"The age of Socratic AI"?
« previous post | next post »
Or should we call it "Delphic AI"?
Alexy Khrabrov suggested both possibilities a few days ago, in "Reasonable AI — the Golden Age of AI Programming":
The emerging techniques are all around the way you construct the prompts and also chain them. Effectively, we’re plotting dialogues.
I call it the Age of Socratic AI, or Reasonable AI. We are engaging in conversations with AI that elicit meaning. We make the most basic assumption that it has the information we need and can provide it in the form we need, e.g. as an explanation or a how-to plan of action. We consider it an imperfect oracle that has to be assuaged, and asked questions in very specific ways to get the reply we need.
Alexy's analogy between the Socratic method and the questioning of Oracles surprised me a bit — though it shouldn't have, since Socrates himself (sort of) makes the connection. But even for those who think that LLMs are verging on AGI, we users are the ones practicing Socratic maieutics, bringing out truth with a clever series of prompts. And at present, really, we're more like Croesus, creating our oracular chatbot leaderboards:
Croesus, king of Lydia beginning in 560 BC, tested the oracles of the world to discover which gave the most accurate prophecies. He sent out emissaries to seven sites who were all to ask the oracles on the same day what the king was doing at that very moment. Croesus proclaimed the oracle at Delphi to be the most accurate, who correctly reported that the king was making a lamb-and-tortoise stew, and so he graced her with a magnitude of precious gifts. He then consulted Delphi before attacking Persia, and according to Herodotus was advised: "If you cross the river, a great empire will be destroyed". Believing the response favourable, Croesus attacked, but it was his own empire that ultimately was destroyed by the Persians.
astrange said,
May 1, 2023 @ 6:52 am
I've seen theories that the Delphic oracles were actually fairly good at answering questions; the reason being that everyone else also came to them with questions, so by figuratively reading between the lines they basically knew what was going on all the time.
ChatGPT similarly has access to way too much info, as I'm sure workers everywhere are typing your private info into it right now. Fortunately, LLMs aren't actually capable of learning from questions like this (they don't have "online learning") so we're not all powering it up by using it.
bks said,
May 1, 2023 @ 9:44 am
Like hope, ChatGPT is a valuable companion but a poor leader.
Lucas Christopoulos said,
May 1, 2023 @ 6:35 pm
Hallucinogenic methods for “prophecies” as in Delphi do perhaps connections with reality that are above the level of “pragmatic knowledge” as in ChatGPT.