Sentient AI or?

[CJay]

The URL is rather amusing as I'd expect most if not all Google engineers to be sentient, howver the article and case is interesting.

https://www.independent.co.uk/tech/google-ai-software-engineer-sentient-b2099232.html

[SiliconWizard]

I can also claim I've become the next pope.

[CJay]

I can also claim I've become the next pope.

Indeed, but as fanciful or misguided as his claims may be, they do raise some interesting ethical questions.

[tszaboo]

So here is the rundown:
Have you ever heard of Price's law? It states that in a certain company, 50% of the work is done by the square root of people. So let's say google has 100000 engineers. Than 100 engineers are doing half the work and the remaining 99900 people do the remaining 50% of the work. So there are some geniuses, and some people that, you need 2000 of them do the same amount of work. So if someone tells you that they work at google, it doesn't mean anything.

[eugene]

Whatever the truth behind the story, the photo of the guy makes it worth clicking the link.

[fourfathom]

So here is the rundown:
Have you ever heard of Price's law? It states that in a certain company, 50% of the work is done by the square root of people. So let's say google has 100000 engineers. Than 100 engineers are doing half the work and the remaining 99900 people do the remaining 50% of the work. So there are some geniuses, and some people that, you need 2000 of them do the same amount of work. So if someone tells you that they work at google, it doesn't mean anything.

The sqrt(100,000) is 316.2xxx, not 100.  But the principal is still troubling.

So say you start with 10,000 engineers.  100 do 50% of the work and 9,900 do the other 50%.  And if you apply Price's law to the remaining 9,900, that means 99.5 engineers do 25% and 9,800.5 do the rest.

Keep going.  Eventually you have 700 engineers doing 99% of the work and the remaining 9,300 engineers doing less than 1%.

Now I doubt if this uniform extension is truly valid (I would hope that there is a "minimum-competence floor" that mitigates the trend), but it makes you think.  And how many engineers would you have to add if you wanted to double your productivity?

[tszaboo]

So here is the rundown:
Have you ever heard of Price's law? It states that in a certain company, 50% of the work is done by the square root of people. So let's say google has 100000 engineers. Than 100 engineers are doing half the work and the remaining 99900 people do the remaining 50% of the work. So there are some geniuses, and some people that, you need 2000 of them do the same amount of work. So if someone tells you that they work at google, it doesn't mean anything.

The sqrt(100,000) is 316.2xxx, not 100.  But the principal is still troubling.

So say you start with 10,000 engineers.  100 do 50% of the work and 9,900 do the other 50%.  And if you apply Price's law to the remaining 9,900, that means 99.5 engineers do 25% and 9,800.5 do the rest.

Keep going.  Eventually you have 700 engineers doing 99% of the work and the remaining 9,300 engineers doing less than 1%.

Now I doubt if this uniform extension is truly valid (I would hope that there is a "minimum-competence floor" that mitigates the trend), but it makes you think.  And how many engineers would you have to add if you wanted to double your productivity?

Oh yes, by the looks of it I added a zero to 10K, and confused myself. The law states that competence scales linearly, and incompetence scales exponentially in a given organization. Being "engineer at google" really means nothing other than being able to pass the interview process. And I don't think there is a floor, research data doesn't show that. At some point the incompetent wan'ts to become a manager, and hire people to do their job while they do nothing, hiring anybody.

[bd139]

No sentient. Not AI. ML with enough corpus to reel off the right answer to an institutionalised ayahuasca sipper who needs his ego fluffling.

(know a lot of Googlers - that's as much stereotype as I could cram into a one liner  )

[free_electron]

I can also claim I've become the next pope.

There's a dead donkey just in line before you... you'll have to skip a round.

[Bassman59]

Now I doubt if this uniform extension is truly valid (I would hope that there is a "minimum-competence floor" that mitigates the trend), but it makes you think.  And how many engineers would you have to add if you wanted to double your productivity?

Seems to me the question should be, "how many managers would you have to fire if you wanted to double your productivity?"

[ataradov]

The whole story is a complete and utter nonsense. A few responses taken out of context do not make a good AI, much less sentient entity. Show me the thing maintaining an hour long conversation without looking like a complete idiot and we have something to talk about. It is still not going to be anything sentient, but it would be a fun toy. But we are so far away from this that there is not even the point in discussing that.

Just like GPT3 is not going to kill writers and DALL-E 2 is not going to kill artists, this is going to be forgotten in the next news cycle.

[radar_macgyver]

Given how high this story has trended, I wonder if this is just a really good SEO/marketing campaign for Google's NL engine. All they needed to do was find one fall guy/"institutionalized ayahuasca sipper" (thanks bd139, I'm going to use that one!)

[ataradov]

Given how high this story has trended, I wonder if this is just a really good SEO/marketing

I think the "tech" news amount is low now. So, all you have is Apple with their reality distortion field and this. And, yes Elon with his non stop whining.

There is no real point in promoting something not available to general public. And I'm sure they can provide an impressive private demo to interested parties.

[pcprogrammer]

What is intelligence in artificial intelligence.

Take a look at this video and see what might look intelligent can be nothing else then just matching pre fed information to a given question, without knowing anything about it.

[ataradov]

I think there are two different things here. One is the philosophical, and can lead to an endless and pointless discussion like all philosophy does. I've even seen some suggestions that we need to start thinking about giving some sort of personhood and rights to the  AI. This is obviously stupid, and only works for people that like to waste time doing nothing.

The second part is more practical. The question is whether we can even make something capable of sounding like an actual human being in most situations. Just maintaining a coherent conversation is enough. Because then we can interrogate that AI at length. This is where all the current AI falls apart really fast. And the test is simple. Take one of those "200 deep questions to discuss on a first date" lists and start going over them. Those lists actually do contain a lot of decent conversation starters.

And the key point here is ability to maintain context of the discussion, not just answer separate unrelated questions, however deep they may be. Some of the things that this miracle AI produces were just straight up google completion results, no need for any intelligence there.

There is also a question of how much computer intelligence should be similar to human intelligence. And I think the goal here is to emulate human intelligence, otherwise we can start calling any random behaviour "intelligence".

And the Mary's room is in a slightly simpler form is constantly experienced by real humans. Just like in the Good Will Hunting - "I bet you can't tell me what it smells like in the Sistine Chapel.". We constantly deal with things that we only know from description. I don't even see the point of this experiment. For example, I recently tried kimchi for the first time in my life, yet I've read the description of the preparation method and I knew what to expect. It was basically exactly what I expected. I guess never seeing the color (or not having any experience with other fermented dishes) is slightly different, but I'm not sure it is a significant difference.

[newbrain]

Take a look at this video

Or read this book, where some depth can be afforded to discuss those thought experiments:
https://en.m.wikipedia.org/wiki/The_Mind's_I

[RoGeorge]

OK, some journalism hype or a PR stunt might be there, but let's not try to dismiss AI.

It was the same with playing chess.  Show me a computer playing chess better than a child, show me a computer playing chess better than a professional player, show me a computer playing better than a chess champion, then they said, OK but show me a computer playing a really difficult game, like Go, and then everybody said, yeah, but that's just a machine. 

Machine or not, the performance is still there.  Progress in AI is uncanny already.  One doesn't need to be a genius to see ways of how to use AI for malicious purpose, to gain personal advantages, to get in power, or to disrupt others.  That's the biggest danger in regards to AI, for now, the danger is us humans.

Only couple of years ago USA (and a few others) pass a law to allow independent AI driven weapons.  This is as real as the invention of gun powder or nukes.  By looking at human societies vs other animals' societies, sentience seems irrelevant anyway.

I don't think humans are anything more than neural networks.

The bigger the network, the more situations it can be trained for, and the more emergent properties that NN will have, like sentience eventually, whatever sentience even means.  So far it was a major difference between AI and natural NN, being that usually the training period for AI was limited, while biologic networks are continuously fed with a stream of data, and continuously adjusting their learning to it.  For AI it used to be first train, then run.  Since the last couple decades, it is possible to do both at the same time, just like we do.

The only problem with the AI is that it is trained with human data, so it'll copy all the good but also all the human pitfalls.  Just that AI can have features we don't, like communicating between units at very high data throughput, or instant reaction times.  So far the only major pitfall of an AI is that it can be unplugged, but I wouldn't count on that in a long run.

The world is changing faster than ever, as always, enjoy the carousel ride. 

Just look at the impact EE has had so far.  Or how the Internet transformed the human society and the human behavior.  Sentient or not, AI will change the world even more.  Also, first Lamborghini was a tractor so I won't underestimate AI.

[pcprogrammer]

I've even seen some suggestions that we need to start thinking about giving some sort of personhood and rights to the  AI. This is obviously stupid, and only works for people that like to waste time doing nothing.

Oh so right, just as giving it emotion is stupid. Referring to the self driving car thread, what if that had emotion and got road rage

There is also a question of how much computer intelligence should be similar to human intelligence. And I think the goal here is to emulate human intelligence, otherwise we can start calling any random behaviour "intelligence".

What is human intelligence. Looking at the so many tests that are out there to measure it shows there is a wide interpretation on what it is. Some are talking about different kind of intelligence like musical and artsy or creative. A very interesting saying I heard long ago was "If the human brain was so simple that we could understand it, we would be so simple that we couldn't", and even though we know a lot about it, I think it is still true, and therefore will not be able to create something artificial to mimic it completely.

And the Mary's room is in a slightly simpler form is constantly experienced by real humans. Just like in the Good Will Hunting - "I bet you can't tell me what it smells like in the Sistine Chapel.". We constantly deal with things that we only know from description. I don't even see the point of this experiment. For example, I recently tried kimchi for the first time in my life, yet I've read the description of the preparation method and I knew what to expect. It was basically exactly what I expected. I guess never seeing the color (or not having any experience with other fermented dishes) is slightly different, but I'm not sure it is a significant difference.

For me it was more about the "Chinese room" that I pointed to the video. Get input, look up the required response and return it without knowing what you are looking at.

[radar_macgyver]

I've even seen some suggestions that we need to start thinking about giving some sort of personhood and rights to the  AI. This is obviously stupid, and only works for people that like to waste time doing nothing.

My beef with the story is that there seem to be far too many people who fall into this category, or at least, they get a disproportionate amount of media attention.

As for underestimating AI, no, that would be foolish. This is just not AI though.

[ataradov]

What is human intelligence.

I would say that ability to maintain reasonable conversation is good enough. There are different levels of intelligence between humas, for sure. But there is still plenty of common ground.

For me it was more about the "Chinese room" that I pointed to the video. Get input, look up the required response and return it without knowing what you are looking at.

And that's what a long conversation is supposed to address. In a conversation you can't get away with canned answers, since correct answer would depend on the whole history of the conversation.

And the biggest issue here is what was used for training data. There are abstract questions that could have answers be derived from generic knowledge, but no AI will never know the score of the soccer game that happened yesterday. And even if it has access to the internet and can lookup this data, it should be able to talk about specifics of the game. Or even "remember when operators zoomed in on that cute couple". Or talk about recent TikToks. You know, normal human conversation, and not an answer to a direct question "are you sentient?"

I for one would absolutely love to chat with AI. But all currently publicly accessible chat bots are dog shit.

[RoGeorge]

I for one would absolutely love to chat with AI. But all currently publicly accessible chat bots are dog shit.

Have you tried GPT-3?  It's the star in the last 2 years or so, and that's only public research AI.

I think the turing test has been passed some years ago.  Look how fun it is, a topic not far than this weekend, here on EEVblog:  https://www.eevblog.com/forum/chat/video-ai-that-generates-code/

[ataradov]

OK, some journalism hype or a PR stunt might be there, but let's not try to dismiss AI.

Even if not dismiss, I would give much lower priority to AI. NNs are extremely useful, but don't give them human qualities. They are not AI, they are just dummy tools.

It was the same with playing chess.

It still boils down to raw processing power. NNs helped a lot and non-NN engines are in the dust, but in the end improvements in computer performance in chess are due to improvements in the hardware.

[ataradov]

Have you tried GPT-3?  It's the star in the last 2 years or so, and that's only public research AI.

It is not publicly available last time I checked. And they heavily filter results, so all the stuff you see from people that have access could be safely ignored. They select one good result out of 100  garbage results.

DALL-E 2 is another thing that would be interesting to evaluate in the open, but for now all we have is filtered marketing hype.

[pcprogrammer]

What is human intelligence.

I would say that ability to maintain reasonable conversation is good enough. There are different levels of intelligence between humas, for sure. But there is still plenty of common ground.

I agree to disagree. For me it is also about understanding and the ability to use gained information properly. It all depends on the level of intelligence of whom you are having the conversation with. To generalize a person with an average IQ (100) might have a reasonable conversation with someone a little above or below but not with another who is way higher or lower. And then there is also language. A very intelligent person that only speaks for instance Farsi won't be able to have a good conversation with an other person that only speaks English, but still might understand each other.

But as with all human knowledge it is based on human perception and interpretation, so what is the real thing? For me nature is tangible and therefore real, but again that is my perception. And like you wrote before that is leaping into philosophy

[ataradov]

For me it is also about understanding and the ability to use gained information properly. It all depends on the level of intelligence of whom you are having the conversation with.

For sure. Bit I'm not talking about maintaining a conversation with me personally (although in a long run it should be able to do that to be useful). I'm fine with finding the dumbest person out there and having it chat with the AI. All I need is unedited logs. If in the end I can't tell which one is AI, it is all good. This is essentially a Turing test, which  some people try to make obsolete, but I still think that it is a valid first test.

Also, this engineer was suspended. Based on his previous statements he seems to be a bit out there. So, I don't think it is a marketing campaign, just one loony going rouge.