Loading slide
Honesty means admitting the edges of what is understood. There are real, important things the field genuinely does not know, and naming them plainly is part of seeing these systems clearly. Four stand out, and you have met the ground under each one.
We do not fully know why scale produces these abilities. We saw that making a model bigger makes it better, and that brand-new skills switch on at sizes nobody could predict, like wetness appearing once enough water gathers. But why simply predicting the next word, done at vast scale, gives rise to reasoning is still not understood at the bottom. The recipe works; the deep reason it works is open.
We do not fully know how to read what is inside the weights. We saw the knowledge is real but smeared across billions of numbers with no index. We can probe a network and learn things about it in aggregate, but tracing exactly why it produced one specific answer remains mostly out of reach. The black box is still, largely, a black box.
We do not know where the ceiling is. Whether the dominant designs of this era keep climbing all the way to human-level skill across the board, or hit some wall close by, is genuinely contested. Serious, well-informed researchers hold flatly opposing views, and that disagreement is honest, not a failure to look it up.
And we do not know how to fully align systems much more capable than the ones we have. We met alignment as the gap between what you said and what you meant. Techniques like preference training help narrow that gap, but nobody has a method proven to hold for systems far more powerful than the ones already deployed.
None of this is a reason to back away from these tools. It is a reason to hold your conclusions loosely, to keep checking, and to stay curious. The honest position is not certainty in either direction. It is clear eyes.