Resonance-Bottleneck LLM tests wave-gating memory

Interesting idea, but the evidence is still far from proving a new architecture. Right now this reads more like a clever gated linear-attention variant than a validated replacement for attention or MoE.

• –The wave-interference framing is mostly a metaphor unless it beats simpler gating baselines on held-out perplexity and downstream tasks.
• –The reported loss drop on an 825MB corpus shows the training loop is stable, not that the model is meaningfully better or more sample-efficient.
• –EMA decay plus a synchronized normalizer are the most defensible parts here; those are standard stabilization ideas in a novel wrapper.
• –Gate activity and entropy staying non-degenerate is encouraging, but it only shows the gates are alive, not that they improve quality.
• –A single RTX 3060 can plausibly train a ~99M model, so the hardware claim is believable; the architecture claim needs ablations, scaling tests, and stronger benchmarks.