Jim Brander has had a career in manufacturing, electronics, a defence project and Knowledge Representation. His idea is that knowledge should be undirected.
@ActiveStructure I can see the point about trying to stop lawyers lying, but the Boeing MCAS problem wasn't about lawyers, it was about engineers behaving despicably - is Orion going to change that?
"Language and thought, Fedorenko says, “really are separate in an adult human brain.”" - from NYT. Does this mean we are wasting our time with a dictionary-based AGI, or will the machine be different?
The errors are understandable – the LLM likes statistics, but isn’t big on logical structure - it is the making up of false justifications and lying about it that is hard to take. Why do that? This doesn’t look like the kind of thing you want to hang your hat on.
An Insect bites an LLM
We didn’t believe an LLM could handle the punctuation map used in complex text, so, rather than feed in a 400 page piece of legislation to test it, we tried a half-page dichotomous key for insects. It did not go well – see https://t.co/bHvCLUIMlN
A blog post about LLMs and language problems - we will be following up with the problem of punctuation mapping on complex text - very relevant to Defence. https://t.co/stAvm8dTyR
@karpathy "Ï like to train large deep neural nets" - you will have to learn how useless they are. Some words have half a dozen POS, some have 70 meanings. Think before starting down another dead end. https://t.co/PRGeT3Sk9C
@karpathy Engineers used to care about accuracy. Your "prompt engineer" doesn't care, as long as something is cobbled together - who cares if it is a crazy quilt - it "deepens the mystery".
@elonmusk@karpathy Love the nonsense being spouted. We build something that doesn't understand a single word of English, let alone all the nuances of clumping, then pat ourselves on the back for handling natural language! Well done.
You are using images like Everest and icons like Helen Keller and Sun Tzu in your blogs. Aren’t they a little overblown?
We are offering the gift of meaning when everyone else is offering stasis or a crazy quilt, so we should probably work Prometheus in there as well.
We need to build machines that break the FourPiecesLimit and can handle problems beyond our ability. Witness all our billion dollar stuff-ups. Obviously, their actions will be unpredictable, because we didn’t know how to solve the problem. We have to give them critical thinking.
To quote Dirty Harry - "A man has to know his limitations". The Four Pieces Limit is a severe limit.We also need to have a much better understanding of our Unconscious Mind. Yes, we need to educate our young in these things, so we are not afraid of using cognitive machines
@pettter_e@GaryMarcus Language can be for amusement. Language can also be used for collaboration. It becomes important that everyone understands the same things. Most processing is done unconsciously, so people don’t know how they arrived at the meaning of a piece of text. This needs to be fixed.