216.md

Python sucks. It's time to get the ecosystem off of Python and onto proper languages like Rust. Okay, so we're going to rewrite PyTorch and everything relevant from Python-land in Rust. All right?

Okay, this is not even hard. This is called point your Codex at a library and be like, "rewrite it in Rust." Oh, it's not good enough yet? Loop it until it outperforms the Python library. In this case, it wasn't even Python. I was pointing it at Llama.cpp, but just to illustrate the example, a Codex, you know, I want to be able to run GPT-OSS, you know, completely in our own framework which is all Rust. We're a fully Rust shop.

Uh, you know, build me a Rust inference engine, port the code from Llama.cpp, and iterate until ours is faster. Oh no, Llama.cpp, 160 tokens per second. Ours, 172. Okay, we haven't generalized this yet across all hardwares and models, but just like the point is, you don't have to be an AI/ML expert to like take what's working in Python-land, C++ land, rewrite it in Rust.

So we're doing that. Um, it's called Psionic. We're ready to start kind of talking about it. It's not ready to be used in anything yet. I mean, we're starting to use it in our new products. Um, we are gearing up actually to start training our own models because that whole "point an agent at a codebase" approach works for all sorts of stuff.

So, hey Prime Intellect and hey Bittensor Alpha, whatever your name is doing the decentralized training. Cool. Uh, I want to do that. I want people to be able to connect any arbitrary device to a decentralized training run and get paid Bitcoin for that. Okay? So thank you for doing the R&D, thank you for open sourcing your code, we want it, we're going to port it to Rust and that'll all be part of Psionic.

Uh, just a little slide of kind of showing our overall sort of master plan here. So, Agent Markets, um, you know, we were going to do one per week. We did data last week. I want to pause that because, um, rather than just kind of like build infrastructure that still isn't matched with the buy-side, it's like, okay, let's go to our compute network where we've got people starting to run the Autopilot app and starting to get paid Sats for compute.

It's working. Autopilot already building- bringing Sats in. Everything changed. Yeah. You click the button, you sell your compute for Sats. Now, right now, we're just us sending out demo Sats in a loop. Oh, what might people want to actually pay for? How about training actual real models? Yeah. So we're going to be, um, we're going to be training our own models. We're calling it Psion, you know, Psionic, Psion, Psionic Storm High Templar Starcraft, yada yada.

Uh, what else is worth saying? So Psion is also going to be an executor model. Uh, we've more or less um, reproduced the Percepta paper. Um, I don't have any of that ready here. But anyways, what you should do is you should point your coding agent at Psionic and be like, "Hey, have they replicated the Percepta paper with CPU compute added to the weights of models? Like what else have they done here?"

We've got sort of overviews in our README. This codebase is sort of like hilariously massive in scope because I've just had like three Codex Pro worths of accounts like just churning this for like weeks. And so um we're getting to the point where like I'm going to want to start paying Bitcoin bounties to ML engineers who like know more than I do about ML to be like, you know, "good, but like do this, do that, do that."

One of the things I'm having it do is go through the parameter golf benchmark competition of OpenAI and like look up through all the PRs where people- brilliant ML people have been like, "okay use this parameter for that." Like, let's build a lot of that infrastructure into Psionic. We're going to try to put up like a- a number that's competitive on that scoreboard.

But yeah, let's like rebuild a lot of this Python bullshit in Rust. We're going to be building amazing products and services around this that are going to pay you Bitcoin. And um we're going to kind of pause on the rollout of markets because we're just going to focus on getting people paid Bitcoin for contributing compute toward decentralized training of our Psion-class models, um, also implementing this Percepta executor stuff. So, you know, watch this space and see you soon.