falling down the AI infrastructure rabbithole

I started building LAIN earlier than I should have, by all measures.

The thing is, you can literally be building at any layer of abstraction right now. Actual, Inc. is building distributed inference and they’re working about as close to the hardware as you can get without a soldering iron. Polsia, a company that got a ridiculous amount of VC funding despite maybe being a fake company whose name is “AI Slop” spelled backwards, is building AI agents that run companies with zero human intervention. Nous Research is building inference infrastructure, models and agents.

I started working at the wrong layer of abstraction not because the foundations are missing or something. I just personally don’t have the right foundational knowledge yet, and I think there are tools missing from the ecosystem that I want to build.

Tursas is my focus right now. Tursas is an AI infrastructure application seeking to build security, observability and guardrails around agent behavior. I’m building Tursas so that I can use it to build agent harnesses for cyber security research. That cyber security research agent harness will, that’s right, be part of LAIN.

I think Tursas will be an incredibly useful company. I’m building it for myself, but the problems of building secure agent ecosystems, building them in a way that’s observable, building them in an ecosystem where identity and responsibility is an inherent part of the data model and not just a bolt on that’s implemented differently at every org, that’s fundamentally important. You should know who built what agent, what revision it’s on, who is responsible for which tools which can be used by which agents and which revisions and what access they have to critical systems and data. Kind of a no-brainer.

That said, if I market the hell out of it, get it in front of users and customers, and nobody uses it… I will. I’m dogfooding my own agent businesses with my own infrastructure by working down the stack.

I’ve got a long way between Tursas and doing low level inference work… but dammit I may end up there some day too.