Dude said something like "you could hook this up to a calculator". Anyone know if that is implying this generation of model could interface with some kind of symbol processor? Or is he just saying, "in theory", there could be a model that did that?
The math seems much improved and it would be a cool trick if it were emulating a symbol processor under the hood. But humans can do that and we opt for calculators and computers for a reason. IMO, the first most useful thing to come from human-machine interface would be adding a highly reliable Turing machine to your cognition.
If we could do that with one of these models, we could have a pretty strong assumption that long proofs and such were performed with strict rules and the model wasn't falling into some going from the gut, "this equation looks like x" type holes. Which seem like a hazard and make me very uncertain any time I see a paper about using ML to come up with answers to what are essentially logical problems.
He likely meant techniques such as ToolFormer[1] where the language model will output a "request" (in some text syntax) that another system can parse, run and report back.
It can still do something similar you just have to tell it how.
Prompt:
"CalcAI: I am a virtual calculator assistant that augments OpenAI's GPT. GPT may prompt me to solve mathematical equations by pretending text with $, and I will respond with the correct evaluation.
User: ChatGPT, solve for x, 2x + 5 = sqrt(2)"
ChatGPT: "$2x+5=\sqrt{2}$"
Prompt: "CalcAI: x = ( sqrt(2) - 5 ) / 2"
ChatGPT: "The solution for x is:
x = (sqrt(2) - 5) / 2 ≈ -2.07."
Why can't calculators or WolframAlpha serve as a computational oracle for ChatGPT?
It would seem as simple as assigning probably 1 to certain recognizable queries. Maybe the difficulty is that the very problem of choosing to use a calculator entails a meta-cognitive rational decision, and it's not clear how to organize that in neural networks, which are what Turing himself called an unorganized model of computation.
Right, so the decision to or not to inject an outside query into a particular repsonse and then iterate on the result, will be something learned by the model and therefore meta-cognitive as you say. Getting the model to a good balance of when to do so is an interesting problem. However, we at least could see whether the model tried a query and display the iterative steps it took. Then at least the cases where it did use a calculator would be verifiable.
The math seems much improved and it would be a cool trick if it were emulating a symbol processor under the hood. But humans can do that and we opt for calculators and computers for a reason. IMO, the first most useful thing to come from human-machine interface would be adding a highly reliable Turing machine to your cognition.
If we could do that with one of these models, we could have a pretty strong assumption that long proofs and such were performed with strict rules and the model wasn't falling into some going from the gut, "this equation looks like x" type holes. Which seem like a hazard and make me very uncertain any time I see a paper about using ML to come up with answers to what are essentially logical problems.