DeepBlueDynamics Autoresearch Adds Weber Optimizer, SDR Entropy

This is a clever research fork, but the real story is probably the stack around the optimizer, not the optimizer alone. The Weber bracket is the kind of idea that could be genuinely interesting, but it still needs clean ablations before anyone should treat it as more than an experimental twist.

• –The Weber update is novel enough to be worth testing, especially since it modifies step size based on both momentum and acceleration.
• –The SDR entropy seeding is fun and defensible engineering, but it is unlikely to move model quality unless the prior seed was unusually unlucky.
• –The reported gains likely bundle together many changes at once: depth, batch size, RoPE base, init scale, and weight decay.
• –The agent harness is arguably the most practically useful part because it turns the repo into a repeatable experiment loop, not just a one-off training script.
• –For anyone evaluating the optimizer itself, the key question is whether Weber still helps after controlling for the rest of the hyperparameter sweep.