Anthropic's Natural Language Autoencoders translate a model's internal activations into plain English, making the black box significantly more inspectable.

Most of an AI agent's capability lives outside the model, across memory, skills, protocols and mediators. Designing agents is about what you externalise.

TTT-E2E treats long context as a learning problem, not a storage one. The model updates its own weights on the fly to keep inference cost flat as context grows.

The MIA (Memory Intelligence Agent) framework argues AI agent progress comes from better architecture and memory loops, not bigger context windows.