I've been working on almost exactly your first idea at https://t.co/ut3TA6AicD.
The prototype uses private AI interviews to understand each party's preferences, then game theory to search for mutually acceptable agreements.
The technology is promising, but I've found the harder questions are around trust, participation, and finding the right use case. Still exploring those.
I made a video explaining how we're using these techniques in @FreenetOrg, including the actual routing algorithm and some real network data:
https://t.co/9Mdk8eG2jK
It's funny how often the right machine learning solution isn't a neural network.
For routing in Freenet, we're using isotonic regression, empirical Bayes, and a few other old statistical tricks.
The advantage is that we can use what we already know about the problem to constrain the model, helping it make useful predictions even with limited data.
I made a video explaining how we're using these techniques in @FreenetOrg, including the actual routing algorithm and some real network data:
https://t.co/9Mdk8eG2jK
@LukeParkerDev We're also using it to build a native desktop client for https://t.co/SND5vrfqt2.
Ordinarily it would be risky to bet on such a new lib but agents change the calculus.
Learning about Lean, a language that's also a proof assistant.
I remember Edinburgh Uni lecturers (ML was born there) describing a language where you could include proofs alongside code, but the compiler ignored them.
30 years later, Lean can verify them.
@awesomekling@jdegoes I'm currently experimenting with https://t.co/t0pgkjl1nG, which claims an order of magnitude speed-up of Rust builds through smart caching of intermediate steps. Not sure yet if there are downsides.