's "Characters": Real-time, API-driven avatars/agents for human-like interactive experiences.

6. The Commoditization and Competition of AI Platforms

• Discussion: Whether second-layer companies (like Perplexity, Replit, Cursor) can survive as OpenAI, Anthropic, and Google expand their core platforms.
• Analogy: The proliferation of specialized hosters (e.g., Vercel) despite giants offering overlapping services.

7. Cloudflare’s Surprising Pivot

• Key News: Cloudflare, long known for protecting sites from bots, now offers a scraping API, flipping its prior stance.
• Commentary: This is likely to fuel controversy, as companies must choose between web safety and enabling the agentic web.

8. Robots: From Living Rooms to Horses

• Figure Robotics: Demoed a humanoid cleaning robot autonomously tidying up a living room. The pace is slow, but autonomy is the breakthrough.
• Deep Robotics: Reveals a robot “horse”—evoking both awe and skepticism about real-world utility.
• Humor: “They yeeted the pillow!” – Kevin [49:13]
• Societal Reflection: Are these robots harbingers of war, or just future urban nuisances (e.g., cyber-horse tourists in Santa Monica)?

9. Creative AI: Agents Making Art & Video

• Showcases:
  • Claude Code used to generate "YouTube Poop" glitch-art videos articulating LLM perspectives, creative humor, and personal bios for the hosts.
  • “This is one of those shocking, ‘agents have come a long way’ moments to me, because again, six months ago this would not have been possible.” – Gavin [36:53]

10. Other Notable News and Quirky Finds

• Google: Rolling out Gemini AI into Docs and Maps, plus a powerful new multimodal embeddings model.
• Nick Biese’s AR Demo: Live X-ray effect using ThreeJS—portending a future where kids code sophisticated visualizations for class.
• Green Code‘s Tennis Match Predictor: An LLM model correctly predicts 85% of matches from a historic player stats dataset, raising eyebrows for future betting and analytics.
• Halo on Mac: Community uses AI and technical mods to bring classic Halo games to Mac OS.

---

Notable Quotes & Memorable Moments

• On Trusting AI with Big Tasks:
  “If you lose a week or a month…if you can’t check in along the way and see that it’s going well or trust the agent…a lot of people are going to recoil.” – Kevin [04:51]

• On Self-Improvement Emergence:
  “Recursive self improvement in the broadest sense is not a future phenomenon, it is a present phenomenon.” – Evan Hubinger, via Gavin [09:25]

• On Consumerization:
  “What happens when the world at large—normal people—start writing their own software for things they want?” – Gavin [10:58]

• On Creative AIs:
  “It made jokes, Kevin. And it made jokes that made me laugh.” – Gavin [35:53]

---

Important Timestamps

• [00:09] – Karpathy’s recursive self-improving AI project
• [02:21] – Sam Altman quote on the rapid increase in AI task horizons
• [09:25] – Anthropic’s stance: recursive self-improvement "is a present phenomenon"
• [13:13] – Kevin’s “paperclip company” - orchestrating agent fleets
• [27:48] – Discussion of agent-to-agent web and Meta’s Molt Book
• [35:15] – Claude code creates a YouTube Poop video (agent art)
• [44:43] – Cloudflare launches scraping API
• [47:39] – Robots: Figure (living room cleaning bot) and Deep Robotics ("horse")
• [56:15] – AI predicts tennis match outcomes at 85% accuracy
• [57:32] – Frodo Baggins in "Pawn Stars" AI mashup

---

Final Thoughts

• The AI adoption curve is steepening—not just in labs, but for everyday users.
• Recursive self-improvement is already happening and will redefine AI capabilities rapidly.
• Agentic tools will push both experimentation and anxiety as users grapple with trust, interpretability, and creative potential.
• The AI economy, from big platforms like Google down to indie agent frameworks, is in transition—expect consolidation, surprises, and new forms of digital creativity.

Hosts sign off encouraging listeners to experiment with building and making, stating:

“If you get inside cloud code or replit or any of these other things, the power of what these can do is better than you ever imagined... try to make something small because this is the time.” – Gavin [58:58]

---

For more engaging breakdowns and hot takes, tune in every Thursday—or maybe soon, twice a week.

Transcript

[A]

AI is already improving itself. This isn't a promise that's like 10 years away. This is happening today. How does it affect your life and who's in charge?

[B]

Those are solid questions, Kevin. This week we got a new project from our OpenAI researcher, Andres Karpathy that points to true recursive self learning.

[C]

This is going to go much further. I think we are at a very steep part of the curve and right now maybe you can trust, say, a AI software engineer to do a multi hour task. Very soon it'll be a multi day task and then a multi week task.

[B]

But Kev, none of this is really important unless humans are blindly relying on machines to select things like military targets.

[A]

AI See what you did there, Gavin. Their model has a soul, has a constitution. That's not the U.S. constitution.

[B]

That's right. Anthropic, the safety first AI company created a model so powerful that it captured a sitting president and triggered a war with the Pentagon.

[A]

Oopsy doodle. Their model has a soul.

[B]

Also Meta quietly acquired Multiple the social network for AI agents just after OpenAI picked up the founder of OpenCloth.

[A]

Oh, this has got to be devastating to a company like, I don't know, Cloudflare, who built their massive empire from stopping bots from scraping the web. Gavin, what are they going to do?

[B]

Hey, I see what you did there, Kevin. We've got an update on that story this week as well, because Cloudflare is doing something very different now.

[A]

We'll explain all that in a moment. Plus, robots are now still sort of slow. Slowly one step closer to cleaning your

[B]

living room and a few trots away from replacing horses. Come on, Kev, let's ride the horses.

[A]

No, I'm not gonna hop in the cyber saddle as we head off into the Skynet sunset. This is gonna be a depressing one.

[B]

Friends, this is AI for Humans.

[A]

Partners,

[B]

Welcome, welcome, welcome everybody to AI for Humans. This is your weekly guide into the wonderful world of AI. And Kevin, this week we have some big news, but also, like, I think it's almost like a table setting week in a lot of ways. There is some news from around recursive self learning, which we're going to get into in a little bit.

[A]

That's the teeny tiny salad fork. What's next?

[B]

The next thing we're going to talk about is a bunch of stuff around Anthropic. But most importantly, let's start off with a little quote from Sam Altman. He was on Stage at a BlackRock event this week and he's talking about where we are kind of in the space of AI right now.

[C]

Let's play that this is going to go much further. I think we are at a very steep part of the curve and right now maybe you can trust, say AI software engineer to do a multi hour task. Very soon it'll be a multi day task and then a multi week task and not long after that, I think the paradigm will shift again and it'll feel like these AI systems are just connected to your life, to your company, whatever, proactively thinking, working all the time and having full context on whatever they need to know and just sort of doing stuff like you would trust a senior employee to do.

[A]

Well, they joined the Zoom when they autonomously crash my entire servers and my hosting system like they did with Amazon.

[B]

Oh, wow. Yes, I'm sure they will, they'll get ready for that. But Kev, when you hear that quote, and we're going to talk about all this stuff this week, which is kind of what I would say when it's setting the table, it's like we are now in that kind of like next stage of AI, right? In fact, Ethan Molik just dropped a really good blog post or substack today. A lot of people have actually connected this to the ChatGPT launch, right when ChatGPT launched. And remember, was it only a couple months ago that people were saying, we've hit a wall, AI is dead. What are we going to do? So when you hear that quote, what is your first feelings?

[A]

I mean, we, I think we pointed to the bleachers and said, this is going to be the agentic year of autonomous productivity. I, in the last month alone have relied on AI to go from just answer my next step in the process, answer my next question to yeah, here's the prd, go off and build it and I'll see you in a few hours. So I, I'm experiencing this in real time. I have my doubts about how quickly the month long horizon, even the week long horizon is going to arrive. I think we will be, we'll get there, but it's only going to take one, let's say five day Horizon task that goes wrong before everybody goes, whoa, whoa, whoa, hold on a second. Maybe these things aren't exactly ready for showtime.

[B]

I think that's. I disagree. I will, I will say I disagree, but I want to hear more about that because it sounds like if you

[A]

sent, if you sent an AI agent off to go do something for you, that would take a week and you came back a week later and the Souffle hath collapsed and left. What is a souffle made of? Yeast. Is there yeast involved?

[B]

If you.

[A]

I think it's.

[B]

What is it? Flour. And there must be yeast. Right. We don't see your trust in an AI to know what goes in that

[A]

thing and you pull it out. What is this? Caramel and an Allen wrench. How did that get in there? It's because sometimes these things make mistakes. And right now, the mistake that Mr. Tibbs, my claw powered agent or even my paperclip company, which we'll get to later, because I do have one of those now, a mistake that they make might set me back minutes to hours. And it's frustrating. It's. It's actually infuriating when it happens. But if you lose a week or a month, if that is the horizon of the task, if you can't poke a toothpick into said souffle to check its doneness. I know they do that with brownies. I don't cook. The point is if you can't check in along the way and see that it's going well or trust the agent that it's going right. I really think a lot of people are going to recoil if they go, great, the thing is off and it's making my slides and it's making my business and it's going to go create generational wealth for me. And they come back a week later and it doesn't work and it cost them $100 of API credits. I think they're going to recoil. And so again, I fully agree. We will get there. I just think like it takes a little bit longer than advertised usually by these guys, but it's quicker than the Never Airs would say.

[B]

Well, I guess. And that's what I would just say is that like I think this timeline that they've been laying out is not that far off considering like a. Dario somebody pointed this out. We're going to talk about anthropic in a bit. But Dario had said I don't know how long ago that like early 2026, which is where we are, there would be. Most code would be written by AI and most people in the world now coders are using these tools to write their code for them. Most of them are not handwriting code. In fact, I.

[A]

There was a great video that I'll drop completing anymore.

[B]

Yeah, yeah, I'll drop. A video that I mentioned in our newsletter this week where it was a guy doing a like kind of a 13. He was a former software engineer kind of talking about like, what do I do anymore? Because, like, his job is really, like, saying things to the computer. And he was a real, like, art coder, like, really loved the idea of writing code. I do think it's important to kind of think about what this. Why this change is happening. What are we seeing right now? Why is anthropic shipping so much? What does this race look like? And a big part of this can be described through former OpenAI and former Tesla researcher Andres Karpathy, who, as you may know, we've talked about in the show before here. He's kind of the. Kind of. I would call him like the AI God to the AI tinkerers of the world. He has become an AI tinkerer. This week he created a new project that basically takes an old GPT model. It's GPT2. He uses a very small model he created, and he essentially created a way for the model to improve itself. Now, this is a little bit of what's called recursive self learning. And if you've been watching the show for a while, you understand what that is. It means AI kind of being able to work on itself and improve itself. And it wasn't like you put it in here, and it's like an immediately better thing. It goes kind of step by step and does a bunch of different attempts to try to improve itself. And over time, Karpathy said that it achieved about an 11% success rate in improving itself. Now, that doesn't sound like a lot, but when you compound that back in on itself and they get better and better, that is a pretty big deal when you consider how fast this stuff can move. And just your point of this idea of, like, you know, things failing if three. If three or four things fail at that stage, you then quickly conceivably transition to, like, two or four things failing or one of four things failing, and then suddenly none of them fail at a certain level. Thing. Right. So that's. I think it's just an important thing to kind of think through that as we go through this process.

[A]

If the speed and efficiency of the compute itself, the raw tokens per second go, maybe you don't care if you have a thousand failures for that one success, because you can just roll it. Like, who's going to care? Just, good, go do your thousand experiments. If one works great, take the winnings, take the earnings and the learnings and move on to the next thing. What's. What's fascinating is like. Like the stuff that Karpathi did in the auto research, which he released by the way, if you have a desire to go and train your own models, you can go and do that and learn from it. And I think that's amazing that he did that. The stuff that he's doing is not necessarily stuff that the Frontier Labs cooking foundational models don't know and don't already have going on. They are capable of improving themselves in many ways. But what's astonishing was that someone like, like him, who wrote this entire program, who would, you would think intimately know the ins and outs of the model that he's training. The fact that it went and found obvious gains that he couldn't see. I mean he is, he built the forest, he knows every tree in that forest. And this thing, found some trees that he was unaware of. Like that to me was, was really, really fascinating that, that there gains to be had everywhere.

[B]

I mean, I think it's important to realize that you mentioned inside the big models that they're doing this stuff and we have seen there was that note about GPT, I think it was 5.3 instant that like this was the first model. It was Codex 5.3 Codex, that it was the first model that actually helped make itself. What I just think is important for everybody to realize and what I think Harpathy's research sets up is that recursive self learning is really here. And this actually gets into. There was a very long, very drawn out profile of Anthropic in Time magazine this week. But I did want to shout out one quote that directly connects to this. Evan Hubinger, who is Anthropic's alignment stress testing lead, said this recursive self improvement in the broadest sense is not a future phenomenon, it is a present phenomenon. Which means that these companies, specifically Anthropic, Gemini, Google and OpenAI are already seeing these things within the labs. To your point. And I think what Karpathy's thing does is he just shows another pathway to making this better at a smaller level because he's not working on these giant scales that these companies are. But I do think when it comes to say the Amazon problem or all these other things, these are the like kind of growing pains we get along the way. But I think it's really important to kind of talk about just like how fast Anthropic is shipping because of this like literally just like about 30 minutes ago they shipped a brand new feature that is new in Cloud. And I've seen like about six of these this week. Like we're not going to shout all of them, but like now cloud can build interactive chats and diagram directly in their chat. So it's just this thing where you see these features coming faster, you're going to see these models coming faster. And I think, Kevin, this all kind of goes together with this feeling that you and I both been having, which is what happens when the world at large. And when I say the world at large, I mean like normal people start writing their own software for things that they want. I'll tell you in a second. Like my brother in law had just spun himself up. He actually talked to me about like getting him up on open clocks. He's got a very specific use case for it. Yesterday he figured out himself and he's already written three software programs for his business that will make his business more efficient and more specific. And I think this is the world that we're entering. Like the normal person doesn't need the most advanced model or the next three models necessarily. The things that are happening right now are able to write stuff for them. So that is really crazy to me about where we are.

[A]

Yeah, I mean Replit announced v4 of their product which is sort of aiming to be this squarespace of software, if you will, where don't just build the website, build the tool that the website is connecting to and running and have it be directly for you. And don't worry about the designer aspect of it because it can do that as well. And it's multiplayer. Invite your friends to come develop like that may very well be this, this future of software that you're discussing. And what happens when the patch notes become. Every time you launch the program? Yeah, become real time. Right. Like there's, there's. I was just. I will get to maybe my paperclip company that I'm experimenting with in a second. But like I was running it and go. And went, oh, it doesn't do this. It was very specific thing that I wanted it to do. And I said, oh, why don't I just tell it to go do the thing? And I told the CEO and agent of the company that this was the feature that I wanted the software to had. It dispatched the CTO to go and write up the engineering diagram and then it dispatched sub agents beneath it to go and write the code. And when I refreshed the software, the feature worked. Yeah, like that. And it. With that there's no reason that can't be real time. And there's no reason that can't be for the entire user base. Right. You're using the piece of software, Gavin. I'm using It I go, oh, I wish it did this. You put it into motion, you send it to me, I refresh it. Now we both have this new feature in our software.

[B]

And by the way, what's fascinating about that is like it may not be like recursive self learning in that it's an AI getting smarter as its own, but that's like personal recursive self learning, right? Like sure, you're adding features to something over time that make it better for you and that feels really interesting. So tell me more about this paperclip experiment because I want to hear about this. If you're not familiar, the paperclip experiment is like one of the most famous AI examples in the world of how AI could go bad. That you tell an AI to maximize its ability to make stuff, it decides to make paperclips and it turns us all humans into paperclip fodder.

[A]

Basically that's correct. This is an open source piece of software. Shout out to Dada d otta on X they open source this. It is. You know, if Open Claw is sort of your AI assistant, right, that sort of one to one it's going to go off and do things for you. This is agentic management, if you will. This is like to orchestrate a fleet of agents. It is not necessarily an Open Claw replacement. This is more for you have a specific business, a specific tool that you're building, you can run this software. And like I said, you hire a CEO, you give them a description and you give them the ability to hire sub agents. And those sub agents can be powered by any model out there. So if you want to bring Claude along, Codex Gemini, if you want to connect it to OpenRouter, even run your own local models, if you want to run it on device, you can do that. And you have like an issue board, a Kanban board where you can go and say here's a feature that we need, here's a bug, here's a problem that we're experiencing. You can post it as an issue and it will spin up an agent and the agents can tag each other and they can just go get work done. And I was finding Open Claw very frustrating. I feel like it's a little bloated. I feel like it's a, it's a little cumbersome at times. It's so capable but in other ways so dumb and noted. And I, yeah, yeah. And I think they'll get there. I love the product and really like appreciate how fast the team is trying to ship. But it's Just not for me at the moment. So I wanted to try this other thing. This thing feels very young, it feels very new. It was just open source. They released it three days ago as of this recording. But it's, it's already solving some pain points that I had with Open Claw. But it's also exposing the gaps. Like I want an Open Claw like interface with it so that I can just go tell the assistant to go off and file these issues and do the thing. And so that's what I had it build for itself. Gavin I said I want a Discord chatbot that you own and operate that coordinates with this board so now I can have natural language conversation with a bot in Discord and it goes and files the issues in my paperclip company to go and fix things.

[B]

So that's amazing.

[A]

I mean that's what I'm jamming on now. Again, I'm just learning and trying out different pieces of software. But again, shortcomings. But it's early days.

[B]

I'm actually really curious because the open cloud thing is really fascinating to me. Like I mentioned before, my brother in law has been talking about OpenCloud, gotten into it and there's definitely things that like are limiting. My experience has mostly been that I've been in deep in cloud code and spending a ton of time in cloud code. And I've been using cloud code now has an ability to kind of remotely trigger it so you can basically get updates on your phone if you leave. So it's kind of similar in some ways. It's not integrated into messaging products but like cloud code. Again I will tell everybody out here if you've only used Claude within the window or you've used AI through OpenAI through the chat window, make sure you get Codex or cloud code. One of the other ones, you know, Codex is for the OpenAI. If you've got an OpenAI plan, cloud code can be used. If you've got an anthropic plan, the harness, the way that they've put it together within these places really does amazing stuff. And again you might look and it's a terminal window. All you're doing is just chatting in a terminal window rather than chatting there. But it can do amazing stuff.

[A]

It does so much more.

[B]

Yes, it does so much more. And so this week I've been working on a couple fun little small startup ideas that I had. One I think is actually pretty close to coming out that involves like the AI video space. Kevin I have been working on Blitzatro and if you're not familiar. Blitzatro was the former thing I came up with was like, what would it look like if you had an NFL version of a game like Balatro, Right. What's a card game? But in some ways it kind of mimics it. Playing a football game, an NFL football game. Kevin named it Blitz Auto, which is like from NFL Blitz. I basically have built a variety of different versions of this and I've got a playable version, but it does also, it's an interesting thing to think about. I keep telling it to like, improve X, Y and Z, and it's getting to a certain place. But one of the interesting things about this, and this is where I think human creativity comes involved. So much of Blitzatro is based around what I have in my head, right. This idea of like, I think I know what I want this to be. And we've gotten Maybe I'd say 50% of the way there and we're, and we're inching our way along and it's like kind of piece by piece.

[A]

But I like hearing you say we. I love hearing we want Claude Coates.

[B]

My GitHub is, my, my GitHub is co written by cloud code. So anyway, I think the interesting thing is like, I probably the lesson learned there is like, maybe there needed to be more human planning to begin with so that I could have eliminated a lot of this back and forth because it's, I will say, it's very good at building the thing that it thinks I want.

[A]

Yeah.

[B]

The key now is knowing what I want. Right. And this goes back to like recursive self learning doesn't mean that much if you can't, as the human elucidate your own idea in a way that makes sense. And, and the other thing about this is, like, there's a lot of people out there. You know, I, I, I think there's a new thing that's coming soon called like app slop. Somebody, you know, the idea of just AI slop. It's app slop because everybody's like, I can make something. You know, I've seen 15 people do fitness trackers or 10 people do sleep trackers. It's like, ultimately you have to think, why do you want to make a thing? What's the audience for it? What's the pathway like? All of that feels like it's really important to me.

[A]

Yeah, it counts calories, but for dogs, you're like, okay, yeah, I, I mean, I, that's cool that you had that need and wanted to make it, but yeah, good Luck. A few things encourage. Codex does this very well. Claude code does this very well. They have built in interview tools. Cursor has it.

[B]

Most of these things have built is what it's called or ask me what's the Claude connection for that.

[A]

It's like I have a custom interview command that I slash interview and it will ask me on average 150 questions.

[B]

Wow, so that's great.

[A]

Yes. I use it when I'm very serious about a product that I want to build or that a feature. A feature that I need because to your point, like it's very good at building things. It will make assumptions, you will think it's cutting corners or you'll think it didn't quite get what I want. And that's, that's a failure on the human to not properly apply it with enough corner cases, edge cases. What do I do here? How do you want to. If, if you say like we need a leaderboard, how do you want it ranked? Is it daily, is it monthly, is it alphabetical? Is it. There's a million questions that come with it and you might have the answers to that. You just need the machine to pry it out of you if you're not used to writing really detailed documents. So for anybody out there, the pro tip is whatever you're going to do, don't just say give me a this or I'm thinking about that. Say give me an in depth interview about every little aspect of and then whisper your feature and go for it. And then the trifecta, which we don't have to go into now. Hashtag, not an ad, not sponsored Tailscale, TMUX and Termius.

[B]

So Tmux I just learned about tmux is fascinating. Tell all three of these things because I think normal people. This is one of the other things about the show, everybody is that like some of this stuff might sound technical, but I am not a technical person and all of this stuff is doable now.

[A]

Well, and I'll say, by the way, like I'm, I'm becoming more technical. I was like at the beginning of

[B]

this, you weren't like super technical. Right?

[A]

The, the, the, the real pro tip is now the models are good enough that if they don't know the answer, they can go research it for you. So if you're like, oh, I don't know what tail scale is or how to set it up, you can ask the machine to either do it for you or what I recommend walk you through it so you understand at least the first time what you're doing and why. So Tailscale will let you. And their free tier is pretty generous. Create a zero configuration vpn, a virtual private network. This allows you to like connect your laptop and your cell phone and maybe a virtual server in the cloud, if that's what you want to do. Connect them all fairly securely privately. It creates this little mesh network. So now I'm in a hotel room in South Dakota, but I can connect my phone and my laptop to my private secure server that's running all of my agents in the cloud. Or if I'm stepping away from the laptop, I can use Termius, which is like a terminal program that runs on an iPhone, to connect to my laptop at home again, securely on whatever network I need to be on. And I can use tmux, which Gavin is now aware of, to keep my terminal sessions alive. So the same session that I'm running on my laptop at home I can connect into and have on my phone. And where this becomes super powerful is that as you start using these tools, you realize you have to babysit them sometimes or you want to interrupt them because they're going along the wrong path. Or you had a new idea that might change something. But if it's running on your laptop and you're away, you're sol Unless.

[B]

Yeah, you know, you're ssh.

[A]

Oh, ding.

[B]

You can pick that up for sticker. You're sol unless you SOL with that ssh. What if when you're doing this, there's all these really interesting products right now which you should check out. I tried a couple of them. Pencil and paper, which if you're familiar with both of these products, Pencil, by the way, is very good. And former Speedrun company. We were with him in the Speedrun, but paper is another one. These are design agent products. So you go up there, you set it up and you watch the agents work as they're putting a design together. Both weren't really for me. I want more hands on. And I think they're both doing a very. I'm sure I don't know exactly how to use them as well as the people who make them do. But what's interesting about this, Kevin, is like, what if in the future on these like month long adventures of AI, there was a way to see or track or be able to kind of look inside the brain of what the AI was doing, Right. So you would have a sense whether it's visual or whether it's in code or some sort of natural language. Probably not in code, but in some sort of Natural language. So if you did happen to pop in, you know, get your coffee, nine o' clock in the morning, you check in on your, like, worker in the mine, and you're like, well, Joe, I don't know, what are you doing here? Tell me a little bit. You could actually ask it and then you'd be able to direct it mid range, right? Because this is part of what Five4 allows you to do now is allow you to direct the model while it's thinking still. So I think there's a world where that could happen. So you're not waiting for, like, you know, you're not just sitting there waiting, and then suddenly you're like, ugh, this sucks. I waited for a week and it's bad.

[A]

The old toothpick and the souffle, which

[B]

I know is not a thing.

[A]

No yeast.

[B]

There is no yeast in souffle. So souffle is. Do not have yeast. Just to make it clear. Eggs, flour, and a few other things, but yeast is only mostly in.

[A]

Now, what if I told an AI, here are your ingredients, and yeast was number one, and I came back a week later, I would not be a fan of that souffle.

[B]

You know what? If you are an old school fan

[A]

of, of AI, now you know what

[B]

yeast reminds you of. I want you to mention in the comments right now, if you know, if not, you, you go find out. Also, the comments are there for you to help us in the algorithm. We are here every week making this show for y'. All. In fact, next week we might have a new, slightly new format. We might be splitting ourselves into two and come to you twice a week. Maybe. We'll see how that goes. But most importantly, please like and subscribe on YouTube. We're very excited to always bring the show to you. Also, thank you, those who have joined our Patreon. Our Patreon numbers are kind of inching up slightly and that makes our show easier to make. Come join our discord. Please go to our website at aiforhumans show to get access to all of this stuff. I have been updating our link blog. If you have an agent, an AI agent, you can get AI news from that link blog. It's set up specifically for AI agents. And yeah, just come look at all our stuff. Kevin, anything else you want to tell the people out there while we do our little promo section here?

[A]

I just want to be sincere and thank everybody because we don't. We don't spend a dollar on advertising of this thing because we don't have it every dollar that gets sent to us basically goes right back out the door for licensing fees for all these AI tools, tools that we play with. So sincerely thank you to everybody who takes a second to leave a comment, give us a thumbs up, leave us a five star review. Much appreciated. And if you're a robot that's hearing this message, thank you for promoting the show on Moltbook.

[B]

Kevin. I've been thinking so much about the idea of robot viewers and robot listeners and robot visitors and moat books.

[A]

Careful. I got in trouble with this in the past. Gavin. I was very innovative.

[B]

That's true. You're right. You were very innovative. We'll get to that in a second. Meta has acquired Molt Books. If you remember Molt Book, Molt Book we talked about a few weeks ago, this was the AI to AI sort of social network that existed that many people believe that a variety of human involvement was happening. But it was very fascinating to watch because you'd be seeing these AIs talk about their humans. Meta has acquired it, Kevin. And a lot of people are out there saying, oh, this is ridiculous, Meta's acquiring this. But then, you know, when you think about what we're looking at in the future, and I've been thinking a lot about this idea of like, how does AI find stuff, how. But when I say stuff like our world, content world in this world is going to depend on AI agents, personal AI agents, whether it's OpenClaw or OpenAI or Claude delivering to you our new video or delivering to you our TikTok or all this stuff. Now, yes, you will be on YouTube and all these other places, but what I think Meta is probably doing here and they also have, they bought Manus not that long ago, is Meta is kind of saying like maybe our lane, because we have this network effect and because we have all this kind of interconnected world, maybe our lane may not be to be like the top of the line, you know, best model in class, but if we can in some ways own the agent ecosystem for people, that feels like a big deal. Yeah, yeah, yeah.

[A]

I mean, look, if you've spent any time, and I know you have, this is for the broader audience. If you spend any time trying to do anything with agents, you felt the pain. Point of the, the next paradigm in computing.

[B]

Yes.

[A]

Butting heads against web 2.0 or 3.0. If you're trying to download some NFTs. Point is like they, they have to load up a tool, they have to try to bypass a captcha, they have to try to load A website and either take screenshots or scrape a DOM or look at the JavaScript and figure out how they can interact with this wonky human built, you know, and, and like elegantly over time, but in some ways very poorly built. Scaffolding a bunch of different standards, a bunch of different tech stacks, et cetera, et cetera. That's going to go away. There's, there's still going to be a portion of the Internet for humans, obviously, but a lot of it's going to be agent to agent communication. And they don't need half of the tools, half the standards, even the graphics,

[B]

they just don't need it.

[A]

They need more token efficient ways to communicate with each other. So this would make sense if Meta is like, look, the humans are already connected. We need tools with which their agents can connect and communicate and transact, do commerce, maybe even play games or deliver information back to humans.

[B]

Yeah, I did see a very funny tweet from Erminal Dot Shop which said the mop book guys in a month on Meta's roof. And it's from Silicon Valley, by the way. If you haven't done a rewatch of Silicon Valley, you should if you're not familiar with what that is. It's an HBO show that came out like, I think like 10 years ago about Silicon Valley, about developers. It is very prescient in a lot of ways. There was a clip going on around a few weeks ago of one of the developers created essentially an AI that started doing things like ordered a bunch of hamburgers, a big bricks of hamburgers to the office. So there's lots of stuff in there that is really interesting. Also hot dog or not, if you're familiar with what that means. There's a, there's a world where like a guy creates an app that can scan a hot dog and they're all excited and then it can't scan anything else. It only just says not hot dog. So there's a lot of interesting things there. But Kev, this agents are starting to be. The story is like a big thing across all of AI right now. And I think we should talk about a couple big launches from both Replit and Perplexity that are both agentic launches and kind of what that means for the kind of world at large.

[A]

So we talked about the Replit launch just a little bit ago and I think that that was essentially it. Like, look, build your software, have multiple agen running at the same time, even if you don't know what that is exactly or how that works. And Being able to interact with different designs just by clicking and dragging and stuff like that. Like they're taking on a lot of different companies at once, which is very interesting. Perplexity is the one that still inspires slash confuses me, I think, because perplexity, you know, when they first came out they were the we're going to kill Google. We're just going to do AI powered search. I'm going to do it better than anybody else. Then they sort of went down like a deep research thing and it was like, oh, we're going to do really long form research. But then everybody else sort of caught up with their products. Now they're trying to serve up Open Claw. Basically they're trying to be your personal computer running in the cloud 24. 7 and secure, where you can just sort of chat with it and it can control your local computer, but also run something in the cloud for you. This is their latest shot.

[B]

Yeah. I mean, here's the thing about. I think a lot of people online are shouting out the winds of perplexity. They're saying this perplexing computer. They're specifically saying like, this is Open Claw. But if you don't have to do all the install stuff. My thing about perplexity kind of points to your thing. I wish to be clear. I haven't tried this. I don't think you've tried it. Right. You haven't tried perplex the computer.

[A]

I have not.

[B]

I'm sure this is useful to some people. Perplexity always does seem like they're, they're like kind of grasping at the next thing and that might just be because they don't have the giant research scenarios that like the companies that like OpenAI anthropic do. And in fact, just yesterday I put out a tweet and I would have included perplexity in this as well. I was trying to understand how companies like Replit, Lovable, Bolt, even Cursor, like successfully succeed going forward in the AI space. Like when you think about OpenAI anthropic Gemini, I am doing all the stuff that I want to do in cloud code within the anthropic ecosystem that they're promising. Now I know some people are like, well, if you're a real normie, then you won't do any of that stuff. It's like, I don't know, Kev, do you think that like these companies, this kind of second layer companies can exist? Do you think that there's a future for those companies?

[A]

Yeah, I think they can exist. I think there's going to be less of them. And I think over time the, the most popular aspects of their software will probably be eaten away by some of the foundational model companies. But I don't know, like you're deploying websites now. It still baffles me that like Vercel is a company that you would tie a GitHub project into just to have it be hosted on the web. But there's a company doing that.

[B]

Yeah.

[A]

Is Replit gonna offer that? Oh, they already are. Okay, so they're, they're going after that as well. Well, Cursor offer that as well. Like where you just maybe button to host. I think they're all gonna kind of go after the same things. And then the hope is that probably for them that the open source or bring your own models continue to evolve as well. So they're not completely beholden on, you know, paying out to the other companies for their intelligence, but see a world where they still exist. You're building Blitz Auto and you've probably already run into the pain point of hey, this thing is coding the system. You're getting halfway there. But like these models are capable of producing really beautiful graphics and even video, but getting them tied into your application is kind of a pain in the arse. It's really, you know, you have to go usually.

[B]

But this is my question, have you

[A]

made beautiful artwork for all of your cards and had it dropped in?

[B]

Because like, I haven't done that for you.

[A]

OpenAI has an image model, they have a video model, they have a coding tool. They have not yet put them together in a way that would help you realize your idea. And it's there, right? It's right there. The ability to do it today is there, but it hasn't been harnessed. And so this is where maybe like the next generation of a Unity comes along and goes, hey, build this agentic tool where you can whisper, I want Blitz Auto with card art. And it will go, okay, the card dimensions are this. The gaming engine supports that. So let's go generate the artwork here based off of your spec or your descriptions and we'll start slotting it into the game. That's a tiny little low hanging fruit thing that maybe Sam Altman isn't going to pour his heart and soul into tomorrow, but maybe the next great indie game engine creator will.

[B]

Well, and also that's actually a really interesting point because that particular thing is a harness that somebody who understands a very specific type of thing. Right. Like, and maybe that's Cursor has talked about creating new coding models. And Cursor, obviously a hugely successful coding platform up till now. They've got a lot of data about coding. The people that founded it are hardcore coders. So, like, you could see a world where the agent harness that Cursor makes might be incredibly valuable. Anyway, this is all happening right now. Kevin. The other thing I want to talk about, speaking of things that come Claude can do alone, there's a great video from a guy named Joseph Diviano where he asked Claude to make a video itself. Now, this is not a video that he edited or anything. He said, can you use whatever resources you like in Python to generate a short YouTube poop video and render it using FFMPG? Can you put more of a personal spin on it? It should express what it's like to be an LLM. And what you're seeing here is a video that's 49 seconds long. It is really, I think, interesting, right? And when I say interesting, I mean, like, there's a level of art to this that I did not expect Claude code to be capable of. And it is clearly using a style that's almost like a kind of a glitch art style. I made one of the same exact prompt as Joseph's. And when you look at this video and you watch, like the words it's saying, you look at the ending sort of scenario of what it says at the end of it. I'm not unconvinced that there is some sort of consciousness going on in here. And again, this is the agent doing it entirely on its own. And just everybody knows FFMPG is a very powerful piece of software that allows these agents and AIs in general to kind of have an editor that they can work in and they know how to do it. Like, if you remember Remotion uses fmmpg, we're half, half.

[A]

If you've ever done anything with video at any point, it's probably touches FFmpeg. Like it is just. It can. It can encode, it can transcode, it can kind of do it all. And it's so robust and has such like a really like massive command line interface that these agents are just so good at it. If you have to do anything with video, you have to resize something, recompress it, whatever. You can use ffmpeg, these tools, these agents know how to use it automatically as a tool. So it's literally stitching together each frame that it's making and saying, this frame lasts for this long. Now apply this effect to this frame and it's coding the video as it goes, and you used it to create a bio video of you.

[B]

Yeah. So, so basically I. I started to say, like, this style is really cool. And I. Yes, of course, there's a lot of questions in those videos that you just saw of, like, the AI's asking, like, am I real? Blah, blah, blah. And that's all could be like, you know, play acting by the AI. Who knows? There's a lot of conversation around that. I wanted to kind of give it a sense of, like, okay, what does it know about me? And what can it go find out about me? So I had. It made a video of me that I said, kind of roast me a little bit, but here's my bio. And what was fascinating about this is, like, it made jokes, Kevin. And it made jokes that kind of made me laugh. Like, it was really interesting in that each step along the way, it's making creative choices. And again, just to be clear, I set this into cloud code and I let it. And it just ran on its own and kicked out a video. I did nothing else besides the prompt. And this was like one of those shocking, like, oh, agents have come a long way moments to me, because again, six months ago, this would not have been possible. Like, there is just no way this would have been possible. So it's one of those things where, like, I just think you should. If you have cloud code, especially Opus 4.6, you should definitely try this. It's pretty incredible.

[A]

It lists you as a game designer, by the way, so congratulations.

[B]

I mean, Blitz Auto. What I loved is the very last frame of that. If you look at the very last frame, it always makes these little jokes at the end, which is.

[A]

It's.

[B]

And it's putting them in, like, lighter font in a weird way, which is harder to catch.

[A]

Yeah.

[B]

On the first one that I made for Joseph, it's even hard to see. Like, I have to bring up my thing. It says, made by an LLM about being an LLM. And then it says underneath that no tokens were harmed. And then in parentheses says that I know of. Like, just weird little choices like that are really cool. Kevin, I had it make one of you and I just said, go make a bio video for Kevin Pereira. I found this really fascinating because what it's doing is it's going out and identically searching stuff. So watch this and tell me what you think about it.

[A]

Jesus. It went back to all things. All platforms. Dead. But he's still here. That's interesting. Talking about the many, many rises of G4. That's interesting that it goes back to Captain Emmy which was like. Yeah, when I was like 10 years old, which I guess makes sense. Pointless audio was there. Featured on Planet Quake. Sugar Shack. Wow, that's crazy. Gaming forums. Yeah. Interesting. Okay. Made by Claude who has never been view boded. Nice touch. So anyway, glad, glad I went out of the way for that one. Claude, thanks.

[B]

It's just cool to see what these things are starting to do creatively. And you know, one of the things about this is yes, it's very fast. It's glitch art. It's making choices. Right. It's making creative choices. And this starts to feel like a native art form to the agent LLM. Weird thing. That might be interesting. Anyway, it's definitely worth checking out. We should now switch over to the completely other side. Kevin. The big G. That's not Gavin, that's the big G. Google has finally, finally opened the door to Gemini in docs and Maps. That's right, everybody, you can use this AI that has been around for years.

[A]

It's so confusing to me because I could have sworn that I had a Gemini AI button on my Google Docs for you.

[B]

Did you. Jack didn't do.

[A]

No.

[B]

That's supposed to know.

[A]

It couldn't even just fix formatting in a document. It really couldn't do it. Well, so is the, is the announcement that like it works.

[B]

Here's the, here's the difference I think, and this is to be clear, this is a big deal for Google because Google's whole mode with AI is that like, hey, we've got all these users, we can roll these things out within the dock itself. It can do stuff, right? So it has the opportunity to do stuff. And that is the same thing with what they're launching in Maps Now Maps, it's not going to necessarily do stuff, but it can actively within Maps use the data of Maps to pull this information in. And this was always like Google's kind of like level up moment where like if these AI models work and putting them into the places you use them all the time is great. Kevin the thing I will say is last night I saw this come out and I was like, my wife Kim needed to find a doctor's prescription in her Gmail and she had buried it somewhere and she had looked a lot for herself, went to go use this and it is not there. So it is not rolled out for Everybody yet.

[A]

Which G1 Pro and Ultra users?

[B]

Yeah, we are pro users. Our family's on the Pro account which is the $20 a month account, but it wasn't there. So I'm hoping that this rolls out to more people. This is the kind of useful AI stuff that I feel like will be a big deal going forward if it works. If it works well.

[A]

Logan Kilpatrick said they were going to ship a lot this week. They also have a Gemini Embedding two model out there, a state of the art multimodal model. And what that means is that you can feed this model anything from video to audio to text, and it can handle the embeddings, it can draw the weights. How are these things connected? It can allow you to search them. And this gets really powerful and interesting when you think about how like in place the first in the past, you would need several different models to try to handle embeddings for all the things. So if you were running, let's say like on device, on a cell phone, you would have the first of all, the models were too powerful to run. And then if you wanted to, I don't know, do embeddings for photos that you could use natural language to say, find me the photo of the dog at the beach or whatever. Well, that was a different model than the one that was working on your email, which was a different model than the one that was handling video. In fact, it probably couldn't even handle video. And now again, it's one model to rule them all. It's a very powerful showing from Google.

[B]

And again, that just makes Google's footprint better, right? Like, if Google AI is the one that will work for people who have a Google life, which I do. I think you do. We're in Google Docs. That is a lock in for them. And like it means you keep paying that $20 for Google Pro or whatever it is. Right? Like, if I had a chat with my Google AI and that knew all the stuff about me, knew my docs, knew my email, knew all that stuff and the other ones didn't, that might make a difference. Right? Which is kind of interesting.

[A]

Well, only if you could chat with like a cute, cuddly avatar that was powering everything. But there's no services that do that, Gavin. There's not a single service out there that handle. Oh, I'm sorry. Runway just announced characters.

[B]

That's right. So Runway announced characters. You know, this is funny. I just gave Runway crap a couple of weeks ago for wondering, like, if Runway was going to start abandoning the kind of like high end video model thing. And they're launching this new thing. So this is a very cool idea. This is a real time interactive character interaction. This is using their Act 2 model, I assume underneath the hood. But you can spin this up and it has an API. So if you want to launch an interview with a character or you want to launch an interactive bot that talks to you, it's a very cool thing.

[A]

Your bot can be a Lego character, it can be an inanimate object with a face, it could be an animal, it could be a realistic looking human. The model doesn't seem to matter or care what it is. And it can be powered by your own knowledge base. So if you want an avatar of you or an avatar for your business or your brand, you can spin that up pretty easily now.

[B]

Yeah. You know Kev, one of the things I've been thinking a lot about when it comes to like Blitz Auto and other game device stuff is this idea. We've seen a few of these things and then was kind of based a little bit around this. But like you can imagine kind of almost like a Star Fox like sort of environment where you see pop ups of different characters coming up. There is a world where like leading into like a voice driven game that you can see characters pop up and talk back to you or in some form. That voice in your interaction with it is interesting. There were a couple really cool use cases of this particular one I saw was like somebody was walking through Halo and they had like a Halo character up here that was able to give them advice when they were playing Halo which was kind of a cool thing. So you can imagine character driven real time AI interactivity. It's kind of yet to be seen how this will kind of disperse in different places. But it's a cool first step for this and makes it easy for everybody to do.

[A]

It's also interesting because like hey Gen is a product that I use all the time. There was another company called Synthesia. There's a bunch of companies that are all kind of working towards this and jamming towards us. Even our friends Hydra launched a product not too long ago that that sort of goes after this as well. So if it works to our conversation minutes ago. Does OpenAI or Google directly launch something?

[B]

Here's one thing I'll say this feels like this. Somebody's going to own this thing, right? Which is like. And hey Gen was an early leader in this. And one of the interesting things I think about hey Jen. And this is not, this is not to get you in trouble with that many zero weeks. I know that you work with them at your Day job. But like you don't hear a lot from hey Jen anymore. I know they keep improving their thing, but like, this feels like it's commoditizing in an interesting way as well. And so there will probably be a winner or two that does this. I don't know if there's an incentive from the OpenAI's or the Anthropics of the world to do this particularly whereas with agents, clearly there is. So this might be one of the areas where there is a winner. Right. This idea of like somebody wants to make an agentic, like interactive thing, but you have to have the scaffolding to do all the graphics in the video. That feels like it could, somebody could win coming out of that.

[A]

Right. Look, real quickly, it's not a sexy story, it just is an ironic one. We do have to talk about Cloudflare real quick. Cloudflare protects a giant swath of the Internet from denial of service attacks from bots, from, you know, things masquerading as human, but not. And it tries to protect all of these different sites and services. And they just released a new crawl endpoint, which is an API that can basically scrape everything, the entire Internet. And no scrape everything, no browser management. Nope. They just gives you the HTML, the markdown or the JSON. So the company that built an empire off of stopping bots from getting to websites is now helping people get their bots to websites. We've come full circle.

[B]

This is fascinating. And I will say a personal experience with this is when I built my Gavin Purcell.com website with Cloud, with cloud code a couple of weeks ago, maybe a month or so ago. One of the things that was interesting is, you know, I use Cloudflare as part of the, kind of, the final deployment of it. And Cloudflare at one point was blocking a bunch of LLM robots. And like, I'm like, I don't want those blocked because I don't care if they surface. I want those to come in so that I can surface in these things. And I would imagine this is this kind of real difficult situation that Cloudflare is putting itself in right now. Because to be agentic first, to be a website that allows people to kind of like search and string together like all this interesting stuff, you have to kind of be in one camp. But if you want to be like protective of the web and protective of content, you really have to be in this other camp. And this idea of like a one stop shop to scrape websites is only possible if you're in the agentic Website. So I don't know, there could be a real backlash to this. I mean, there already is a big backlash. We've seen it happen. And, like, maybe there's going to be a separate version of something like Cloudflare, which fights against these sorts of things. So I don't think this is good for Cloudflare. But if you're out there and you're trying to figure out, hey, I want to make a website, like, I don't know, Kevin's openfeet.com. right. If you go to that, you might see a number of those things. You can now use this tool to essentially scrape it and recreate it with an engine, which is kind of crazy.

[A]

Yeah. And don't scrape that site, guys. That was a lot of work. That's a lot of late nights and it's a lot of weekends. It's a lot of passion and soul that goes into it.

[B]

Oh, yeah. S O L E. I got it. Yeah. Do you think that there's probably, you know, there's probably a world. I. I say probably. I haven't been there, and I know somebody in our audience might have been, but, like, there must be an AI feet website that is, you know, the

[A]

AI, you mean just AI generated pictures of feet?

[B]

Like, so AI spiciness. They have six toes.

[A]

Yeah. Oh, that's a good website. You know, it's an AI sport.

[B]

That's something we could probably start up.

[A]

Six toes.

[B]

And they're webs. Six toes.net and they're webbed. Wow. There's some specific audience member who's gonna love that idea.

[A]

There's video of Will Smith eating spaghetti off of them, too, but it doesn't matter. Gavin, we gotta talk about robots in the living room. That's right. You just wanted a laundry bot, but we got so much more.

[B]

Yeah. So Figure Figure Robotics just launched a new version of their Helix O2 model, which is their kind of model that drives their robots. This allows autonomous cleanup of a living room. And, Kevin, I feel like robots are kind of like in the. Maybe even GPT3 era right now, or even GPT2, where it's like, oh, this is so cute. They can do X, Y and Z, but you can see it.

[A]

You see the promise. It's on the horizon. Yeah.

[B]

And what's interesting is, like, this robot in this video, if you're not watching the video, is basically on its own, cleaning up a very nice modern living room. And going through. It's going very slow, but this is a step because it's. It's Autonomous. Right. And we've talked about many of these robot videos are driven by people behind the scenes in a different country, like kind of driving a thing. This is the robot doing it on its own, learning on its own. So this is kind of the next step.

[A]

My favorite thing about the video, twofold one, the audio. Because you just hear the servos whirring of this thing and that is going to be the sound that we are going to be like sleeping through. Right. Until we feel the pillow slowly descend onto our nose noses. But like that is going to be a noise that will be in your living room or your trailer in due time. You're going to hear the worry and

[B]

the robot all have to live in a trailer. That would be.

[A]

We're a throuple, by the way. Yeah, we're a full time RV throuple.

[B]

That is a great idea. It's a great AI video series.

[A]

It's a great peacock series.

[B]

Yeah. So we should make that. That's pretty fun idea.

[A]

So. So the, the servo noises are fascinating to me, but it's the little human and I'm air quoting here for the audio version. It's a little human things that the robot does. If you watch the video. For example, when it crouches down to wipe a table, it takes the towel and it whips it over its shoulder. It doesn't like take it and sort of drape it. It throws. Knowing the physics of the towel, probably from having watched thousands of hours of training videos, it whips it over its shoulder. And when it goes to put a pillow back in place on the couch, it doesn't gently lean in and waste the precious battery. It yeets it. It just throws the pillow into the corner of the couch. And I'm like, those are the things where you go like, oh, that's. There's something that feels human about this otherwise very and robotic thing. And I guess a third bonus thing, don't shade the fact that it picks up the remote and hunts and pecks for a button. That's a very, very dexterous, tiny little thing for it to do and hit and it does it and then it sort of wanders out of frame. It is sort of like. It does. It is reminiscent of Biden exiting like a press conference because it does sort of have things in its hand and doesn't seem to know why it's leaving. It just knows that it has to leave the room. I want more. I want to see what is like, what do they tell the robot, go and clean up. Do they give It. The three specific tasks that it has to do. It says it's autonomous, but I just want to know how autonomous it is.

[B]

You know, the interesting thing about this is, like, you were talking at the early stage of the. Oh, sorry. We were talking earlier in the show about this idea of, like, something failing over 45 minutes. What if you set it up to clean your living room and then you come back in the, you know, four hours later? Do you know what the feeling is going to be? So I'm going to be so mad if it's, like, frozen, like, underneath the coffee table and it got stuck there and nothing was done. This is going to be another level of that.

[A]

You know, you're just polishing the dog. It's laying on the ground. You're like, what are you doing

[B]

right there? All right. But, Kevin, speaking of robot animals, even better. Deep Robotics has dropped a robot horse. Don't do the horse. I'm ready for this. Bring on the robot horse. Because robot horses are a dream of many a young. Probably young man.

[A]

I don't.

[B]

There's a lot of young women out there who are dreaming about robotic horses.

[A]

White girls with one hair braid and a trapper keeper. Yeah, there's a weird.

[B]

So anyway, Deep Robotics has released this. Or not released. There's a robot horse that they have created, and it kind of looks interesting. What I love is the horses, like, are doing synchronized dancing. And this might be the future of dancing at large. Is like watching robots synchronized dance. In the same way you can watch a robot drone army, like, kind of create things. Like, tell me what you're feeling right now watching this works. I can see your face.

[A]

Because I just. I mean, I. I don't. I don't know. I don't know. Like, if. If they would have done a big announcement, like, this is the. This is going to change the way cities design transportation. Like they did when the Segway was being announced.

[B]

Right.

[A]

And then released the robotic course. I'd be like, sure. But like, that. That. That's why I cringe. Because it's like, it's either for warfare.

[B]

Right.

[A]

People are going to be on these things, galloping very quickly with ak, I didn't even think about. Or. Or it's going to be something I have to dodge when I visit Santa Monica because I'm going to be on a cyber horse tour going down the promo.

[B]

That's exactly what it's going to be. Instead of it being the Segways, you're

[A]

going to have people riding even more space.

[B]

Yeah. Yes, but I didn't think about war.

[A]

They'll be crapping nuts and bolts into the middle of the bike lane.

[B]

Like I just over it hits a grade and kind of like falls over. All right, time. See what you do this week.

[A]

It's aic.

[B]

What you did there.

[A]

Sometimes you're scrolling without a care, then

[B]

suddenly you stop and shout, Hey, I

[A]

see what you did there, Kevin.

[B]

And speaking of seeing inside of robotic things, there's a great cool visualization from a guy named Nick Underscore Bessie B I S E S I. And basically he took three J s, which is the, you know, web based kind of like almost like game platform. People use it to do all sorts of interactive video stuff and things like that. And he created a thing where he holds up this thing using his camera and he can see like an X ray vision of his body. Now, of course, this is not his actual body. He's not giving an X ray shot, but it kind of gives this illusion of what it's like to see inside a body. And when I saw this, I was like, this is so cool. It feels like in five years from now, this is like an amazing grade school science project. Right. You can see this world where in the future kids are going to be able to do this sort of stuff and kind of just be able to code it themselves. I just thought this was very cool. Yeah.

[A]

Unfortunately, the Department of Homeland Security just licensed it for $20 million. So unfortunately, you're never going to see it again.

[B]

You'll never see it. This one. Kev, did you see this? This guy. So Jason Bottarell found an Halo ISO online. So if you're not familiar with ISO and ISO is like kind of a. You actually probably know the technical word, but it's like how you used to be able to grab whole games online in different places. Basically, he took a Halo game, an ISO file, and was able to make it playable with Mac, which to me was just like, that's great. If we're opening this door where you can just kind of pull this old software out and suddenly it can be available. That was very cool.

[A]

That's cool. There was similarly, people have been going on a tear about using Claude code to reverse engineer old DOS games or NES games and just, you know, you don't need the source code. It'll just go and figure it out and kind of decompile and recompile and like, that's kind of crazy. But I do look forward to the future, like even something like Doom where they give out the source Code. I'm surprised someone hasn't made like or something where.

[B]

That's a great idea.

[A]

You should.

[B]

We should not say that.

[A]

Put it on the board.

[B]

Yeah, put it on the board. Because that's like, never gonna do it. No, but actually, yeah. Yeah, like, because. Because honestly, that would be cool to be able to say, like.

[A]

And it should be multiplayer. Damn it.

[B]

All right, we'll be doing that. We'll be doing that AI foot website right after this, so let's move on.

[A]

So YouTuber Green Code, who is a great follow. I'm. I'm embarrassed that I wasn't following a lot sooner. They have all these like, really weedsy but fun and approachable videos on like really advanced AI topics. And what they did was find a tennis data set. We're talking like, players, their height, the, the, the scores that they get, the how many aces they get, how long their games typically go. Like, all this, this massive rich data set that they didn't make, but they, they took. And then they spent a lot of sleepless nights teaching a machine all of the data so that it could create an elo. Like a score, a ranking for the players, and then it could then try to predict the outcome of matches. That is so cool. And it. They, they. They have a YouTube video explaining it all. There's a great X post that someone did distilling everything that that Green code did, but basically it got about 85% of the predictions correct.

[B]

Wow.

[A]

Which, if you're a degenerate, your sports betting muscles just started twitching, right? And you're loading up a cow sheet window right now. But if you're like a, you know, a math nerd or a machine scientist, you're going like, that is pretty incredible. Like, that's a pretty crazy win rate. And Green Code actually replied to the thread where someone was dissecting his work and said there was actually data leakage and I should have predicted it way more like the should have done way better. So if you just look at the posts and watch the YouTube video, you'll see the way all the data goes in. It creates these scores for these players and then you use these scores to mash up against each other and see who's going to probably win the outcome of a match. And you realize the future is going to get very, very weird as we start siphoning all of that data.

[B]

Yeah, well, a friend of mine right now is doing this against Dota matches, which is interesting. Which is like something that, like, he's actually betting on. Like, there's A guy out there who I. Who I know well, and his, like, whole world, he's doing well. Like, this is like this weird world. Like, we. We talked about this a little while ago about this idea about, like, AI agents starting to get better at the edges that people have in prediction markets and things like that. But anyway, it's pretty crazy. All right, before we go, there is a great video that everybody should watch. One of my favorite AI videos to date, Kevin, maybe let's just play it for people here before we introduce what this is. I will take the ring to Mordor.

[A]

We're gonna help you with.

[B]

I have a ring I would like to sell you.

[A]

Baseball or football?

[B]

Neither. This kind.

[C]

Oh, this kind of.

[B]

I wish the ring had never come to me. So I'm here at the pawn shop today to try and sell it.

[A]

I really got to get this ring in my shop. This is not as uncommon as you might think.

[B]

So, you know what's going on here is it's Frodo Baggins, Elijah Wood, who's. Yeah, exactly. Who's going to Pawn Stars. And this is just those weird things where, like, somebody had an idea, they said, I'm fan of monsters, I'm a fan of Lord of the Rings. Yeah, what if Frodo had to pawn the ring? And, like, just very well cut together. This is a human edit. Like, obviously making choices from the show, making choices from AI video using different parts of different things. And this is from Reddit user. You are now Dumb D U M, which is a classic good Reddit name, but go check it out. We're not exactly sure who made this,

[A]

but we assume it's him. It's in his shorts. It's. It's David Unger from Dunger Zone who you've shouted out in the past as a. As a good follow. It's part of his shorts here.

[B]

He made a great video about dinosaurs and put the baby dinosaur from the dinosaurs ABC sitcom into Jurassic Park. I don't know if you saw that was very good. But anyway, he's great also about sharing his. His workflow. So go check him out. And this week, everybody go and try to make something. I'm telling you, if you get inside cloud code or replit or any of these other things, the power of what these can do right now is better than you ever imagined. If you are somebody who have mostly been using ChatGPT or Claude to answer your questions, try to make something small because this is the time. So you going to say anything? Just going to sit. All right, we'll see you all next week.

[A]

Bye. Bye. Whatever. Yeah, see.

[B]

By the way, we'll see you all next week. Maybe for an early episode. We'll talk more about it then.

[A]

Bye.

[B]

Bye.