The Likeliest Next Thought
I don't think machines will get to the kind of thinking that arrives before anyone has room for it. What they do is find the likeliest next thought1. That is most of what thinking is and it is worth a great deal, but it is anchored to what people already believe, because that is what it was made from. And it has no reason to hold a position once you push back on it. Germ theory came from a man who died in an asylum insisting doctors wash their hands. Group theory came from a boy who spent the night before a duel writing out what he knew, and died in it at twenty. Neither was the likely next thought. Neither man could let go of it. Thinking isn't only mental, it's also emotional. Some of it is bred by envy or greed, some by rage or shame, some by love for the thing itself.
I don't think AI needs any of that to be useful, or to create real value. But it's worth noticing which way the borrowing runs. The leading lab is called Anthropic. The root is anthropos, human. We keep asking how much humanity to project onto these systems so they'll be more useful to us. The better question is how much of ours they've already eaten: millions of people reasoning in public, digested, returned as the reason to stop.
What's eaten doesn't come back. And the places where people reasoned like that in public are closing.
- A language model is trained to predict what comes next. Given the text so far, it produces a probability distribution over possible continuations and samples from it. "Likeliest next thought" is that mechanism described at the level of an idea rather than a token. ↩
Subscribe