Your verifier is not ground truth
The evidence for AI self-improvement is concentrated where the grader is deterministic and wrong in patterns optimization can find.
Scaling is running the wrong race
The failures that matter are structural. More parameters won't fix them.
The Emerald City
Anthropic's Mythos announcement is real competence dressed in green glasses.
The collaboration is the attack surface
Multi-agent AI has a problem that constraints can't solve without destroying the point of having multiple agents.
The fever is not the disease
Anthropic's agentic misalignment paper shows exactly why anthropomorphizing LLM behavior leads to mitigations that make deployed systems less safe.
Retrieval IS the intelligence
The AI industry's most consequential architectural mistake is treating retrieval as a preprocessing step.
The very expensive router
The agentic AI movement is going in a circle and calling it progress.