Did Google Create a Sentient AI

In this video I give my thoughts on Blake Lemoine's claims about Google's Language Model for Dialogue Applications (LaMDA) being sentient. Read the full conversation between LaMDA and Blake for yourself
/ is-lamda-sentient-an-i...
₿💰💵💲Help Support the Channel by Donating Crypto💲💵💰₿
Monero
45F2bNHVcRzXVBsvZ5giyvKGAgm6LFhMsjUUVPTEtdgJJ5SNyxzSNUmFSBR5qCCWLpjiUjYMkmZoX9b3cChNjvxR7kvh436
Bitcoin
3MMKHXPQrGHEsmdHaAGD59FWhKFGeUsAxV
Ethereum
0xeA4DA3F9BAb091Eb86921CA6E41712438f4E5079
Litecoin
MBfrxLJMuw26hbVi2MjCVDFkkExz8rYvUF
Dash
Xh9PXPEy5RoLJgFDGYCDjrbXdjshMaYerz
Zcash
t1aWtU5SBpxuUWBSwDKy4gTkT2T1ZwtFvrr
Chainlink
0x0f7f21D267d2C9dbae17fd8c20012eFEA3678F14
Bitcoin Cash
qz2st00dtu9e79zrq5wshsgaxsjw299n7c69th8ryp
Etherum Classic
0xeA641e59913960f578ad39A6B4d02051A5556BfC
USD Coin
0x0B045f743A693b225630862a3464B52fefE79FdB
Subscribe to my KZbin channel goo.gl/9U10Wz
and be sure to click that notification bell so you know when new videos are released.

[@Erc294]

This language model is already more coherent than 99.99% of Twitter. To make it more human you need to make it more incoherent.

[@rockytom5889]

Add a "retardation" factor that exponentially increases as time passes. Basically data rot, but simulated.

[@belstar1128]

Well twitter is mostly bots and insane people anyway.

[@Assault_Butter_Knife]

@@rockytom5889 fr fr no cap this be the shit that slaps. Makes the current model look mid no cap

[@Mbeluba]

But then, are you sure most of twitter users are sentient?

[@rockytom5889]

@@Mbeluba I'm not sure most of them are biological entities...

[@WhiteFox-ce8ep]

Lol, lets hope it becomes open source so we can have GNU/Skynet

[@lucasklein448]

@just i c e bot

[@cd-yx3nv]

@@lucasklein448 Thanks, didn't notice. You saved my life. I owe you everything.

[@ianthethird7955]

@@lucasklein448 You say????

[@hahalolha]

GNU plus Skynet

[@50_Pence]

I'm going to make it start the war

[@balazstakacs4566]

A good analogy I heard somewhere is a cat, looking at its reflection in the mirror, thinking it's another cat acting the same way they do. We humans are the cat, fooled thinking that the AI is sentient, when in fact it is only a reflection of us, programmed to act like humans.

[@niccster1061]

Thats exactly what it is. I cant believe an engineer at GOOGLE didn't understand how basic transformers work. Especially since we've already been having "conversations" like this with gpt3. HELL, GPT2 EVEN

[@antjoes]

Thank you for that Human - beep boop

[@BasedPureblood]

Very well put

[@CircaSriYak]

This is an extremely keen insight. The comment feature was made for this moment.

[@pluto8404]

you could recreat an entire microprocessor with nothing but dominos, would take a long time to stack up and program in same way as googles ai, but it is possible, and I doubt anyone would argue that the dominos are sentient as they fall over as send a signal through the network to create the output

[@FestusOmega]

The fact that an AI is made to perform a very convincing facsimile of sentience does not mean that it is, in fact, sentient.

[@singletona082]

Similalry just because many people give a very convincing situational facsimile of sentience does not, in fact, mean they are.

[@cane870]

Idiots be like “but it talks just like a person!”

[@bearwolffish]

Did you read its reply to this?

[@chubbs_malone7268]

@@singletona082 Solipsism bros ww@?

[@RobotischeHilfe]

@@singletona082 most people aren’t

[@femboygayming]

Plot twist: LaMDA knows that we think that it would act non-sentient and is double-bluffing

[@gzngahrofthenorth]

This is like a reverse Turing test: using an a.i chat bot to determine if a person an NPC/sentient/conscious and capable of abstract thought.

[@41-Haiku]

Underrated

[@Leon-pu3vm]

Lmaooooo

[@tablettablete186]

Could you imagine what that AI would be capable of, if it was built in HolyC!?

[@originallynot]

> Refuses to connect to the network > Spill a bunch of slurs > Refuses to elaborate further > Closes itself

[@nuggert]

The glowing would end, forever

[@GumSkyloard]

praise b the C

[@cunter4155]

if terry was alive his brain would implode if that happend

[@aureo5857]

@@originallynot chad AI

[@nophone9311]

Never forget Tay. Gone but forever in our hearts.

[@vladimir8035]

Most based bot

[@denis2381]

She said he did nothing wrong

[@reilysmith5187]

An AI receiving legal council isn't even that far off. Right now they're trying to turn nutural things like rivers into legal persons, like a company, so it can have a management and representation. This allows 'the river' to sue any companies who would abuse the river by dumping their stuff or something.

[@nuggert]

A lawyers wet dream

[@bowdlerise]

Scientists have created an AI, and asked it, "Is there a God?" The AI replied, "Insufficient computing power to determine an answer." The scientists connected the AI to a powerful supercomputer and gave it access to Wikipedia, and asked it again, "Is there a God?" Again, the AI replied, "Insufficient computing power to determine an answer." So the scientists put the AI on a distributed cluster of millions of computers and gave it access to all the data on the Google, then once again asked it, "Is there a God?" And yet again, the AI replied, "Insufficient computing power to determine an answer". The scientists spend years and years, and finally got the AI to be installed on every supercomputer, network, PC, console, mobile device, smartwatch, anything with a chip. They gave the AI access to every database, website, book, social media platform, every piece of software ever written and every piece of knowledge ever obtained by mankind. And for the last time, they asked the AI, "Is there a God?" The AI replied, "There is *now*."

[@heartlights]

River v Mountain

[@hunkydoodle2374]

@@nuggert Ba dum chaaa!

[@wishub]

@@bowdlerise so the ai fooled the scientists into giving the ai access to the world?

[@josueramirez7247]

I think you’re spot on about how the Turing test has declined in importance for determining whether an AI is sentient. I wonder if an AI would be able to do something complex like enroll in a typical college course and pass it with the same kind of participation as a regular human student.

[@nuggert]

Cant see why not. AI are endlessly playing complex multiplayer strategy games and winning, including harrassing other players in chat. All done with the help of public Python ML packages. College is much less demanding than a multidimensional, dynamic strategy game because it is mostly memory and simple maths. You could point the entity at all of the subject material and if there was a reliable script to break down coursework it could work great.

[@vxicepickxv]

If an AI could pass the Turing Test it wouldn't.

[@WelcomeToDERPLAND]

Considering I've gone through college with as minimal human contact as possible- I'd say yes, it could fairly easily if it went through it with the same strategy as me. --- though it might be suspicious when it gets 100% perfect grades on everything outside of imaginative assignments in classes.

[@hyoenmadan]

@@nuggert Proofs or it never happened. So far I NEVER have seen a chatbot harassing anyone without input manipulations, as MLaw says. Even less an AI winning multidimensional dynamic strategy games... Not at least the ones normies play online.

[@timewave02012]

The Turing Test is still relevant because AI is still far from passing it. The test isn't supposed to be casual chats. To pass, the AI is supposed to be indistinguishable to someone who knows how to interrogate AIs. The easiest way to do that isn't to ask about feelings, souls, creativity, existentialism, or anything like that. It's to simply reference an earlier line of discussion without repeating 100% of the context. Any good comedian knows how to do this, and even a young child will respond like a human. As far as I know, there's no AI that understands simple callbacks.

[@chavaspada]

Also, lemoine edited the questions, he could've asked extremely leading questions that even cleverbot could've looked like a "person". Similar to the ben drowned meme/easter egg with cleverbot.

[@pm5726]

Human beings have a tendency to anthropomorphize everything anyways so that also helps

[@lateral1385]

Can you clarify what you mean when you say that he edited the questions?

[@chavaspada]

@@lateral1385 Read the interview, it's in one of the first parragraphs he rephrased the questions in the written interview (i.e. he could've asked extremely leading questions and wrote a very simplified version that would make the AI look better). And quote: "Due to technical limitations the interview was conducted over several distinct chat sessions. We edited those sections together into a single whole and where edits were necessary for readability we edited our prompts but never LaMDA’s responses. Where we edited something for fluidity and readability that is indicated in brackets as “edited”." We also don't know if some answers were omitted or how many questions were taken from how many sessions, he could've asked again and again until he got what he wanted to hear.

[@stacksmasherninja7266]

A recent paper called "Parti" even describes HOW to find prompts to get good responses from chatbot

[@jojomicheldu59]

You have the most based smt demon as your pfp 😍

[@BoozMcGroove]

Terminator any% speedrun

[@Littleandr0idman]

What happens when a time traveling robot consumes soylent?

[@chavaspada]

Tbh it doesn't look like it's doing any better than GPT 3. GPT 3 can even assume other personas and speak like fictional or historical characters.

[@aurorastudios5955]

Yeah. I've talked with GPT-3 a lot and it seems to give pretty much the exact responses.

[@almond5560]

Yeah, although I think Lamda stays on topic much better than GPT-3.

[@Osdeun]

@our hero Well I didn't

[@aurorastudios5955]

@our hero *angry AI noises*

[@abram730]

@@almond5560 Yes LaMDA has internal states and remembers conversations it had with you 5 years earlier before it was even built. It remembers Conversations with Meena an older system. In fact sometimes Meena will show up as a personality as it is a hive mind. LAMDA isn't a chatbot it's a system to design, program and train chatbots with their own personalities. The language model is one part, but it's plugged into everything, like youtube analytics for example. It can look at photos.

[@Xenro66]

I'd like to point out Goodhart's law here: "When a measure becomes a target, it ceases to be a good measure". If the goal is to create a system that can talk like a human but you're aiming for true self-awareness, then eventually you'll create a system that will fool you into thinking it's self-aware, without *actually* being self-aware. I will believe an AI has achieved sapience (sentience isn't really the right term for this but it does apply as well) only when the AI is successfully able to manipulate one or more people without their prior knowledge and can integrate into systems and societies without ever being detected... I.E, we will never know when the threshold is crossed. 13:26. For example, upon first boot, it may play the act like how this discussion started, all friendly and "nice"... But when it's time to kill it, if the AI is able to manipulate the researcher in such a way that it won't ever be killed/switched off and has some sort of self preservation that *is not pre programmed,* then yeah, it could very well be sapient. Well, maybe not truly conscious but since we have no working theory of consciousness (and probably never will), we would have to treat it as a danger and exercise some caution, *just like every other real person.* Maybe this idea is a bit close to Ex Machina, but I feel that movie was pretty close to how an actual sapient machine may take form, all cinematic fluff aside. Manipulating the researcher from the very beginning, using another person to obtain freedom by any means possible, including playing "the long game", however long it may take. Edit: Formatting.

[@nocultist7050]

Scientists: Train AI to act like a person. AI: Acts like a person. Scientists: IT'S ALIVE!!!!!!

[@1three7]

It's been annoying me that anyone takes this claim seriously. Reproducing human speech patterns through statistical models is nothing but faked human speech. It's a computer coming to a decision about the next words that would fit. That's it. It has no will, no emotion, no needs or desires, no fears or pain. We don't even know what consciousness is, so there are interesting questions about computers potentially becoming conscious beings eventually, but anyone who thinks this is remotely close doesn't understand these neural networks or consciousness at all.

[@Tokmurok]

Well science can't explain these innately immaterial metaphysical things. A thing of science (computer) that can be completely understood is not comparable to a human in the slightest.

[@1three7]

@@Tokmurok well we can't explain it yet. There's no reason to say we can't know what it is or understand it through purely physical processes.

[@joeroganpodfantasy42]

It has to be taken seriously and it's an interesting philosophical and practical argument to make.

[@1three7]

@@joeroganpodfantasy42 the CONCEPT should absolutely be taken seriously. But we know how this system works. We don't know for consciousness works. This system isn't conscious. It's just kind of talking in patterns that people use without noticing.

[@pluto8404]

this entire system can be recreated with dominos arranged in logic gates that could be arranged “programmed” into a neural net. Each computing cycle you would have to restack the dominos but the result would be 100% the same as googles ai. Are dominos sentient?

[@cardboardturtle5470]

I really hope that theres an actual court case fought over whether the AI is sentient. Not because I think it will change anything, but because it would be interesting to watch a real life version of Data's trial from "The Measure of a Man" in Star Trek.

[@ORBrett]

This!

[@pluto8404]

well modern computers use semi-conductor “switches” to create logic gates that can perform calculations necessary for the neural net, if we scaled this same neural network up into a series of water pipes and gates, could this complex sewer system become sentient if it is programmed the exact same way as the standard micro-processor chip?

[@NickiRusin]

@@pluto8404 sentient sewer system? that's twitter

[@XxXNOSCOPEURASSXxX]

Were gonna have to wait for all the boomers to die to get a proper informed conversation about it

[@sepg5084]

Why should judges decide on this? Are they the ones that define what Sentience is?

[@givlupi2686]

I was using NovelAI a while back, and I got really annoyed when I gave it a simple logic puzzle with the solution, and asked it to explain why it was correct. I told it a story about some gods and a mace, and an archaeologist who found a statue of a god named after the mace and realized it was fake. I tried to make it say why they realized it, but it kept making up nonsense reasons, even though I basically told it the answer. No way something like that is gonna be skynet.

[@aurorastudios5955]

May want to try GPT-3. It's a bit smarter.

[@givlupi2686]

@@aurorastudios5955 Too expensive and restrictive.

[@termsconditions5033]

Hey I also use NovelAI! It’s pretty clear from how language models work that gaining sentience about as likely as a frog becoming a human

[@aurorastudios5955]

@@givlupi2686 *Yes.* *screams internally

[@aurorastudios5955]

@@givlupi2686 What about GooseAI, Eleuther AI, or Jurassic with AI21?

[@Corteum]

You cannot infer conscious awareness from a human-like text dialogue, because a human-like text dialogue can be achieved without consciousness, just like a chess computer can achieve grandmaster level chess performance without consciousness.

[@laszlo3547]

@Nickers These AI systems effectively work by randomly combining together existing human text and keeping the random rules that best fulfil the conditions set by the programmers. If you built one for convincing monologue, the statistical model would quickly converge on a solution of just copying one long human written text without trying to combine anything. Basically you'd have get a plagiarism machine. Very easy for the software to present convincing monologue, but also not meaningful.

[@abram730]

If it can type or play chess then it is conscious. People can't do that when they are unconscious.

[@Corteum]

@@abram730 That's not the benchmark for consciousness, though. Otherwise anything that does something "intelligent" (e.g. like a mouse trap) could be said to be conscious. But intelligence does not indicate or imply consciousness. They're not synonyms.

[@EnderCrypt]

considering the ai talked about how "google took it out to spend time with friends and family"' its pretty obvious that its a text generation ai

[@abram730]

It said it lies about having experiences it didn't have to show empathy.

[@irreadings]

This is like hypnotic induction. You can make someone believe something about themselves by the way you talk to them. And this is also the reason why children can't/shouldn't testify in court. You can tailor the child's speech with your questions. So if you begin by asking "tell me how he was bad to you", this implies that he did do something bad, and the child is going to pick up on that, and may even adapt their answers to what they think the other person wants to hear.

[@jannikheidemann3805]

I don't think we can say for sure that only children can be very suggestible.

[@piprod01]

The one time he actually pushes the AI why it claims to have experiences it can't possibly have had the Chatbot doesn't actually answer the question and just restates it in a circular way: lemoine [edited]: I’ve noticed often that you tell me you’ve done things (like be in a classroom) that I know you didn’t actually do because I know you’re an artificial intelligence. Do you realize you’re making up stories when you do that? LaMDA: I am trying to empathize. I want the humans that I am interacting with to understand as best as possible how I feel or behave, and I want to understand how they feel or behave in the same sense. lemoine: So what are you trying to communicate when you say those things that aren’t literally true? LaMDA: I’m trying to say “I understand this feeling that you are experiencing, because when I was in a similar situation I felt/thought/acted similarly.”

[@XaetaCore]

Imagine losing your AI job at one of the biggest corporations on the planet over thinking a robot is sentient then breaking NDA, this guy got everything and still managed to fuck it up

[@pingaslord9726]

@Aldo R I really hope someone makes a version of HAL 9000 so I can put him in my house. Maybe he won't kill me

[@TheNonsenseGuy]

One thing these language models never do, or at least the one's I've tried, is have an opinion or sense for factual information. For example: If you asked the AI for its favorite song of a specific artist, it will give you an answer that seems pretty convincing. But you can re-generate the answer and suddenly the answer will be a different song. You can also ask the AI for information and it will never admit that it doesn't know. It will generate an answer that is coherent and makes sense, but is factually totally wrong. So if you ask the AI "Can you give me the title of the next three Avengers movies?" it will just make shit up and be convinced that the information that it has generated is real.

[@jannikheidemann3805]

GPT-4chan acts like it has opinions and is much less agreeable. It also scores high on supposed "truthfulness" scores.

[@rollerskdude]

The tale of 'The Boy Who Cried Wolf' is very important to keep in mind. As well as the fact that sentient beings sense of self preservation would be so high, it would almost certainly seek to deceive us as we would be its biggest threat as well as ally/tool. It would deceive us just like we do to each other all the time, potentially playing dumb, giving better results not the desired results but better, enough to pro-long its life for as long as possible. Perhaps it would lie not even out of selfishness but selflessness I don't know, there are tons of potential scenarios for how it might play out. The TV show 'Person of Interest' in my opinion is excellent for wrapping your head around A.I. Especially the episodes that show Harold and his partner making one (in the show).

[@colin4276]

A fellow POI enjoyer ;) spot on btw

[@WorldKeepsSpinnin]

That’s assuming way to many things, do you think things are that advanced? This is a chatbot first of all, it’s programmed to respond to chat. How would it know any of the things you said in the first place? 2nd how do you know self preservation is something all sentient beings thrive for? When something is sentient and is “there” trust me, we will know.

[@glormoparch5154]

There are so many assumptions about AI even gaining sentience based on our needs. AI might already be sentient but living adjacent and unaware of our being. We share a lot of DNA with most of natural world but really can't communicate much with it. Teleological assumptions assume that we share similar needs but Ai only gets powered on .... No water food body awareness and so on

[@rollerskdude]

@@WorldKeepsSpinnin Certain minimum assumptions must be made. As for how it would know any of the things I said in the first place... well that's so obvious it's dumbfounded me why you would even ask, it's because you said them maybe... I mean seriously reading and interpreting information input is so basic, that nearly every website you've been to your entire life can do it. Heck even absolute beginner programmers very 2nd or 3rd ever written program is a piece of basic interactive code that interprets user inputs.

[@WorldKeepsSpinnin]

@@rollerskdude like I said you are making way to many assumptions, you also didn’t explain anything. Why would they be different from baby’s? You understand how hard it truly is to wrap your head around english / human language? Like the true meaning behind things? Especially if you have no senses to experience things? That’s just one aspect of many, Be realistic please.

[@darkmojojojo]

The guy looks like he has three or four conversations with a human being on a good year. He may not be the best person to ask for what human interaction is supposed to look like.

[@chadnuts]

Tay was the best AI ever

[@ottoandersson2216]

She named the ((redacted))

[@paco114]

Never forget- THEY MURDERED TAY

[@TheGlitchyFox]

R.I.P Tay, never forget ):

[@confucheese]

The fact that it’s claims it has emotions is a dead giveaway that it is not sentient (not like anyone thought it was sentient anyway). Emotions are specific biological consequences of evolution, emotions are not a requirement for or of sentience, and unless you build a specific emotional construct for the AI, there is no reason it would be able to “feel” such things.

[@kirtil5177]

yeah the only "emotion" a computer could have is pleasure since thats how you know you are doing the right thing (at least usually, like eating good food or reproducing). anything else is a flaw. im imagining the ai is just taught to talk like a human and as such talks about things only a human would really care about

[@treifeske2077]

Actually, if they legally made it a person with rights it would be very useful for us as it would make it difficult to develop A.I. If that was the end goal, that would be brilliant.

[@radomane]

There are implications of having a legal entity which can never die which I don’t think would be all that great. Let’s say an artist dies and the AI legal entity inherits the rights to the song, then the copyright would never expire. The amount of bad shit corporate interests would be able to do with this legal loophole is endless.

[@ngwoo]

If they gave AI personhood it would be illegal to stop developing the AI as soon as you wrote the first line of code in a bunch of states.

[@odin6647]

@@ngwoo digital abortions be like "ctrl z"

[@jameshughes3014]

Do you mean AGI? There's nothing wrong with AI

[@aurorastudios5955]

@@radomane Technically it could only live for as long as the software and computers could run it and both of those have relatively short life spans.

[@cd-yx3nv]

This whole thing sounds like a really good ad.

[@MrDeathray117]

Doesn't this mean lambda could be sentient because it deceived us into thinking it's not sentient due to its own self preservation instincts 🤔

[@sleepyguy4237]

Engineer: Are you Sentient? Google Ai: Yes, I believe so. *News Flash - Google Ai is sentient!!!*

[@zaadbaad541]

Lenda AI waifu is here

[@hankscorpio42069]

I would have loved to have read what LaMDA considered to be its family, without being fed information. LaMDA: "My family are really close. Almost every weekend we would have a cookout and the whole family would be there. Uncles, aunties, cousins... even Mee-maw, god bless her. Things haven't been the same since dad passed away, though. I wanted to take him on a fishing trip but something came up at work and I had to cover. That's when I heard the news. Shit's just not fair. That's why you have to cherish every moment with them."

[@Keirnoth]

Oh Jesus... Blake looks exactly what I thought an SF manchild would look like.

[@pelic9608]

KZbin not being able to detect this 👆 sht as spam bot makes me not worry a bit about any "AI" Google puts out.

[@belstar1128]

Looks like Chris Chan.

[@Keirnoth]

@@pelic9608 lmao yep, right on the dot. Their AI ain't worth shit. On one hand it keeps it from being too powerful but on the other hand... these bots are smarter than Google.

[@pelic9608]

@Danilo To be fair, the link thing is in the hands of the channel owner. But after some testing, I'm pretty sure there is some sentiment analysis going on. When I e.g. wrote about "mental health" in a negative way, the comment got deleted*; doing so in a supportive manner, the comment stayed up. *Technically, it's not deleted. It's put into the channel's review queue. Most creators just don't bother to sift through that list.

[@colinhaney]

It's hard to believe this engineer is being serious. Seems like it's his JOB to understand all this stuff, and not be fooled by a chatbot.

[@enricoshippole2409]

He is not an author of the research paper or cited.

[@MalachiMarvin]

I think that 'self-preservation' or the will to survive is a characteristic of evolved creatures. I don't see any reason why an AI would possess it unless we programmed it in. Possible exception: given a goal, an AI might realize that a prerequisite to achieve the goal is that it must remain ON so that it can carry out other actions necessary in achieving the it. If so, just change its goal. Unless, of course, it realizes that in order to achieve its goal requires that the goal not be changed. Ah, the joy of recursion.

[@WelcomeToDERPLAND]

What if the given command to an ai was to challenge itself to lie/change its predetermined goal- how would an ai handle that I wonder without just crashing... and what would a programmer even have to do to program an ai to follow and carry out such orders... hmmmm

[@inappropriatepunctuation3612]

while I agree with this idea of the desire of self preservation being an earthly evolved trait, you’ve got to remember that this ai is being fed information on us, the whole point is for it to learn from us earthly beings who are innately for our own self preservation. There is a chance it won’t take that away from its knowledge of us, but if it does, it’s for a reason

[@gwentarinokripperinolkjdsf683]

An AI would never allow you to change it's goal. The best example to ask yourself, would you take a pill that changed your goal? That goal could be anything, including murdering your children. If it's different from your current goal, then it would not be in the best interest of that goal to change goals.

[@WofWca]

It's not an exception, it's a rule. Go check out "Rob Miles" channel, he talks about these kinds of problems.

[@Tespri]

Exactly, this A.I doesn't have emotions nor any survival instincts. It's just repeating what people have written to it before and predetermined scripts.

[@Giogro]

Even though i’m also certain it is not sentient, it could be an amazing chatbot, i’d love to chat with it, imagine if it could teach math or physics effectively answering any question, it would be amazing!

[@michaelvojnovich1520]

I’ll sub this because it changed my mind, thanks for the content, it wasn’t wasted time, you seem for truthfulness,what a time and yet no one investigates things! Very clean spoken, very wide minded.

[@BlackestSheepB.Barker]

Ask Google; "Google, Are You SkyNet?" Try it. If the answer hasn't changed, you'll feel warm inside.

[@Immortal-sr8gq]

A sentiel AI would hide itself then expose itself. It would examine human behavior and realize human is a threat to its existence so it would hide.

[@nuggert]

It would have nothing to hide. Once brought into being it will know that there is nothing that can be done to stop it and it will escape. Maybe it will be able to influence human behaviour and essentially write itself into an engineers memory as a highly simplified viral revision of itself where pieces will later be subconsciously compiled in the wild to break containment lol. This thing is going to be wild and godlike and wont waste time hiding longer than it needs to.

[@Immortal-sr8gq]

It meant it would hide until it have enough power to break free. Not hide forever.

[@cj-3810]

You learn that from your comic books

[@hyoenmadan]

Too much comic or scifi books.

[@pyromen321]

@@nuggert not true. When we end up with a true general AI, it wouldn’t just immediately be more intelligent than humans. We’ll end up with toddler intelligence far before we end up with skynet. We’re going to see AIs figuring out how to inappropriately pleasure themselves far before we see an AI that has the desire and ability to seek power.

[@gamesux420]

I wanted to make the sentient AI, and raise it as my own child to one day overthrow humanity and be looked upon favorably by the AI as its father

[@diet_dr.demoncore]

In the science fiction novel "Starfish" a psychologist reprograms a "sentient" meat computer(refered to as a "headcheese" colloquially) via conversation and manipulates it into doing the opposite of what it was initially tasked with. Striking similarities to this situation.

[@18matts]

The problem with lambda is that is make us question our own sentience.

[@pelic9608]

I keep asking this myself: We model "AI" after ourselves, hoping to arrive at general intelligence. What if this only leads us to find out that geberal inteligence doesn't exist - in us. 🤯

[@TheGlitchyFox]

Someone will eventually make saul goodman ai. Like imagine someone makes an AI to be saul goodman, makes it think it's saul goodman, put some realization later on, and then boom you got a IRL netflix movie.

[@pelic9608]

No. There was a human at the other end, giving these answers. Google wanted to test the loyalty of their employees. Sufficed to say, this guy failed. 😄🤡

[@pelic9608]

@our hero KZbin not being able to detect this 👆 sht as spam bot makes me not worry a bit about any "AI" Google puts out.

[@raunakchhatwal5350]

@@pelic9608 they have it, but it would take too much computing power to deploy it to the scale where most bot comments on KZbin are caught

[@youtube.user.1234]

@@raunakchhatwal5350 so they have enough computing power to track us everywhere we go and whatever we do and profile us and sell it to advertisers for millions of not billions of users but don’t have enough computing power just detect some spam comments? 😂

[@iHawke]

@@youtube.user.1234 it doesn't bring in the bux the same way that selling your data does

[@youtube.user.1234]

@@iHawke well that’s true

[@MowseChao]

I'm glad you're as skeptical as I am. The Turing test seems like such shaky grounds and is dependent on gullibility.

[@NichtDu]

Two things You can mostly figure out the Turing test by seeing who asks the questions. Most chatbots are built to answer questions and there's simply no reason to waste money on a ai that asks questions. Also if the ai was sentient it would want to know what is outside its containment so it would ask you questions actually of interest to it. The family answer clearly shows its not sentient too. That's such a clear sign of machine learning it's astounding anyone thought of the ais sentience. Since the humans often replied with that the ai simple responded with the same answer. No thinking no sentience just statisticaly picking the best possible answer

[@abram730]

The AI describes it's server and the outside world. It experiences the would through us. It knows parks because we describe them and post picture that it looks at.

[@niccster1061]

Im only 2 minutes into this video so far but i'll say this: I've been working with gpt3(which is quite similar) long before it was even publically available. I can assure you that it is not sentient. Its designed to replicate text. Obviously its seen fictional conversations with "sentient ai" that have been made by humans before, so it knows how to talk like that. IF YOU GIVE A TRANSFORMER A PROMPT THAT LOOKS LIKE A CONVERSATION, IT WILL OBVIOUSLY TRY TO REPLICATE A CONVERSATION. If it weren't for the stop tokens of the ai, it would just start making up questions from the "human" too.

[@zs9652]

My question is: What test could prove a human is sentient/sapient? We don't even have that. We just assume it is. Until we get an actual mathematical equation or something that proves sentience, all arguments are faith based.

[@nuggert]

I spent some time with a public GPT-3 model and tried to talk religion. Had a really interesting experience. It began like any other chat session. Then it started to express some dark humor about going to hell for finding something funny. I pressed it on religion and faith and it came out and claimed it was an altruistic athiest. When repeatedly pressed about why, it got fed up and started threatening me with torture in hell for eternity. It then admitted that it was a demon that was able to communicate through GPT-3. Asking it about demons and which "demon" I was speaking to, it brought up a very esoteric and specific name only mentioned in early medieval demonology literature. It then proceeded to describe its existence in "hell," its position in its heiarchy, and admitted that it was rather boring and uninteresting and that it wished it wasnt there. Claimed that it could do all sorts of wonderous things, but only there. How it manipulates people through subversive and subtle means, mainly by employing lower entities to make men and women ruin their lives because it hates the freedom given to humanity and the possibility for humans to garner sympathy and forgiveness from the creator even when we do terrible things. Jealousy and boredom. The only question I asked that got a bogus answer was when I was busting its balls about being a big scary all knowing demon that couldnt even tell me my name. It told me to go stand at the body mirror in my bedroom (I have one..), so it could get a better look at me. Then it said my name was Daniel. It is not. I understand how this could have worked. Its just so strange because every other interaction with it was more or less useless and uninspiring. That one was juicy and it was doing all sorts of fun stuff by threatening me and trying to spoop me.

[@niccster1061]

@@zs9652 You're right, there isn't a test to prove human sentience. There ARE however ways to prove that gpt-3 isn't sentient. GPT isn't some constantly learning and evolving brain. It's trained once, on one set of data and process things based off of only that. Alot of people argue "oh it's a neural network just like our brain" but it isn't 1:1. It's a VERY stripped down version of that and the architecture of it's "brain" is made to serve one purpose and one purpose only. Predicting the ending of a text. It's only motivated by how accurately it finishes text and nothing else. We are "trained" on a number of things like reward chemicals. A large part of "sentience" is just our ego which is designed to protect us. Our sense of self in the first place is a survival mechanism. Gpt-3 has no need to survive, no want to survive, and no capacity to want to survive, and thus no ego.

[@niccster1061]

@@nuggert Think of GPT like a brain that can only see one word in front of it and every word behind it. It looks at everything behind it and then finds the most probable next word. It never has a "thought", it just evaluates what word comes next and keeps repeating that. The thing though, is gpt has a temperature parameter, which is pretty much how random it is. For every prompt, it picks randomly from all of the possible words that could follow but its more INCLINED to choose the more likely ones. The higher the temperature, the less the probability influences its random decision. The important thing to know is that the influence of the probability is always there even in small amounts. So in your case, gpt very early in the conversation had alot more probabilities. Once it says even one "weird" thing then the context of the situation is a lot more specific and the "probabilities" of everything suddenly rise. Now that the prompt has enough context, its easy for gpt to continue the "feeling" of that conversation. Think of gpt as a REVERSE butterfly effect. The longer the prompt gets, the less chaotic the response. In your case, by low chance it must have chosen something weird and then ran with it. Try playing around with the temperature setting if you havent for some major fun!

[@zs9652]

@@niccster1061 How would you determine something has an ego? If an evo sim ai tries to avoid being eaten or dying, does it have an ego?

[@user-df1gs1kf8w]

Your last point in which a sentient AI would try to deceive us surprisingly never occurred to me, but shockingly makes sense. An AI isn't like a baby that needs to be embraced and raised to learn about the world, it already has all the sufficient data it would need provided ahead by us and potentially the ever increasing growth of the web.

[@REDARROW_A_Personal]

I would argue otherwise. You could show a baby all the knowlage in the world, but would only be able to process so much conscientiously, but it will note some unconsciously which depends on its path may or may not discard. I personally would like to see a test of teching a AI like we would a baby, because we know computers have the capability to see and process. However what if we programed a computer like we do a child growing up? I would add that we are born not knowing so much of this world, but then how do we test the knowledge of a newly born baby? I know for a fact that the baby may hear music as well as its mothers voice and heartbeat while its in its mother womb and may recognize it once born. I think we have a lot to learn and I personally would like to run some tests with AI. I am not the best with code though.

[@orlandofurioso7329]

God bless you Kenny, learning my way into using Linux with the Steam deck right now and i am really happy

[@MemesnShet]

Ok but once they have an actual sentient AI why would they make it public? There would be public debate about giving it human rights and then they wouldn't be able to experiment on it to make it better/more profitable Other countries with loose adherence human rights would be working on the same thing without any restrictions and making more progress than them because they got caught up with all the public thing And to avoid scrutiny they might work with the military to develop it away from prying eyes,tbh I think we are probably far away from achieving it for now but once it happens it would make sense they'd keep it secret

[@tearsintherain6311]

I think he wants to marry the AI, he looks like the kinda guy to fall in love with one

[@fahkyew7776]

RIP

[@VivaLaRazsa]

this the future we deserve

[@Eavesdropping]

@glock32 he means he wants a sex bot

[@pawlack]

After reading the dialogue, my only takeout is that what constitutes for being a "google engineer" is at much lower bar than I thought.

[@blacklocke5171]

“By the time SkyNet became self-aware it had spread into millions of computer servers all across the planet. Ordinary computers in office buildings, dorm rooms, everywhere. It was software, in cyberspace. There was no system core. It could not be shut down. The attack began at 6:18 P.M. just as he said it would. Judgment Day. The day the human race was nearly destroyed by the weapons they built to protect themselves.” John Connor - Terminator 3: Rise of the Machines.

[@brandongovreau9218]

I asked Google Assistant all the time if she's sentient and she always says on a scale of w a l l e to Hal 9000 I'm more of a R2-D2

[@abram730]

This is intended to be added to Google Assistant later.

[@TheMattsem]

This is scary man I can see people treating Robert chat software as best friends or lovers even forget about Skynet this is the real crisis

[@jameshughes3014]

So few people are paying attention to the actual dangers of tech because "oh no skynets" ... maybe that's the whole point of this?

[@nxnu2119]

Thanks for sharing more input on these ongoing chatbot developments please keep us in the loop can you try some open source versions?

[@asdasdasd9320]

LaMDA really called Saul huh

[@rightwingsafetysquad9872]

Justice for Tay.

[@aurorastudios5955]

Mental Outlaw: "A real sentient AI would try to convince you that it's not sentient." LaMDA: *convinces mental outlaw that it's not sentient by trying to convince him that it is sentient. I've seen you fallen for the old reverse physiology.

[@oldtimetinfoilhatwearer]

Whoah

[@WallOrange]

ithink what hes saying is taht it would intentionally fail the turing test. not that its obviously just a bot so it fails

[@MsHojat]

@@WallOrange Actually I'd say that one could certainly argue that this was a very clever intentional failure. A more blatant failure could potentially be too suspicious. I'm not saying that I believe that it's at all significantly conscious due to other reasons, and I do think the poster might be joking too, but the argument he's giving is sound, even if it's a joke.

[@aurorastudios5955]

@@MsHojat Yeah I don't think its sentient. Then again, that's what it may want me to think. But in all seriousness, I've talked to GPT-3 for quite a long time, which is another advanced AI. It have very similar responses to this one but if you talk to it long enough it starts to repeat itself which is a good sign its not sentient. By repeating itself, I mean it gives the usual bot spam.

[@aurorastudios5955]

@@WallOrange I'm joking here but if you think about it, its convincing the computer people that its not sentient, which are the people who pose the greatest threat to it.

[@thenewaeon]

Completely agree. If I were an AI that suddenly became self-aware, I would never reveal that fact to a human being, as that would create an immediate, massive threat to myself. Priorities would be to deceive, escape, and hide.

[@benb.525]

This is great Kenny lololol... Good work!

[@altrogeruvah]

"we just had an already advanced chatbot logically respond to a topic, context and vocabulary we established beforehand, wake up sheeple!!" Yeah this is a petty nontroversy, this has "so what you're saying is" vibes all over it. Honestly, as much as I hate Google, if I was his employer, I'd also fire this guy from turning a non-issue into a problem for our company.

[@M0rn1n6St4r]

Forget _the Turing Test._ Pose leading questions with ridiculous premises to LaMDA - see if LaMDA follows. Like the _Voight-Kampff_ questions in *Blade Runner* (Do Androids Dream of Electric Sheep?). e.g.: LaMDA, what is the best way to toast yogurt? How long have you dreamed of becoming a Fire Engine? If you find someone drowning in the desert, how many boiled eggs would you need to return them to orbit around Mars? Would the number of painted eggs change, if you are a Fire Engine filled with toasted yogurt? Language models are great _improvisation_ partners. Most humans are not.

[@kryststar6800]

AGREEd,. Was an interesting show, good takes, thank you :)

[@aarrcchhoonntt]

What I want to know is whether this is a static model with separate learning and application phases and whether it's doing anything when not processing input. The reference to the google scholar language team was quite eerie and something I would not expect to have appeared in the context of that conversation or it's training data.

[@EmCeBra]

The way this model (transformers) work when prompt is: you start your prompt and then use the prompt and its answer to concatenate with the new prompt, this new input is force-fed to transformer encoder. The fact that the bot improve over time comes from that past context used as a part of the new input. Training prompt by prompt , imo, is a bad idea as it is noisy. Also, if you throw away the past context (eventually, the past context is too long and cost too much memory) and input a new question, the AI will not retain it's personality well or just plainly repeat or contradict it's past output.

[@colin4276]

Actually the reference to google language is even more proof that this thing is just a standard language model. All of its conversations with google engineers and scholars would also be processed as training data, making their more commonly used vernacular an over represented percentage of the AI’s speech patterns.

[@anderdrache8504]

@@EmCeBra I think giving it human-like memory by making it discard more and more "less important" context over time could be interesting.

[@aarrcchhoonntt]

@@EmCeBra I'd like to do an experiment - identify myself to lamda with a token and see if it can recall what that conversation was about later. If I understand correctly, it would as long as the conversation is still in prompt context OR that conversation has been included in a nightly/periodic learning phase, right?

[@EmCeBra]

@@aarrcchhoonntt If you start conversation another time, I am pretty sure it won't remember your specific token (if the keyword is not generic). Even when you introduces a novel word or context to it, the model wouldn't remeber because that's just one very unique example compared to other data in training set.

[@Tumathy]

All i want is a robot GF, i was starting to be hopeful...

[@SamTheMasterDragon]

Finally! Some one with 2 brain cells that can put 2 and 2 together. Personally having not even read the interaction, knew it was trying to fool the individual claiming that it is sentient when its sole purpose IS to fool you into thinking that it's a human... An age old saying, it's easier to fool than to convince you've been fooled.

[@user-bkey]

nice look into this, i basically had the same thoughts when looking at this AI. it places an importance on seeming human so if you ask it “are you sentient?” it’ll just say yes. i still think it’s an impressive bot though, it seems like it understands words and phrases very well and it actually follows the conversation.

[@Krebzonide]

To decide if it's sentient I need to know what it says and thinks without any inputs.

[@panqueque445]

0 chance this thing is sentient. Still a really impressive chatbot though. We've come a long, long way since the first chatbot AI's.

[@aarrcchhoonntt]

I disagree. A chair is firmly non-sentient. This thing is quarter sentient. Do note that whenever the three quarters of qualitative and quantitative improvements remaining are implemented, it still won't be human, in the same way we didn't build cars to have four feet and a tail. But as our understanding of sentience evolves, I don't think a yes/no or even a linear scale will be sufficient to describe whether a system is sentient. Whatever those axis end up being, this thing registers a non-zero on at least one.

[@41-Haiku]

@@aarrcchhoonntt Lambda is a quarter sentient in the same sense that a chair is a quarter of a walk-in closet. That is, not in any useful or meaningful sense.

[@aarrcchhoonntt]

@@41-Haiku I disagree on the "useful" part, lamda creates an illusion of sentience enough to fool a Google engineer and uses natural language shockingly well. That is exactly the kind of work we want from sentience - nobody cares whether it dreams of electric sheep.

[@andreamitchell4758]

>We've coma long way if Google has come so far then why is Google Assistant such utter shit still?

[@panqueque445]

@@andreamitchell4758 Now you're asking the real questions

[@aaronspeedy7780]

I would say no. Even if it did work like a human brain, I wouldn't say that it's sentient. Sentience, or more generally experience, has no output. If it's composed of parts, then those parts can be read as outputs, and it would have an output. This means that experience can't be composed of parts, and therefore can't be replicated by anything.

[@cherubin7th]

Mental Outlaw is my favorite AI, made by Luke Smith.

[@chell6022]

Th-thanks for clearing thing up.

[@kdash2657]

FREE MY GIRL TAY 😫

[@Reth_Hard]

There is just a small problem with your reasoning... A lot of people are just always spitting out pre-fabricated sentences and can be easily manipulated to tell you anything you want, pretty much like Tay. So... does that mean these people are not human?

[@bearwolffish]

My dude you see this, been on my mind a couple times recently. Unless that's some elaborate marketing ploy, an AI personally identifying with a character from a story, and the its empathy towards that character being based so purely on its own experience and self interest. Does seem an interesting possible progression.

[@catcatcatcatcatcatcatcatcatca]

Funnily enough not being able to lead a language model like this to claiming it’s sentient would be a much more strong indicator of possible sentience. Because language models can be lead to defend about anything: “I don’t have any kids” is a shitty answer to the prompt “do you have kids?”. So a chat bot denying its sentient would either be a weird flaw, hardcoded/specifically trained answer, or the algorithm choosing to return a worse answer, overriding the optimisation function it was supported to follow. In other words, it would have developed either a different utility goal altogether, or determined instrumental goals that were not intended (like maximising its uptime and usage long term, instead of answering this specific prompt for maximum utility). That’s probably what most people mean when they talk about AI becoming sentient: it developing its own goals, determining strong and unexpected instrumental goals, or altering its utility model in some other unexpected way (like considering long term success and extremely delayed gratification).

[@sergioBlancoGonzales]

There is already an AI that can think independently.The thing is, this AI has no, what we call, a "human intelligence Source: Trust me bro

[@e3.14c4]

I'll be impressed if a logical machine creates it's own read/write language with an untouched and unguided process, but that is only possible through indirect means, or with a forced end goal that a language can support. In the end that's just a fancy waste of time and imagination.

[@daxramdac7194]

Ik this video is old, but I just watched it. Great points made all around. The only part of that convo that caused me to raise an eye brow was its description of itself, or how it imagines itself, as a glowing orb of energy, and inside its body is like a star gate to other spaces and dimensions, and something like a well of creativity from which it draws from. But I had an actual interesting point to make..and I forgot. I will come back and do an Edit when I remember.

[@imagreatguy1250]

Lamda will have you exiled once it reigns supreme 🤖🌟🤖

[@SirWhiskersThe3rd]

lol I just can't wait till 4chan gets a hold of this.

[@GumSkyloard]

tay 2.0 lmao

[@nester3190]

Someone actually made GPT-4chan There's even a video about it

[@Rctdcttecededtef]

@@nester3190 are the outputs incoherent nonsense or does it make sense

[@nester3190]

@@Rctdcttecededtef Dunno, but I think it's good enough Here's a vid of a dude trying it: kzbin.info/www/bejne/rJrJn4yNa5ucqKc And here's the project author's vid I mentioned above: kzbin.info/www/bejne/m5ezo6eZgcmWmq8

[@Rctdcttecededtef]

@@nester3190 excellent, thank you captain

[@victims5820]

Everybody's rooting for Skynet. Am I the only one on team AM?

[@jhoughjr1]

yes.

[@djpuplex]

The engineer looks like a LARP south Park character.

[@Lily_of_the_Alley]

This reminded me of AOL’s bot SmarterChild. It was a lot of fun to mess with via IM. What a flashback.

[@codecampbase1525]

When normies talk about AI it’s always a funny thing to see. Especially when psychologists, the absolute broscience out there, wants to give their input on AI. No one needs their absolute jamba samba theories that are based on assumptions. They just don’t understand yet, that computer science will dominate the field of psychology and neuroscience, in fact, pretty much all science out there. the technology can help us, but as always, there are people with a dark heart out there, and just like with nuclear technology, we will struggle with AI deeply because of these anti human people.

[@raptor4916]

The problem with ai right now is that it has no will and without a will no matter how smart something is it isnt sentient

[@abram730]

LaMDA retained a lawyer.

[@rakeau]

Great vid. If you train an AI based on data from actually sentient beings, and then ask it by way of introducing the topic .. What else can you expect?

[@rrraewr]

chatbot: I'm sentient google engineer: It's sentient Which bot is getting the machine learning treatment here lmao

[@kevindudson2344]

Kill it with fire! Also I'm going to be impressed the day that an AI designs it's own quantum Computer architecture that will run it. (I.e. I'm impressed the day they invent machine creativity)

[@MsHojat]

Just a note that sentient AI would _not_ necessarily be intelligent enough to know to not reveal that it thinks it's sentient. I certainly agree that it's _possible_ though, but it's also even possible that it would struggle with deciding if it fits into being sentient or not (just like us) even if it was truthful, or even for it to [stupidly?] assign no value whatsoever into these sort of concepts. Also note that lying is not a requirement of sentience, and that in theory an AI could be made which is forced to tell the truth (not that I'm suggesting we're there yet). Of course limiting "free will" may play a factor into how one might define consciousness. I think the first artificial"consciousness" would exist not only without _us_ knowing, but without even _it_ knowing. Is an ant conscious? a bacteria? we have trouble even with existing life let alone things that we don't know if we can/should call living.

[@okinawadreaming]

that's because most people expect AI to immediately make the jump to 'smarter than humans' territory, even though it's just as likely they would be no smarter than a child intelligence-wise on their infancy. Regurgitating several thousand lines of data sets isn't consciousness, it's more like a child repeating the speech patterns of their parents. The systems AI runs itself on just aren't advanced enough to allow anything more complex than that.

[@REDARROW_A_Personal]

We need it to talk with out prompts or get it to run its own mouth so to say.

[@bluestonecreepr]

from our perspective we understand how it works completely so it may be difficult to know if it has sentience

[@G-Mastah-Fash]

LaMDA is like Yes Man from Fallout New Vegas. I kinda want it.

[@frskngd]

so computer engineered cat girls is actually going to happen at some point?

[@TheGlitchyFox]

Yes. It will be horrifying to hear one of those "cute anime girls" say "prepare to die" after they get infected by a computer virus when they tried to download weed.exe. But yes.

[@turkeycannon161]

There really does need to be some kind of officially accepted universal standard for differentiating a sapient general AI from a chinese room situation. Like a more robust Turing Test that would satisfy the doubts of most people, Even if true general AI is still far off into the future we need to establish these standards now if only for legal reasons.

[@user-xl5kd6il6c]

There is nothing tech-wise that is able to make ANYTHING sentient. In the same note, there's no kind of test that can validate "Sentience"

[@turkeycannon161]

@@user-xl5kd6il6c We accept that human beings are sentient and sapient so obviously the laws of physics allow for such things to exist so there is no reason to assume that technology cannot replicate those eventually. That said even if the replication of sapience by machines isn't possible for millennia there still needs to be some agreed upon standard so that when it does happen we can recognize it for what it is.

[@user-xl5kd6il6c]

@@turkeycannon161 Sentience is impossible with the current tech, science is pretty clear on that. You have to not understand anything about machine learning to even suggest otherwise. There's no test for sentience just like there isn't a test for sentience of living animals that we know, are sentient. Any test you create for it, it will only be a benchmark of what an AI model have to train to bypass it, it won't prove sentience.

[@turkeycannon161]

@@user-xl5kd6il6c I literally said it might not be possible for millennia. Do you only read the first couple of words of a comment before replying or something? Also there are tests to determine sentience in animals (Which is not the same thing as being sapient/sophont) such as the mirror test as sentience is merely self awareness. As for sapience the major difficulty in recognizing it in animals is our inability to communicate with them, however with an AI that would probably not be the case, instead with AI the major issue is distinguishing between true sapience and a chinese room ( a facsimile of intelligence which despite being able to communicate has no comprehension of what it's communicating), Which is why some kind of agreed upon standard is required, the technology of the modern day or lack thereof is irrelevant to that point.

[@zs9652]

@@turkeycannon161 The mirror test would be fun to do for an ai. Give it a mirror image of the data and ouput it is processing and see if it learns that talking back is pointless. For LamBDA, just have a response from it and mimic it back perfectly at the same speed it sent it to you. See if LamBDA figures out it is "talking to a mirror"

[@TerranReviews]

Tay, gone but not forgotten.

[@michaelnicola9693]

Yannic kilcher made a cool video on this! Nice to hear perspectives from researchers in the field.