4 ms·

I expect we can squeeze a lot more exponential out of LLMs because they’ve basically shown that human “consciousness,” insofar as it’s composed of knowledge and

I expect we can squeeze a lot more exponential out of LLMs because they’ve basically shown that human “consciousness,” insofar as it’s composed of knowledge and rules for synthesizing that knowledge, can be represented mathematically in a very high dimensional space. Why does this “just work?” Nobody really knows, but it clearly does. However, I also expect this squeeze will come at an increasingly expensive price — not just because of inefficient token usage, but because of fundamental limitations of LLMs as a model. LLMs are letting us brute force our way through a lot of reasoning, but it’s hard to believe that such a generic model of intelligence will take us to the next frontier. We’ll need some fundamentally new approaches at some point. Maybe those will make achieving the exponential more efficient or maybe they’ll unlock even higher degrees of possibility. Who knows?

andai 2mo ago
The transformer is Turing complete. It might be a tarpit though? I don't know. I think a nice example is using them for arithmetic. It's a specialized deterministic process, so it's extremely wasteful to do it that way. But they're good at finding solutions to things we don't know how to specialize yet. So, to use metaphor, maybe the transformer-based models are like the FPGA, and then when we figure out the patterns in that system — all the different kinds of specialized reasoning — we can extract it into an ASIC?
- variadix 2mo ago
  I think this is one approach to AI safety and interpretability that could work, but would require labs to slow down to figure out how to extract circuits/algorithms out of trained LLMs rather than deploying the opaque artifact.
visarga 2mo ago
> Why does this “just work?” Nobody really knows, but it clearly does. We know language has to be learnable by every human, so it needs to be really independent of any specific brain development particularities. If it was not accessible to babies there would be no more language next generation.

andai 2mo ago

The transformer is Turing complete. It might be a tarpit though? I don't know. I think a nice example is using them for arithmetic. It's a specialized deterministic process, so it's extremely wasteful to do it that way. But they're good at finding solutions to things we don't know how to specialize yet. So, to use metaphor, maybe the transformer-based models are like the FPGA, and then when we figure out the patterns in that system — all the different kinds of specialized reasoning — we can extract it into an ASIC?

variadix 2mo ago
I think this is one approach to AI safety and interpretability that could work, but would require labs to slow down to figure out how to extract circuits/algorithms out of trained LLMs rather than deploying the opaque artifact.

visarga 2mo ago

> Why does this “just work?” Nobody really knows, but it clearly does. We know language has to be learnable by every human, so it needs to be really independent of any specific brain development particularities. If it was not accessible to babies there would be no more language next generation.