[bonsaibilly]

> MY feeling so far is more that laymen are less impressed of them than experts - because for people not from the field, the fact that they produce so much bullshit seems to trump all other aspects.

The NYT literally just published an embarrassingly credulous account of how the Bing AI wanted to seduce the author away from his wife and commit various acts of violence, that fully took all of the interactions at face value.

The non-experts are imagining full-on sentience where the experts correctly recognize mere word association shenanigans. Your feelings, in short, are ass backwards.

[anonzzzies]

> The non-experts are imagining full-on sentience where the experts correctly recognize mere word association shenanigans. Your feelings, in short, are ass backwards.

The current model (which is an early version; computers are new on earth and exist only for a statistical error length of time on humanity scale, on earth scale, let's not talk universe scale) is already vastly superior than many humans I know in almost every way (outside manual dexterity and some other fringe stuff you can easily fix with external systems, like we humans do). We have no good definition for what sentience is either; maybe our brain is word association shenanigans; connect up 2 chatgpts and call one 'inner voice' and the other 'external voice'. It will start claiming sentience in no time flat; the same as you. Why are you right? It's a feeling yeah?

[bonsaibilly]

> We have no good definition for what sentience is either; maybe our brain is word association shenanigans;

This is just stupid.

We've pumped more English through GPT-3 than any existing English-speaking human has absorbed in their entire lifetime, and what we've ended up with is something that very very clearly has so utterly failed to generalize even the most basic level of understanding of basic concepts that if you ask it to count the number of letters in a word it will cheerfully pump out the wrong answer ("there are thirteen letters in the word 'twelve'") because its dataset correlates the two words with one another and it has learned precisely sweet fuck all about what it means to count, something a child's brain picks up with exposure to many orders of magnitude fewer language examples.

To imagine your brain is just an LLM is to mistake your reflection for another person in the room. Utterly daft. Get off the LLM hype train and start looking at these things objectively and critically. They're nowhere near what you're imagining.

[fiso64]

>if you ask it to count the number of letters in a word it will cheerfully pump out the wrong answer ("there are thirteen letters in the word 'twelve'")

Not to be too blunt, but you seem to be talking out of your ass. I'm tired of the overly dismissive comments here on hackernews by people who didn't bother to do the bare minimum of research. LLMs do not work with individual characters, they use tokens (i.e multiple characters, or sometimes entire words).

[BaseballPhysics]

Wow, you... completely misunderstood the purpose of that example.

Seriously, slow down and read it again, all the way to the end of that sentence.

[fiso64]

Uh, no.. I don't believe I did. Asking it to count the letters in a word is like asking a human to listen for a 30kHz tone without them knowing that humans can't hear it. I'd expect a lot of false positives.

[bonsaibilly]

Which reinforces the point I was making. Stop trying to win an argument online and think for 5 seconds about what it means for the claim "the human brain is just an LLM" if you're arguing that an LLM is naturally ill-suited to this task human children can do without issue.

[anonzzzies]

Yep. It’s not called tokens for nothing.

[williamcotton]

Here is an objective and critical assessment of large language models and how to empirically improve the results of questions involving calculations like the lengths of strings:

https://github.com/williamcotton/empirical-philosophy/blob/m...

[fullshark]

That author was playing dumb for the good of his story, in particular to help it go viral. This is what news media is in 2023.