Signed in
We value your privacy
This site uses cookies to improve your browsing experience, analyze site traffic, and show personalized content.
Welcome to Pump.fun!
Pump lets anyone create coins, giving everyone equal access to buy and sell from the start. Prices can move quickly, so trade carefully.
By clicking this button, you agree to the Terms and Conditions, Privacy Policy, and certify that you are over 18 years old.
We value your privacy
This site uses cookies to improve your browsing experience, analyze site traffic, and show personalized content.
With the launch of the Matrix model back in March, I was fairly close to a Jev-like architecture. Here is where I was - Frozen pretrained backbone + small trained heads. A Judge head, a domain-router MoE, a complexity head, etc., ~278M trainable params. That is exactly the architecture the Jev teardown points to (a pretrained decoder + a decision readout). - Score in one forward pass, no text generation. The Matrix model did not generate, it read a score straight off a pooled representation in a single pass. That's the defining Jev principle ("don't generate a verdict, read it from the representation"), applied to reranking instead of action-choice. - Reading decisions from internal representations, not generated tokens. Jev teardown framed the whole proposition as "retain a pretrained LLM's knowledge but replace generated confidence claims with decision probabilities read directly from its internal representations." That's literally what Matrix's Judge head and reranker do. Here is what Jev taught - RLCD, Calibration as the training objective: training the probabilities to be calibrated via proper scoring rules. - A general "typed questions over shared state" surface. Jev generalized: encode any state once, ask many typed questions sharing the KV cache. Matrix stayed domain-specific. - Options that interact. Jev scores the whole option list at a final-position head, so adding a 5th option shifts the others. Matrix's reranker scores each query-doc pair independently. So I was on the right track: frozen backbone + scoring head, no generation, before Jev launched. 🤷♂️ What it taught is the objective (train for calibration) and the generalization (typed questions over shared state). Jevons Paradox Now I am taking all these learnings, for a local, on-device Jev-like model as part of the Hyperspace agentic system. It's called Paradox. Already built it's v1 to validate the possibilities: that a local, API-free, Jev-accurate decision engine runs on a laptop and decides real pages correctly in a hybrid setup; measured that its weak spot is multi-step decision quality which will be solved by Paradox v2. More abundance of this type of a model, from Jev's datacenters to running it on-device, the better it will be for society. h/t NielsRogge, harshagundal, and 4rcherhume (architecture deep-dive) and CompleteSkeptic 🫡 Fees to @UtaAoya via UsePaid