How close are we to NAIs?

[Tom Mazanec]

The book says they got NAIs in 2015.
How close is the Real World to NAIs in 2023?

[Prince Charon]

Originally Posted by Tom Mazanec

The book says they got NAIs in 2015.
How close is the Real World to NAIs in 2023?

I think that depends on how limited or generous the definitions we use are. Does ChatGPT qualify? (Of course, some of the limits on ChatGPT are imposed by the developers, and there's a difference between legal/social limits and technological ones.)

[Fred Brackin]

Originally Posted by Prince Charon

I think that depends on how limited or generous the definitions we use are. Does ChatGPT qualify? (Of course, some of the limits on ChatGPT are imposed by the developers, and there's a difference between legal/social limits and technological ones.)

NAIs are supposed to be capable of sentient behavior and learning. I don't think ChatGPT does either. It recognizes certain patterns and emulates them but doesn't understand at all whether it's done something the right way or the wrong way.

NAI-5s were also supposed to be available in 2022 and cost $500. Even the most usable interactive programs (probably things like Siri and Alexa) no doubt cost millions and run on server farms.

I'd say we're orders of magnitude away from anything like the AI in Transhuman Space.

[Prince Charon]

Originally Posted by Fred Brackin

NAIs are supposed to be capable of sentient behavior and learning. I don't think ChatGPT does either. It recognizes certain patterns and emulates them but doesn't understand at all whether it's done something the right way or the wrong way.

The 'P' in ChatGPT means 'pre-trained,' so it is capable of learning if you don't impose limits on its ability to do that (which is what I was talking about). In short, we don't and can't know how close we are. Certainly, I think we're behind where the TSverse was in 2015, much less 2023, but that's explicitly a setting where they spent more on AI and related developments, and didn't impose the same limits we do at this time (though they likely imposed different ones).

[Anaraxes]

Originally Posted by Fred Brackin

ChatGPT... doesn't understand at all whether it's done something the right way or the wrong way.

This is a key point for anything that's supposed to be a self-motivated intelligence. ChatGPT exhibits zero understanding of the content of the text it generates. It's great at generating that text and parsing user questions, but there's nothing higher level than that. It can't tell if it just stole the right or wrong answer from a web site, whether answer #2 matches answer #3 that it just gave, and so on. There's no stateful model of what it's talking about, expectations, or internal motivations. So, nice as it is as a natural language processor, it doesn't exhibit (to me) any "intelligence" behind the words.

Kinda gives me the same feeling as a mentalist / con man doing cold reading. The output looks good, but there's not really any "there" there. It's more about selling that output while eliciting the actual thoughts from the mark and playing the odds while rephrasing that info as vaguely as you can get away with into the next response, rather than actually generating the answers.

I don't have THS, but the definition for NAI in the GURPS Wiki is "Programs that follow basic paths, ranging from smart tools to videogame NPCs". So by that definition, we've had NAIs for a while. And whatever ChatGPT does, it isn't one of those and in fact is less N-AI then, say, those videogame characters, which at least have some minimal self motivation (even if it's just "patrol this path and alert to anything you see"). I don't think ChatGPT does anything at all if you don't keep prodding it.

[Phil Masters]

Originally Posted by Fred Brackin

NAI-5s were also supposed to be available in 2022 and cost $500. Even the most usable interactive programs (probably things like Siri and Alexa) no doubt cost millions and run on server farms.

That first availability date doesn't have to imply that they cost the same when they first appeared as they do in 2100. Experimental software with a huge development cost in 2015 could have become the oft-the-shelf, consumer-market-priced standard in 2100. And a NAI-5 is by definition Complexity 5, which means (per High-Tech) it needs a fast microframe, mainframe, or slow macroframe at TL8.

Anyhow, that first availability date is mostly a measure of how old the software could be, and software in TS is clearly often enhanced and upgraded over time, with lots of legacy code still buried in its depths. So what you could find is that the slightly eccentric NAI-5 that runs your building's internal systems actually includes chunks of Alexa, behavioural patterns dating back 80 years, and even fragments of memory from great-grandpa's time.

(None of which means that the TS timeline actually fits recent history. But still, one can kludge.)

[malloyd]

Originally Posted by Prince Charon

I think that depends on how limited or generous the definitions we use are.

Always a problem in AI debates. We can't [define] intelligence clearly, and it's hard to duplicate something you can't define, or for that matter tell if you have succeeded.

Does ChatGPT qualify?

No. It strikes me as Google with a paraphraser on top of it. The paraphraser does seem like an achievement on the path to natural language processing, which is often considered a key component of AI (I think that's debatable) but probably doesn't involve anything intelligence like. I actually think it's further from an AI than, say, Watson, that at least seems to have had some sort of effort at having the AI deal with semantic meaning.

[Fred Brackin]

Originally Posted by malloyd

The paraphraser does seem like an achievement on the path to natural language processing, which is often considered a key component of AI (I think that's debatable) b.

That's entirely possible though I am struggling a little with uses for a "true" AI that couldn't handle natural language. You could maybe do some sort of self-directed math or physics researcher but could you understand its' output?

[malloyd]

Originally Posted by Fred Brackin

That's entirely possible though I am struggling a little with uses for a "true" AI that couldn't handle natural language. You could maybe do some sort of self-directed math or physics researcher but could you understand its' output?

Perhaps I can't, but the High Priests of the Computer....

Seriously though, needing to format your questions in some sort of formal language and needing an expert to interpret the results isn't particularly problematic. Indeed some fictional AIs have required this (Merlin for example in Piper's Cosmic Computer) perhaps on the model of the way early computers required specialist card punchers. Or, possibly the experts required to interpret oracles. Much the same thing right?

[WingedKagouti]

Originally Posted by Prince Charon

Does ChatGPT qualify?

That's a hard NO.

ChatGPT and other algorithms are basically predictive language models that look at a large set of data, trying to find the most apropriate output given the current input (which includes the previous output in the session).

Thinking of them as Expert Systems is far more accurate.

Your smartphone is much closer to being an AI than ChatGPT.