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?
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.
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?