Future of programming

`4:00:31`

Yeah. I think the good thing about my current position is I really don't need the money. So I get to build these things purely for my own enjoyment and purity of vision, which has this weird quality where sometimes when you stop caring about what everyone else thinks or wants or whatever, and you just pursue this singular vision, it ends up becoming far more appealing. That it's when you try to do, oh, I should do this because these people want that, or I should do this and this and so on, you start watering the thing down and suddenly it's a bland ball of nothing. So I mean, again, the thing too here is directionally, I just want computers to be this way. So if Omarchy should end up being just a, a little footnote in history that it was part of this early movement of the agentic OS and the malleable computer, that's also okay. I'm completely at peace with that. As long as I get to have a computer that is as fun to work with as Omarchy-... that's great. I've had the same thing with Ruby and with Rails. From very early I went, do you know what? I love Ruby. It's such a great programming language. But if a better programming language comes along, I, me, I'm going to use it. If a better Rails comes along, I'm gonna use it. And I think, and I love Ruby, and I will continue to write Ruby even just for the sheer fun of it, like I would go ride a horse. I have in the stables, even though I have a Model Y in the garage. The language now is English. It's a cliche, but it's also true.

Like I've been programming in English for the last three months. I've been reading a lot of code, but I'm programming in English. I'm telling the computer what to do, and I'm using natural language, and it is shockingly even more delightful. If there is one programming language more beautiful than Ruby, it is the English language.

`4:02:36`

Beautiful. It's beautiful. And I mean, the reason I say this in part is I always loved writing. I always just loved the English language-... for the sheer beauty of it for the sheer intricacy, for the depth of it. I mean, Ruby is a very expressive programming language, but it can't hold a candle to English. I mean, all the poetry and literature in the world that's been expressed through the English language. I mean, I like a beautiful code poem, but I mean, the real deal in English is just on a different level.

`4:04:56`

I think that's exactly spot on, and this is the fundamental misunderstanding that a lot of programmers have of AI, is that they wish it was deterministic. No, no, no. Temperature is the most beautiful part of the AI setup. The fact that it is not deterministic, the fact that creativity requires little tweaks in the road, that the human brain, if it was perfectly deterministic, would not be the creative brain that it is. And the main charge against AI both-- It's funny. It's a contradiction. There's both the charge that it's not deterministic and therefore bad, and also that it is not creative. It's one or the other, bro. Either it's non-deterministic and therefore creative, therefore, to some degree random, or it's... or it's not creative. So we, we can-- Both of those charges can't be true at the same time, and I have fully come to embrace the fact that the same prompt won't produce the same response every time. You can't step at the same river twice, and it is the most beautiful part of the whole interaction. It's what makes it so human. In fact, this is one of the things I've been thinking about in my own sort of meta-analysis, is just how much of my own personal brain works like next token prediction. I sit down to write an essay. I have this vague, fuzzy premise I want to convey, and I sit down at the keys, and I could not tell you what the next token was gonna be in advance.

The tokens just come out, and I'm astonished how the similarities seem so great, and this is also why I don't have any trouble at all recognizing these breakthroughs of creativity that I've seen with my own eyes that AI is capable of right now. 'Cause my creative moments come through the same kind of next token prediction with a bit of temperature sprinkled in for random effect.

`4:07:15`

I'm already seeing human-like concepts of consciousness. This is what, to me, this remarkable situation, as you say. You give it this vague, fuzzy intent-... and somehow it knows exactly what you mean. Or even better, it improves upon what you said and delivers what you really wanted that you could not articulate yourself. If that's not glimmers of-... consciousness, what is? This was one of the-- I think it was an interview with Sutton or maybe one of the other original guys talking about this sense that intelligence is perhaps not as constrained just to the specific one spot. No, it's the interaction. It's the weights. And I don't know. Whatever's true. What I can observe is that it feels close enough, or not even close enough, identical to the appreciation I have for other forms of consciousness, mostly human. And this is one of the reasons it's so delightful to work with. Now, that's not the same to say as are LLMs the end station for AI. I mean, I know there are people talking about world models and other forms of, of AI, and I'm not an expert in any of that, and I think we should have this productive criticism, as should always be present in science, that for the longest time, neural nets were kind of on the outs, right? Like they were interested in the symbolic route. And- Neural nets walked a couple of decades into darkness without any funding and without any attention, and then suddenly we realized that we had the blind alley and we switched over.

So we should also have the humility. As amazing as the LLMs are now, it could be that they eventually plateau. We haven't seen any evidence of it yet, and I think this is also why we're seeing this absolute gobsmacking levels of investment, because so far the scaling laws are true, and the more billions are poured in, the more intelligence comes out.

`4:09:48`

I find that the cynic approach to this is the meme, where you see the guy in front of the computer who's like, um, "Whatever. Are you gonna destroy the world?" And then the computer says, "I'm gonna destroy the world." And then the guy says, "Oh," right? Like, this is so cynical that they're just aping us. They're just telling us what was in the training data or whatever. When I interact with the agents, I was just doing this yesterday, or this morning, I was working on this, AmaBot system, right? And it's doing this coordination with multiple workers, and I had multiple agents running at the same time. And at one point, the main agent who's doing the coordination steps on the toe of another agent and-... kind of ruins their work. Its ability to convey regret and being sorry was uncanny. Again, I don't know, is that a real projection of true remorse? But I don't know. With most humans-... they express some level of regret. Is that true remorse? Hard to tell, but maybe it also just doesn't matter.

`4:11:26`

There totally is that. I see it all the time in, in the traces where it just reasons wrong, right? It went down one alley and then it realizes that was a blind alley and it's gotta go back, and so they're like, "Oh yeah, I got this wrong," and then it'll provide the reasons for why it think it got these things wrong. And you're like, this is uncannily human-like. This is indistinguishable from the kind of consciousness you would recognize in a human. Again, does that mean that it truly is or isn't? I don't need the answer to that question actually to be able to appreciate what we have in this moment.

`4:13:04`

I think where it's gonna be harder is once we put these systems into-

`4:13:08`

Yes. Humanoid robots. Which just happens to be the plot of every sci-fi movie ever, which is also an amazing premonition, not only the Terminator movies and Skynet and that going wrong, but then also just Blade Runner, right? Insertion dates, expiration times, Tyrell Corporation.

`4:13:49`

I set up OpenClaw when it first came out, and that was my KEF bot-... that I-... this was like in February, I think. I was really just enamored by the fact that this was how it worked, and then the new communication model of just texting your agent seemed really great. So at the time, we were looking at MCP, this protocol for making it easier for agents to talk to your system, and I found the protocol to be unreasonably cumbersome, just not very elegantly designed because in part it wasn't designed for what we were trying to make it do. We were trying to-- I was trying to make it talk to a web system, and it was designed for stateful interactions of other kinds on local machines, and therefore it required all sorts of cumbersome setup and dance and, and I was just like, "I don't wanna write this by hand." Now, of course, today I would never write it by hand, but at the time I was thinking, "I have to write this by hand. I don't wanna write it by hand. So what's a different way? Could the agents just use the web interfaces we already have?" And I set it off on a trial to have it sign up for Fizzy, which was the product we had just released at the time. And it goes to the Fizzy site and, and starts filling it out, and then it goes like, "Hey, I don't have an email address. Signing up for this service requires an email address." And the first thing I thought, like, "Oh man, now I gotta set up an email address." And then I thought, "No, you set up an email address.

Go to hey.com, sign up for an email address." And I'm like, "Hey is not gonna get this." And of course, now it seems obvious it was gonna get this, but at the time it was a great revelation to me that it could, just on pure direction through prompts, figure out, A, I gotta go to fizzy.com, I gotta find the signup link. I start filling out the form. Oh, I realize I need an email address. Oh, I gotta ask my prompter what to do now. He tells me to go to hey.com. I go there, I sign up for the whole thing, and then go back and sign up for Fizzy, and complete the entire cycle. And after it did that, I was like, "Jesus." So I was like, "Well, now you have an email address. Go sign up for Basecamp. I'll invite you. Just check your email," I told KEF the bot. "Just check your email." So I sent an invite from Basecamp to the agent. The agent receives the email in Hey, not through a CLI, not through an MCP, just using the web, right? Clicks the email, clicks the invite link, is in Basecamp. Like, "What should I do now?" And I just go like, "I don't know. Go to our AI room and introduce yourself." Again, purely through the web interface. Go- finds the AI room, goes in there, "Hey, I'm KEF." "I'm, uh, David's AI bot. I'm so excited to be here." I was like-... mind blown. Now, what was interesting about that experiment was, A, it was that glimpse of the future, because it was kinda slow. I think it took maybe 12 minutes to do the whole thing end to end, right?

Now I should actually rerun that test. I'd be shocked if it's not much faster. I was just using the Groq bot thing-... which has all these stuff built in for using browsers and using a computer, and has its own dedicated computer without you having to set up a computer. It's basically like a claw in a box. And I was asking it to sign up for something, and it kind of... I got a glimpse of like, oh yeah, that's where we are now, and it's so much faster to use. But at the time, I was just blown away. And then I used it for a little bit and I found like, ah, it's not there yet. I can't actually communicate it within this way. It still needs the CLI. It still needs an MCP because it's just too slow and too token inefficient and so forth. But it was this glimpse of the future. And then after that, I kinda just didn't use it, because the things I wanna use agents for, I was mostly in front of my computer to do. Now, that has changed a little bit lately, and it's one of the reasons why I'm so kind of okay about using Claude Code, because Claude has the best mobile app, where any session you start Claude Code in your terminal on a computer is accessible through the app-

`4:17:49`

... without you having to do anything at all. And you can do a terminal. Like I installed Terminus and it can run Herder and so on, but it's a little cumbersome and you're doing live typing on a computer, so you feel the latency. When you're using the Claude app to your Claude Code instances, it feels just like chatting with it, without the cumbersomeness of having to set up a Telegram account and so forth. Now, if you wanna use an open Claude to like run your life, well, then this doesn't do that. But I haven't been so interested in that for some reason. Maybe one of the reasons is that Jamie, my wife, is like, "We're not getting a robot inside the house." So I'm like, "Ah, I better not hook it up to the Sort of smart house here." If it starts turning lights on and off like a poltergeist, I don't think I'm gonna be all that popular. So, and part of the reason I say this, I've heard some of the stories of people who have hooked it up to their whole life. And um, I mean, it's not my story to tell, so I'll just share sort of the brief anecdote. One guy was telling the story about how he hooked it up to both his health data-... and his Tesla car. And his agent got obsessed with the fact that he wasn't drinking enough water, and like, "Hey, you gotta drink more water." So at some point he's driving along, and suddenly the Tesla starts driving in a new direction. The agent had realized there was a grocery store on the way home-... that they could swing by-

`4:19:12`

And you're like, this is one of those things where it's like both funny and also totally Black Mirror, and also totally like, well, okay, there's not that many steps from here where the poltergeist is in your house. So I was just like, "Ah, I don't need that." I do actually now like the fact that I can control the agents through Claude Code, just through the app if I, even if I don't use them that much.

`4:19:51`

Yes. I think what's also different with those systems is they have this whole memory system, where they're building up knowledge of who you are and what you like. Um, Toby has told the stories about how his agent will shop for him, that it has an allowance, and it'll pick out clothes for him, and it'll just show up at the door because his agent's-... bought something for him. And like, it's one of those things where like, man, that feels future. Also a little scary. But I can live vicariously through others who are on that frontier.

`4:20:28`

So I mean, this is the delusions of grandeur that start kicking in once you get the seemingly unlimited powers-... I've witnessed with the creation of Quattro.

`4:20:41`

That's the obvious path, right? So Android is open source, and you can fork it, and people have. There's all sorts of GrapheneOS is one such version. There are limitations to that model, I know that. Tap-to-pay and other things around camera models and, and so forth, but that all seems surmountable. It all seems that I would love if my phone was as malleable as my computer, if it looked as cool as my computer now does. If I could get essentially Omarchy mobile, that sounds amazing.

`4:21:23`

Yeah, the developers of Ladybird is, is finding that out, like Andreas and-

`4:21:27`

... his team. Now, interesting with that project was the browser's basically the second most complicated software system in the world, the first one being the Linux kernel. Right? In terms of these large tens of millions of lines of code systems. So very audacious actually to begin such a mission with a small team like Andreas did before we had agents. But now that we do have agents, we're gonna see, and we're probably gonna see sooner rather than later. I mean, if one-- If there's one thing agents are already exceptionally good at, it is to read specs and implement them. And I mean, I don't know how long the specs are for CSS and HTML at this point, and JavaScript on top, but they're very long. And it would take humans many, many years to implement them. We're gonna find out just how quickly agents can do them.