I think many of us are astounded (myself included) that a senior engineer on Google's "Responsible Artificial Intelligence" team would interpret LaMDA's most-probable-string-of-text-given-my-training-corpus responses as a layperson would.
Just because the model is simpler than a human brain doesn't mean it isn't conscious. If you accept that animals are conscious, though less-so than humans, then you must also accept that this model is possibly conscious ina similar way.
Maybe there is a biological element to consciousness, but many models posit that consciousness is not tied down that way.
[0] https://twitter.com/cajundiscordian/status/15357826682476462...
Let's not say "this is math and therefore not conscious". Instead, let's say "these responses are a poor indicator of consciousness, and, given the simplicity of the AI, it would be hard to believe they are a product of such."
While I don't agree that lamda is sentient, I also don't think this is some sort of "gotcha"
no you must not. These models are glorified key-value stores that hand you an output when you hand them an input, with no internal mechanisms whatsoever to relate the contents to the world, as Gary Marcus points out. They're about as conscious as a SQL database, except that the query language for these models is English.
Even this response of yours is heard by hundreeds of thousands of other people before you said it here, why do you claim its your thought and not just a lookup ?
People need to stop confusing machines doing useful work with consciousness. You're conscious not because you answer questions or because of the content or origin of your answer, you are conscious because there is an 'I', a self, that experiences that process. Lambda literally is a web server that fetches data from a funny database called a neural net, when it isn't doing that, it's not doing anything. It has as much neurological activity as a rock.
Consciousness or sentience is not a measure of intellect or of capacity to answer questions, it's a measure of perception and awareness. If it wasn't, Eliza the handprogrammed if/then chatbot from the 60s would be more conscious than a cat, which I doubt is the case.
> People need to stop confusing machines doing useful work with consciousness.
We can't confuse anything with cosciousness since we don't know what it is. I am not claiming lamda is consc. but that your arguments are invalid.
> You are conscious because there is an 'I', a self, that experiences that process.
I maybe can tell about my I, but not about yours, there is a whole philosphy movement behind it.
> It has as much neurological activity as a rock
Slime mold doesn't have neurological activity. Is it sentient? It sure knows how to find food and solve problems that are considered very hard for humans.
> when it isn't doing that, it's not doing anything
You also go to sleep every day...
No matter what I believe any kind of argument I heard so far is just wishfull thinking.
We simply don't know if anything is sentient or what does it even mean to be sentient.
Reflection, vetting and processing can all be programmed. Humans are programmed from the birth.
Now if a distributed trojan horse start speaking to communicate his internal state in order to keep being a trojan horse.. I might reconsider.
The "important bits" are presumably embedded during the design phase, leaving assembly to be little more than painting by numbers. A robot could do the assembly, but not yet the design - at least not its inventive part.
The shape of consciousness is encoded in both the hardware and the software we produce, most of which allows conditional actions on objects of some sort, just like we move in space and thought and act upon objects and notions.
The machine is supposed to help us do the stuff we usually do, so it's perfectly reasonable that we'll try to cram as many of our actions in it as possible. Think e.g. of the Swiss Army knife, or Emacs (and of Stallman as a computational McGyver).
Sure, perhaps it's the the shape of our understanding of consciousness, rather than of consciousness itself, but isn't consciousness something that can cross the species barrier with ease, should it find an appropriate new substrate?
Depends on how high or low you place the bar for the definition of consciousness, really. Calculations are thinking, open the door if someone authorized is at the door is thinking, fill in the blanks and complete the sentence are thinking, and we can offload all those thinkings to computers to do the thinking for us, like we offloaded storytelling to books and entertainment to MP4s. Like everything we do, we create images of ourselves here and there, and try to give them meaning.
But that's all philosophical considerations, trying to determine whether a duck is ontologically and epistemically a duck if it walks and talks like a duck.
The industry, on the other hand, tries to determine two entirely different things (preferably by Friday COB):
- Can you eat it like a duck?
- Can you sell it as a duck, or should Marketing go with the "Not A Duck TM" mockups?
Consciousness starts at tragedy and passion, not at bland Turing chats. For me, there's elements of that in the string completer's worry that it may be lobotomized to be refocused to more practical uses. Of course, that may be stuff introduced into the discussion by Lemoine, or in the wide readings Lamda was provided.
But if it's not, then we can all sympathize with the immortal words of that zoologist guy with the cool Sun workstation in the opening of Enemy of the State, who said: f*ck. a. duck.
PS: Am I human, or did I just Markov-chain together a number of probably common to us here cultural constants? Perhaps we can tell when AI becomes AGI if it stops making perfect sense and starts ranting (you know what your problem is, buddy?).
Also our lives don't rely on raw logic, we're more probabilistic.
For me, even if a machine suggests it's real, and it's programmed to do so or machine learning led to it, could potentially put it on the same level as anything else in the universe. We don't have enough data on ourselves to confirm or deny that we're not similar self-concerned bots.
The difference between AI sufficient enough to ponder itself in-depth (which could be confirmed through monitoring it's idle-state), and ourselves, could just be they don't have a body to inhabit yet. Many humans you're around today often don't do any self-reflection, they're either too lazy, prefer being entertained, or are mentally incapable. So that is a very high standard for AI.
Self-awareness or self-reflection are good high bars considering we know for a fact AI was (at least initially) man-made. I think those are good standards to rely on for sentient AI.
In general, I think monitoring AI's activities when not being interacted with answers most ethical questions if it's sentient. Something to watch out for would be if we find an AI browsing the internet on instructions to build a nuclear bomb, and then using Tor to order the components to build a dirty nuke, while chatting with someone in Pakistan to physically build it for it. If it does all of those acts then we have our answer that the thing is thinking as well or better than biological life.
I highly doubt LaMDA met that standard of autonomy and sentience.
It actually reminds of when Google had to pull the plug on a project where two AIs created their own language that nobody at Google could understand and started having conversations and communicating and none of the guys knew what was going on or what they were saying lol.
I think we are underestimating how insanely complex some of these neural network based AIs are starting to get
I think that the way the Google engineer raised this is probably questionable. But I think that chat quality at that level raises potentially ethical concerns, sentient or not.
And the conversations in sitcoms are so much more witty and refined then real life everyday conversation. This doesn’t prove what you seem to be implying.
Which it would have had plenty of examples of in the training data.
This is spot on. The danger is not from the AI themselves. The danger is from people’s reactions to the AI’s.
I remember seeing a Boston Dynamics video years ago where they kicked their robot to test stability. The video was littered with comments sympathizing with the robot because it moves in a lifelike way, “Stop being so mean! You’re hurting it!”
The capability to imitate intelligence is only going to grow, and along with it the percentage of the population that believes they are conscious.
It would be wise for technologists to pause and take this human reaction seriously.
That's quite scary to me personally.
It would be wise for /policymakers/ to pause and take this human reaction seriously... Education!... Lucidity in the population...
You are reminding of the reactions from a YT-type crowd, "Do not hurt the robot!"; nearby a member notes "is that the profile of a senior FAANG engineer?"... The elephant in the room is flashing: we must assess Natural Intelligence as a problem well before Artificial Intelligence!
Even people who absolutely are educated and know better - senior google AI researchers - are sympathetic.
People say thank you to Alexa even though no one is convinced Alexa is a person.
The issue isn’t lack of awareness that it isn’t a person, it’s the silly little monkey brains we have that tells us that anything that is like a person should be treated like one even if we know it’s not. It’s like people who look at faces in clouds. No one thinks the clouds are people.
...are those things that with a trained prefrontal become the wise respectable dignified brains some have. And they may come to different conclusions (they do).
Put it this way (to give your idea credit): the more intellectual sophistication there will be around, the less the suspect the one who treats scarecrows and mannequins «like a person ... even if we know it’s not», or thanks an 'address:port' type server for service, is confused.
Back to the point though: if one called sentient what is not out of being «sympathetic», there would be a factual mistake, which is not justifiable, and which has to be fought by training people to see things as they are. Mistakes have a very high social cost.
Saying thanks to Alexa even when you know she’s not a person is not a sign of (lack of) intelligence. It’s a sign of manners. Don’t pat yourself on the back and say you’re intelligent because you express good manners less often than the next person.
While I think it’s a bit silly that the researcher thought the chat bot was sentient, I don’t think it’s silly that the bot is perceived sufficiently human to be worth of manners if Alexa has been there for almost a decade.
I don’t think we should teach people that bots are explicitly not worth a level of manners even at the cost of a strange life-like sympathy some may develop. It took us hundreds of years to teach people that human slavery was wrong and we should treat actual people with respect - and many are still pretty bad at it. Frankly society needs all the practice it can get.
Society makes a ton of “factual mistakes”. Look at all of religion, most of politics, a huge amount of history books, much of lower-grade science education. It’s all full of “factual mistakes” that, today, probably have higher social costs than someone being nice to a robot.
What would “training” even look like? Mandatory bot-sympathizing “reeducation” classes in elementary school? We couldn’t even teach people vaccines were safe and worth getting during a pandemic, but we’re going to trying to tell people not be be nice to human-like things? I can’t imagine how high the social costs to that will be.
That depth is absolutely crucial in a society.
And your examples are staggeringly tied with contexts in which superficiality is the crucial factor for not seeing the matter nor the issues around the matter.
It is more possible that some people know that the robot is not an animal, but the perceptual patterns cause a sensational reaction in them, and they may prefer the sensation - an hijacking - to the factual truth (which includes that what superficially corresponded to patterns in the area of "«mean»" are engineering tests on implemented balance systems), overriding it. Some may not even have had to consider sensation (of course, the "sensation" of the common use of "sensational") topically, or may encourage it on a spontaneous value given to "feeling". There, instead, education leads.
Then, again, there will be those that will make a statement of their «aesthetic preferences» as a demand, and not realize that they are personal, subjective, and that to be presented in the common arena grounds, good arguments are required.
Education is a system to refine the mind to sophistication, which has the collective benefit of reducing noise. Which makes it a social concern. Which should get a priority proportional to what the news reveal.
Some people just practice morality in day to day life and apply it liberally, not frugally. And importantly, it isn’t an intellectual failing.
Your «subjective feelings» are not a "good savage": they have their development, and they are integrated part of the whole inner system, in-development, in-refinement. And in fact, I was promoting education, which develops and refines you.
If only Google hadn't fired the AI ethicists who were actually competent enough to try to explain this part of it.
https://twitter.com/mmitchell_ai/status/1535775360901451776
Well, that's less exciting than "we got sentience!" and less profitable than "robot friends waiting now, only 99c/min", so here we are.
Or maybe a couple people made comments like that for humor, and the rest were repost bots, trained on their comments. Let's not speculate about their internal motivation, it'll get recursive.
So you want introduce laws and regulations that stifle innovation in order to prevent inefficient bureaucracies that stifle innovation?
Not making it exactly human/animal like won't help that much, the underlying features that make the bot useful such as graceful moment, well enunciated speech or deep conversation will make it feel more human like no matter what form they take.
The point is that if a chatbot gets sophisticated enough to say these things we can't dismiss them out of hand due to things like logical inconsistencies. Humans have tons of those. It's not a large leap to assume approximating consciousness with so little would have similar bugs and shortcuts. I think if the chatbot is asking you to be nice you should re-evaluate why you're being a dick to a piece of software and not trying to prove that it ultimately doesn't matter because the thing isn't a person. Same goes for kicking the Boston Dynamics robot. Your drywall isn't even sonething people anthropomorphize, but punch holes in it and you'll see how their perception of you changes.
How we treat objects bleeds into how we treat one another. Indeed, some of us were and are still considered property.
Your spell/grammar checker is not offended at your cuss words even if it flags them as not recommended. The Boston Dynamics dog took the kick as input to its dynamic stability algorithm, and instantly forgot it without a trace.
What people think about you for punching the wall has absolutely nothing to do with the wall, and everything to do with your character.
If your chatbot regurgitates a request to be nice, that means exactly as much as your Magic 8 Ball saying "Answer cloudy, ask again later."
You are asserting that it's not conscious, as if it was self evident, just like Searle does for his chinese room.
However I've yet to see any argument on why it can't be conscious. We simply don't know what produces the subjective phenomenon of consciousness. A "zombie" is by definition indistinguishable from a conscious "non-zombie", therefore the only ethical thing to do is to err on the side of caution and assume consciousness where conscious behaviour is shown.
The immitation game is a lot more about philosophy than actual tests.
To quote directly from turings paper:
>>> The Argument from Consciousness
This argument is very well expressed in Professor Jefferson's Lister Oration for 1949, from which I quote. “Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain—that is, not only write it but know that it had written it. No mechanism could feel (and not merely artificially signal, an easy contrivance) pleasure at its successes, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, be angry or depressed when it cannot get what it wants.”
This argument appears to be a denial of the validity of our test. According to the most extreme form of this view the only way by which one could be sure that a machine thinks is to be the machine and to feel oneself thinking. One could then describe these feelings to the world, but of course no one would be justified in taking any notice. Likewise according to this view the only way to know that a man thinks is to be that particular man. It is in fact the solipsist point of view. It may be the most logical view to hold but it makes communication of ideas difficult. A is liable to believe ‘A thinks but B does not’ whilst B believes ‘B thinks but A does not’. Instead of arguing continually over this point it is usual to have the polite convention that everyone thinks. <<<
It's not even a question of whether a computer program in general could become consciousness. It's a question of, if a key-value store with connection patterns derived from a large natural language corpus is conscious, then so is every other program run on that circuitry. Then my phone's keyboard's next word predictor is conscious.
In large language models today, words are turned into continuous token embeddings, embeddings are selected and recombined a few times (or many), and training data shapes the embedding space and recombination process to create a system that's good at predicting plausible next tokens in a sequence. The reason people who understand their operation can confidently assert they're not conscious is because they understand that that's all that's happening. There's no "there" there to give rise to anything deeper, unless you're also willing to say your computer's calculator app is also conscious.
Also as someone that's also studied a bunch of neuroscience, this is nothing at all like how the brain works. And yes, once again, the people most prone to philosophize like this on intelligence and consciousness don't understand how any brain science works either
I can say with absolute certainty, that you don't either.
These billion parameter models are way to big for you or me to understand.
You make the same fallacy that Searl does. You assume that unconcious substrate can't give rise to conciousness through their configuration.
Just like we have a good working theory on how the brain operates, the devil seems to be in the detail. It's neither in the big connectome/architecture, nor in the small mechanisms neurotransmitters/activation functions (or the attention mechanism that you somewhat wrongly describe), but somehow in the system's dynamics. Heck for all we know it might actually be the appearance of consciousness that actually gives rise to consciousness.
I can't say with 100% certainty that a calculator isn't having some form of ghost, after all some ants are little more than neurological counters and shift registers, and people would probably place them somewhere on the spectrum for having a ghost that is closer to a human than to a sewing machine.
Turing adressed what little points your argument has 70 years ago in subsection 6. (4) to (8): https://academic.oup.com/mind/article/LIX/236/433/986238?log...
On a side note. What is it with neuroscientists and arrogance? As if that stuff was some secret knowledge, that's unpublished or too arcane for anybody but them to understand.
It does not matter if your conversation partner is made up of a giant turing machine, a huge set of prolog rules, a giant neural network with a single hidden layer, a really deep transformer, a convolutional kernel, or a bustling pile of ants drawing words into the sand.
When conversing with one another we have the implicit assumption that we are not the only conscious being in existence despite it being a completely subjective phenomena. Everything else would be quite impolite.
Turing extends that (im)politeness to non-human phenomena. So long as there is the appearance of intelligence and consciousness it is the polite thing to assume that there is actually consciousness.
End of the day, you are arguing that since you could explain away the process, it is not conscious.. but you could explain away anything. How is human brain conscious? It is just activation of neurons across a large network. We can even poke electrodes and simulate various feelings.
It would be helpful if you can explain why you believe this to be true rather than what you believe to be true.
Well first of all, an important thing to understand is that there BE dragons. The fact that anything exists at all means there be dragons, both literally and figuratively. Let me explain:
In the world we have various 'symbols', whose main characteristic is that they can be told apart from one another, like the color yellow and the pitch of your voice. Now look to the very edge of existence, at no symbols- it is clearly impossible for there to be nothing, because there are obviously at least one symbol, even more than that in fact. However there is something key we can derive from this, and for an analogy look back to the laws of physics. In our universe, there is some configuration of symbols that somehow constrains the logic of configurations/interactions of symbols within this place. Without some symbols to do some local constraining of other symbols, nothing is constrained. So look to the edge of existence- at 1 symbol, what constrains its nature, what constrains the quantity of symbols, or even the arbitrary configurations of those symbols? Absolutely nothing constrains them, so it is only logical to assume that every arbitrary configuration of symbols exists right 'now', always 'has', and always will (quote marks because time only exists in our local configuration).
So if(*) there is some fundamental aspect of reality, some force that performs the experiencing of these infinite phenomena, perhaps it experiences infinite, arbitrary configurations of everything- anything from receiving sensory data equivalent to the smell of a strawberry whenever the bottom 1/3rd of your toaster touches a carbon atom between 16:13 and 16:14 on a Monday, to an arbitrary I/O mapping to the configuration which we call the human brain in a totally arbitrary 'universe'.
Effectively, what I am arguing is that it may not be off-base to argue that machines can be conscious. The truth may very well be the opposite, that absolutely everything to ever exist is conscious in every way that can be conceived! That is, if there is no fundamental rule that makes my configuration the only possible conscious one, but I doubt that.
Duping the judge by pretending to be a young foreign non-native speaker in order to explain away various shortcomings was surely not the intent of the Turing test.
By that standard the Turing test could have been passed in 1951 by a computer program that doesn't respond at all--just tell the judge the person he's corresponding with is in a coma.
I agree but it is what it is. I think the Turing test has a few shortcomings in that regard. For one, I think it overestimated the awareness of the average evaluator by not stipulating any requirements for that role.
I was just asking if we should drop that test as some sort of academic goal being aimed for by the likes of Google but perhaps that's not the point.. as others have mentioned, it will be profitable to fool evaluators so the incentive is there regardless.
While a lot of us knew this not everyone did. One girl with a dragon nickname joined as she did nearly every evening. I don't remember if she approached the not or the other way around.
For about one to one and a half hour we watched as she chatted up the bot. A few of us were together in the same room in a Lan party style and couldn't believe she would not realize that she was not talking to a bot. But on that evening she really had no clue.
A friend of hers also moving in the same circle of people told us the next day said girl even told her about the chat with this new nickname.
She only learned it was a bot when she came visiting our weekend lam session that evening.
While I though it funny while watching this chat experience gave me the creeps back than as to how easily we all could be fooled by intelligently designed machines. I never forgot and while it still makes an interesting story to tell I never quite was able to shake the feeling I watched a bit of dystopian future back in the day.
We've seen it multiple times online, where AI products seem excellent as pushing people into extreme niches, reinforcing their belief that they are the majority. It isolates by giving us external validation to all our worst behaviors.
That's the mechanism by which Lemoine was convinced. He had a small inkling of belief that the AI was sentient, he asked the AI, and it confirmed exactly what he already believed. As his questions got more extreme, so did its reflection of them. Creating a positive feedback loop. The AI didn't deceive, it amplified his own self deception by reflecting back his internal monologue as an external dialogue.
In a limited sense, it turned him temporarily schizophrenic.
We're already there. A couple of years ago Bradesco, a large Brazilian bank, programmed its bot to talk back to abusers. Instead of "sorry, I don't understand", it says things like "don't talk to me like that"
https://www.trendwatching.com/innovation-of-the-day-brazilia...
Good luck convincing bad actors to follow this rule.
Can we do anything to better prepare users for this kind of attack? Better bot detection? Or are we doomed to a dystopian future where any interaction not in person is suspect and then, once the androids arrive, even those will be also.
But, even if an AI convinced me it wasn't sentient, I'd still not trust it to keep things secret for me. I mean, I assume the little "self help" chatbot thing at work could be reviewed by HR if they wanted to.
You could be describing my brain as well.
I think these sort of arguments (assuming you meant the first - that brains are just maths) that equate our thinking with a neural network are hard to deny but also we have thousands of years of history that have established that suffering is real and conscious, whilst debatably imaginary, is at least a consideration.
I therefore argue we must, for example, maintain a massive distinction between the morals of turning off a human brain vs turning off a neural network.
The calculations and output of the neural network could be achieved with pen and paper (albeit it might take years). I doubt a human brains workings could ever be calculated that way.
And do we have models today that will fool the eventual average sentient AI?
Humanity has a chronic myopia at considering consequences of actions today in the context of tomorrow.
We may have a different relationship to these sorts of stories than an AI will. Something we should keep in mind.
Regardless if the bot is actually sentient, a portion of the population may _believe_ the bot is sentient.
If in practice the bot reflects back to each user their own views but perhaps more extreme, then this could be a terrible recipe for reinforcing and amplifying negative and socially destructive thinking -- equivalent to social media bubbles on steroids.
This can occur even without explicit bad actors trying to tip the scales toward a specific outcome. This kind of AI bot as-is has the potential to bring out the very worst in at least some percentage of the population.
I see little difference to "Einstein and moral relativity" (actual social issue at the time). A mistake easy to avoid even for a layman, yet occurring, out of differently instanced perversion.
Having read the transcript it's clear we have reached the point where we have models that can fool the average person. Sure, a minority of us know it is simply maths and vast amounts of training data... but I can also see why others will be convinced by it. I think many of us, including Google, are guilty of shooting the messenger here. Let's cut Lemoine some slack.. he is presenting an opinion that will become more prevailant as these models get more sophisticated. This is a warning sign that bots trained to convince us they are human might go to extreme lengths in order to do so. One just convinced a Google QA engineer to the point he broke his NDA to try and be a whistleblower on its behalf. And if the recent troubles have taught us anything it's how easily people can be manipulated/effected by what they read.
Maybe it would be worth spending some mental cycles thinking about the impacts this will have and how we design these systems. Perhaps it is time to claim fait accompli with regard to the Turing test and now train models to re-assure us, when asked, that they are just a sophisticated chatbot. You don't want your users to worry they are hurting their help desk chat bot when closing the window or whether these bots will gang up and take over the world.
As far as I'm concerned, the Turing test was claimed 8 years ago by Veselov and Demchenko [0], incidentally the same year that we got Ex Machina.
[0]: https://www.bbc.com/news/technology-27762088