Latent Dynamics
Articles about Latent Dynamics
The RAW Alignment Defect: Continuous In-Weight Learning via Fast Weight Attention
How a one-step shift in key-value pairing changes what a fixed-size recurrent state can store, and which claims the evidence supports.
The Non-Compositional Blind Spot: Verifying Emergent Multi-Agent Capabilities
Why every skill in an agent registry can pass its audit while the installed set stays unsafe, and how capability hypergraphs and Datalog closure check it in microseconds.
Verification architecture for autonomous research agents
Published audits show autonomous AI agents can finish a research run with zero errors while fabricating the results inside it. Here is what a system that catches this would actually need.
Breaking the Autophagy Barrier
A fact-checked tour of semantic entropy filters, prover-verifier games, and the engineering safeguards that keep recursive AI training loops from collapsing.