And the structural ones

The rest are about how these systems are built and run.

Concentration of power. The previous two slides, in a phrase.

Environmental cost. Training and serving these models consumes serious electricity and water for cooling. Real, measurable, and worth weighing against what the model is being used for.

Dual use. Capability that helps a chemist helps anyone who wants to do chemistry, and the same is true of a system that finds vulnerabilities in software. This is the hardest one, because you cannot remove it without removing the useful version.

Compounding errors in agents. From the last chapter. An error in step three of a forty-step task becomes the foundation of steps four onward.

None of these requires a machine that wants anything. Every one of them is available with the systems that already exist, and most are already happening.

Whether the longer-term concerns about far more capable systems turn out to be well founded is genuinely contested among people who know the most. Holding that question open is more honest than settling it in either direction.