3 ms·
Google AI Chief Shane Legg Says There's a 50% Chance We'll Hit AGI in 5 Years
- ClassyJacket 3y agoIs ChatGPT not already that?
- anigbrowl 3y agoNo, although I am more generous than most in according it some degree of sapience. The problems with existing LLMs are twofold: They're not build with a sense of memory, but rather pop into existence as autistic savants with no real agency, and (allied to this) are bad at asking questions, unless explicitly instructed to that end. With a little effort you can get an LLM chat bot to introspect its text evaluation process, create memory structures to be filled with variables or decision criteria and much more, but it's very hard to get them to 'push back' on the user in any way such as asking for clarification or context absent in the original prompt. An interesting analogy for this failure: imagine a spy (LLM) is impersonating a military officer (quasi-autonomous agent). An underling suspects a substitution and challenges the officer for proof of identity: our inept spy bends over backwards trying to explain that they're really who they appear to be, instead of bossing the underling about.
- brucethemoose2 3y ago> instead of bossing the underling about. A roleplaying LLM will absolutely do this instead. One can of course step back and see that LLM spy text completion and human spy text completion are mechanically different. What people seem to get hung up on is the end results being essentially the same, but this doesn't bother me at all.
- anigbrowl 3y agoA roleplaying LLM will, but the point of the analogy is that (absent specific instructions on how to personify such a role) a general purpose LLM won't.
- hindsightbias 3y agoSo a 50% chance in 12 Friedman Units