AI video generation and filmmaking

`2:57:34`

What, what more could there be? And I mean, I can quibble at the margins, but on the big picture, I've seen it able to do these long run tasks and coordinate these multiple aspects of it that we're so far beyond that early moment where I'm telling it, "Go do this, package that." And now it can carry out a whole task from these vague instructions. That feels very AGI-like, and that's what's so inspiring about the moment because you catch these glimmers. And then you go, "What if everything was like that?"

`2:58:20`

So, so this is... I'm in that meta phase right now. So I'm building this Amabot system that's managing the development of Omarchy, and I'm using AI to build the bot. And I'm seeing that recurrent loop where I started it on, like, "This is what I want. I want you to be able to have these isolated VM workers and so forth," and then it can do the self iteration. Then it starts running it. It starts noticing its failures. It starts optimizing it. And a lot of that system, Amabot in particular, AI has driven the majority of the design decisions. Like, I had some vague ideas of where I wanted. I wanted to use this brains and hands pattern where the model is running from where the code is executing-... which is actually a really interesting pattern. Again, Toby alerted me to that, where you have a coordinator that runs the model-... but it's manipulating a safe VM where any untrusted code you're pulling down from pull requests or issues or whatever can't contaminate the model. And even in that loop, I keep seeing it catch itself, spotting vulnerabilities, where it's like, "Oh, I actually just took some feedback from a test run. If a clever attacker had embedded a malicious payload in the response of the test run, that could have polluted something. I took it for good thing. I better start treating that as as outside data." And you just go like, it is just... I get why San Francisco is so obsessed with fast takeoff-... because clearly they've seen this 100 times-... more.

They've seen it with all the models that are in-... training now, and we keep getting these little glimpses, like the security breach with Hugging Face, where OpenAI was training this model, and the model starts inventing ways, essentially sending smoke signals to itself through embedding messages in a package manager, where you just go like, first, that's damn clever. Second, okay, that's also a little scary. And third how amazing would this be if we could harness this level of intelligence-... and ingenuity towards productive ends? So this is the it's so over, we're so back pendulum that keeps swinging back and forth-... that AI has just delivered that swing back and forth like nothing else. Like barely a week goes by where it has to swing one way, and then the next week it swings back the other way. Exhilarating.

`3:01:22`

That's exactly what it is. That's exactly why it's so exhilarating and why, as we talked about before we started, it can sometimes be difficult to come out of hyperdrive. That's also one of Toby's terms here. That when you're working directly just one-on-one with agents and you're not looping in other humans, you go so fast, and when the agents are really understanding your intent, or even better, they revise and improve upon your intent to deliver something that was greater than what you asked for And you're doing this with 16 parallel threads. It can be quite difficult to come out of hyperdrive and have to deal with squishy humans. They just, they don't think as fast. They don't move as fast. And I mean, I'm trying mostly to make light of it, but you can also see a dark version of that where someone does end up in an AI psychosis where they're not interested in other humans at all anymore. I mean, certainly don't have that at all, but I do have this sensation at times where I'm really grateful for just the rush of just me and the agents. Like, can I just not talk to anyone else but agents-... for four hours? Like, that's an incredible run.