What we don't know

There are things the field genuinely doesn't know. They're worth naming clearly.

We don't know why large models are as capable as they are. Scaling laws predict performance improvements, but they don't explain why certain abilities emerge at certain scales. The theoretical foundation for why next-token prediction produces general reasoning is still incomplete.

We don't know how to fully inspect what's in the weights. A model can produce surprising, impressive outputs, and we often can't trace why. The weights are interpretable in aggregate — we can probe what the network has learned statistically — but the causal story of any specific output is hard to reconstruct.

We don't know where the ceiling is. Whether current architectures will scale to human-level performance across all tasks, or whether there are fundamental limits nearby, is genuinely contested. Researchers with serious credentials hold opposing views.

We don't know how to fully align powerful systems. RLHF helps. Constitutional AI methods help. But for systems that are significantly more capable than current models, we don't have alignment techniques that have been proven to work.

Uncertainty about these questions isn't a reason not to use these tools. It is a reason to hold your conclusions loosely, and to stay curious.