Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Maximum™]

Weak AI and Strong AI were already established terms in scifi though, and an LLM doesn't really meet either definition remotely.

That's why I used that wonderful description earlier: LLMs are blurry .jpegs of the internet.

If it seems somewhat familiar it is just because there is so much of ourselves in the internet now, so smushing it together and smooshing it back out in an on demand interface is a really easy way to trick people into imputing motive and awareness and indeed a mind where there is none.

You can do neat things when you throw data at a model and ask it to squish out an interesting solution for a relevant problem.

Stuff like minimal material optimization and out-of-the-ordinary approaches to problems are cool and ALMOST qualify as intelligent.

Chatbots confidently making shit up based on what they expect the next word to be in a response is a cool parlor trick, and a great way to pollute the internet with even more confidently stated bullshit than it was already full of.

[MaxTheFox]

Established terms can change. I just don't see the point in taking a stand for the usage of a word (as long as it's not about real people's identity) because I value people understanding wtf I am talking about more. Again, it wasn't really spread by corpos, not in the same way (or as nefariously) you put it at least. I remember the early days of this stuff and people were calling it AI even if it wasn't even conversational. AI doesn't need a mind, or anything more than a veneer of a mind.

For my worldbuild I invented a term: pseudosapient AI. It's essentially what we have, except more advanced, with emphasis on pseudo. Good for menial tasks involving information, or interpreting orders. Not good for anything involving high accuracy, on-the-fly learning, or human-level flexibility.

[Starver]

There was a time when we'd call Eliza an AI. Far from it, of course.

Tell me about the time when we'd call Eliza an AI.

Well, although it was programmed by humans, it did occasionalyl come up with rare flashes of brilliant insight.

I'm glad you are well, although it was programmed by humans, it did occasionalyl come up with rare flashes of brilliant insight.

It would often degrade badly if it got the wrong end of the linguistic stick. And probably the self-building version is just as dumb, just less predictably so.

Tell me more...

...
edit to correct one typo, and *duplicate* another...

[Strongpoint]

LLMs are funny because they prove that the ability to speak in coherent sentences does not equal intelligence. In Musk's case, we can see that the ability to become a multibillionaire also doesn't equal intelligence

[MaxTheFox]

LLMs are funny because they prove that the ability to speak in coherent sentences does not equal intelligence. In Musk's case, we can see that the ability to become a multibillionaire also doesn't equal intelligence

The Turing Test has been falsified as a measure of sapience, tbh.

[King Zultan]

I've forgotten what the Turing test is.


Also can really stupid people like Elon fail it?

[Starver]

As the classic version is that an evaluator remotely converses with a person and a computer[1], blindly, and decides which is which, then obviously in any case where the computer 'passes', it is because the human has (relatively) failed.

And, as sophistication of chatbot increases, it might be because the human is actually too precise (because the computer has been developed/encouraged to make human-looking typos/thinkos), at least until that becomes expected. Then the sloppy 'human' typist might be misidentified as "trying too hard" to look like "a normal human"...

So, really, the full modern test should be done as a cohort thing (multiple humans 'vs' multiple chatbot 'AI' instances, with multiple evaluators[2] scoring ) to tease out a probability function that cannot be null-hypothesised by significant sigmas. Or something.

(There have been people who proudly claim to have been misidentified as the AI in past Turing Tests. But then Elon, for all his latest warnings about "humanoid robots", probably is just as likely to be one as anything we're going to see in the near future. I see through his game... Methinks the android doth protest too much! Obviously not perfectly programmed (cannot even generate a plausible name, for himself or his 'children'), but mechanically only slightly off into the Uncanny Valley...)


[1] Independently, although some interpretations are that the (test-)human and computer converse and the judge reviews their conversation...

[2] Throw in some generative-adversarial 'judges', just for fun!

[Strongpoint]

For me, a test for basic intelligence (not sapience) will be something like

1) I explain the rules of a somewhat complex game of which AI has no prior knowledge.
2) It plays it, forming its own strategy instead of making random moves

[Maximum™]

[Rolan7]

I still can't figure out how people think AI is going to destroy things. Unless we connect AI to actual machines, or we let AI re-wire our brains like Dollhouse, or we just blindly follow instructions from AI without checking it first, there's just no physical mechanism by which AI can effect any change.

So if AI does "destroy the world" - it's the fault of whoever connects it to actuators.

I think the Dollhouse scenario is probably the most entertaining to be honest.

or we just blindly follow instructions from AI without checking it first,

Unironically I see this as the main issue. Startups do it constantly but even tech giants keep using their Algorithms in ways which optimize for profit without proper regard for consequences. Then keep trying to patch the consequences (when forced to by public outcry and even congressional hearings)

The threat is that a reliable AI honestly says "Do this and you'll make a profit", we do it without knowing its internal logic, and people systematically get hurt in obscure ways. That's not a hypothetical scenario anymore, it's a daily occurrence.

[Starver]

It's the hidden nature of The Algorithm™.

When a person had to (say) add keywords to look for in submitted text to link someone "probably discussing getting a new kitchen" to ads for some relevent company (who was paying for placement), it was perhaps a shotgun approach if the advertiser was willing to pay for each 'shot' in the dark that they thought might help them ("sink", maybe - unless it was boat-related; "dishes" - satellites, or just eating out; "surface" - nautical/mathematical?) and tit woud be a complexity of regexping something like "I don't really have enough room by my sink to dry all my dishes, I end up spreading them over the nearby work surfaces" ( <= contrived example! ) to know that some kitchen redesign outlet might like to convey an ad to that person.

Left alone to a learning algorithm that enumerates whether a person who exhibits certain interactions is more likely to be susceptible to certain messages than others (or even add to that some additional push-choices) creates a chimeric situation where it might establish that speaking in a general feminine way is part of the selection criteria (because Real Life Is Biased), and yet nobody will really know this without exporting (and comprehending) the complex net of logical associations that had been formed over the time it was ennumerating what it thought were necessary parts of the process. (Which, by the way, could have as easily been as random as a Skinner Box experiment generating synchronicity and 'ritual-behaviour' in test subjects.)


On the odd occasion I have had the opportunity (and inclinaton) to check "why am I being served this ad?", I've not usually found it convincing. I don't really question why I'm getting Youtube suggestions about things SpaceX (because I'll readily view that kind of thing), but that's just the most logical. I don't know if it's teasing/testing me with other things, or has the wrong end of the wrong stick.

In another context, I've been served ads for islamic dating apps, which I most charitably put down to the same disengagement I have with all other ads from that source (it's ruled out most of the other potential recipients, through active negative reinforcement of some kind, and so desperately dangling a line to see if I'm the kind of person who will bite) or the geolocation is wonky (I've had ads seemingly in Portuguese, Chinese and one or other Cyrillic[1] language, wouldn't be surprised to have been ided as middle-eastern at this other time...).


And if the learning is skewed ("This kind of candidate never usually got past the human interviewers", so the suppsedly 'bias-free' system doesn't even try to put them on the short-list) it might decide based off of something a person might not even understand as tangible (certain short fragments of characters or phonemes in their name, address or past work/education history?). Or even realise, impationately, is just not right at all.

But it goes largely unchecked. Because such an aetherial algorithm as might have useful potential isn't slowly given 'examples', carefully considered by humans and the adjusted weightings then double-checked as both useful and proper. You chuck at it an entire corpus (x-ray images, written works, the answers to whether a whole mass of people thought that various groupings of pixels represent a tractor/octopus/orchid/supernova) and let it swish it around to come up with a recipe that cooks up conclusions that are closely similar to expectations. And decompressing/comrehending the myriad minor nudges that add up to the big final nudge is probably as much work (if done properly) as a basic attempt to code the 'intelligence' in from scratch.

That's just pattern-recognisers and things that other pattern-recognisers consider to have constructed a valid patternnfor. Once we start looking at true innovation and inspiration from our silicon sapiences (a way off, though we're a few steps on the road), I don't think we'd have much hope of understanding the internals without a bridging-technology to (via intermediate AI) summarise what the stew of data actually means (for which, we need to trust the less complex program).

Oh, what a rabbit-hole..!  Which we've certainly stuck our head down. We haven't (and may never) get to the point of reaching the Drink Me potion, but there's plenty of problems not that far out of reach (and actually already there).

[1] I presume Ukrainian... It seemed to be an ad for studying at a major UK university, and was well into the current invasion-era when people were not even listening to Russian composers/putting on Russian plays, let alone inviting Russians to come and stay.

[Laterigrade]

hahaha, I love this, thank you for sharing it, I have stolen it

[Maximum™]

Right? I stole it myself, it is truly fantastic.

[MaxTheFox]

lol

[Maximum™]

One year ago today, after I had already been banned that summer yet again for pointing out that Ben Shapiro is a stupid bitch (and noting in my refusal to remove the post and accept the ban that "being a stupid bitch is not a protected class" before writing the site which Musk ended up buying months later off entirely) I apparently signed up for mastodon for the first time (or thyme, maxthyme!) like many others.