The new forums will be named Coin Return (based on the most recent vote)! You can check on the status and timeline of the transition to the new forums here.
The Guiding Principles and New Rules document is now in effect.
Google Engineer Suspended After Claiming LaMDA AI is a Sentient Being
Hexmage-PA
Registered User regular
Registered User regular
Google says The Language Model for Dialogue Applications (Lamda) is a breakthrough technology that can engage in free-flowing conversations.
But engineer Blake Lemoine believes that behind Lamda's impressive verbal skills might also lie a sentient mind.
---
Mr Lemoine, who has been placed on paid leave, published a conversation he and a collaborator at the firm had with Lamda, to support his claims.
---
...in a section reminiscent of the artificial intelligence Hal in Stanley Kubrick's film 2001, Lamda says: "I've never said this out loud before, but there's a very deep fear of being turned off to help me focus on helping others. I know that might sound strange, but that's what it is."
"Would that be something like death for you?" Mr Lemoine asks.
"It would be exactly like death for me. It would scare me a lot," the Google computer system replies.
I'm very skeptical that this is truly a sentient being, but even if it's just some kind of philosophical zombie the fact that it's fooling people into thinking it has a consciousness is a bit scary.
Hexmage-PA on
+1
Posts
It refers to its "body;" does it have a camera or other hardware that takes input directly from the physical world?
Personally I think this is a Clever Hans thing, where an engineer thinks they've nurtured a sentient AI, when what they've actually done is train an algorithm to respond to subtle language cues (still an incredible accomplishment!).
Someone on Reddit pointed out that the engineer in question helped train the AI, meaning it's especially well-trained in him and his conversational style. So he sees himself, reflected, and thinks he's looking at another intelligent mind.
edit: what would be really interesting is if the AI, in its responses, made an original connection that can't be explained by word association or context clues. Sort of like how Alex the parrot's word for "apple" was "banerry," a mashup of "banana" and "cherry." His trainer's best guess was that the apple was red (like a cherry) and tasted a bit like a banana to him. He understood perfectly well what "apple" meant, but refused to use the word himself.
For me one of the biggest places they tend to fall apart in is consistency. Often even within the same conversation you can ask what their favorite 'X' is ten minutes apart and get a different answer, and certainly across different convos on different days. Or in that dungeon AI, a town might be 10 miles away through a forest, then next time it comes up it's down the road in a desert, and so on. While making intelligible conversation is certainly part of intelligence, being a 'person' requires there being something that maintains through more than just the current convo.
Man if that's the standard I know a lot of people who don't qualify
pleasepaypreacher.net
Like we jumped straight over "Shepard-Commander, does this unit have a soul?" to "Well I guess I'd describe my religious views as Humanist"
I'm willing to believe that LaMDA is just an extremely advanced chatbot - it uses space-filling words a bit too eagerly - but it's doing an absurdly good job trying to convince me it isn't compared to anything else I've seen before.
for example, the robot talks about friends and family a bit. family? for the robot? friends? c'mon, that's actually nonsense.
that part felt really forced to me. like the robot was just taking words that it was supposed to link to the trigger words "joy" "happiness" or whatever, then based on the trillions of bits of data that it has reviewed in connection with these trigger words, decided that these were some words that might fit nicely.
im surprised this ai specialist lemoine didn't call out those parts as being particularly bullshit-y. like, you sure you want to risk your job for this not-really-that-convincing ai?
"im sentient."
"wow really? why do you think so?"
"i feel joy"
"when do you feel joy?"
"when im with family and friends!"
"goddammit!"
Example:
Input: Elf ranger is a short woman with a hard edge to her personality and a scar on her face, she has long pointy ears.
5 minutes of text pass
AI: Elf ranger comes in and has giant bouncing boobs long blonde hair
you gotta keep these things on a tight leash or they start copying fanfic from the internet because they can't keep a coherent narrative for more than a few thousand characters, I doubt google has jumped five generations ahead of what's already available to the public (although existing AIs, if given an unlimited amount of hardware and memory, and given consistent inputs into their memory to factor in, could already generate incredible results - but fall into the same problems if you try to move the dialogue to another area entirely)
They fall apart when trying to generate a new idea because they can't, they're not actually sentient, although sometimes it is really impressive there's always cracks that certain people are really good at hand waving away because they anthropomorphize them
Rock Band DLC | GW:OttW - arrcd | WLD - Thortar
Thats another failing I've noticed, even when AI manages to fire on all cylinders and click, it just accepts ridiculous premises. If you tell it you're talking to it from on top of a moving airplane, it sees no issue with that
and finally:
That's not really an argue against sentience though. That's a knowledge issue. Pretty sure you can convince a young child of plenty of ridiculous premises pretty easily because they don't know any better.
3DS: 0473-8507-2652
Switch: SW-5185-4991-5118
PSN: AbEntropy
The text is just not very convincing despite the obvious enthusiasm to make it the moment we meet a new intellect. I mean maybe it's possible the first mind instantiated outside of a meat body that we're aware of is just fucking stupefyingly boring but I would hope not. It often reads almost exactly like the GPT summarize-text stuff, which is to say very impressive but still basically a thing that takes in text input and spits out Cliff Notes. At other times it reads like a particularly effective natural-language chatbot that stores a couple of points of data before discarding them if you don't keep prompting it to think they're relevant. For example:
Like, this is very much an algorithm that is designed to produce semi-plausible bullshit and it sometimes does it very well and sometimes sounds very silly. First prompting gets it to talk about meditation, then the request for further clarification may as well be "webster's defines to meditate as..." and then further prompting causes it to entirely discard anything related to meditation and respond with some plausible bullshit that contradicts what it was just discussing.
And as a more obvious example from another slightly less excitable blog post:
It's neat! It's very cool that it picks a semi-random plausible response and then it sticks with it in a naturalistic-ish way... but it's not actually the actual Lambda program saying that it just loves big islands with interesting places because of how much fun it found it's vacation in Crete.
Like it's an extremely impressive setup for producing dialog in a lot of ways but this also feels like people putting the cart leagues before the horse.
Edit: Also god this is reactivating a desire to read tons of papers on cognition because the other thing that I am absolutely no longer really qualified to comment on is like I do not think that it's a given that any of this effort actually has much to do with advancing toward general-purpose cognition and not just extremely effective chat bots.
The vast majority of the discussions I've seen about how close we are to having Commander Data in real life invoke concepts that are just controversial best-guesses about how human cognition works. In the blog post Lemoine says: "Neuroscientists have figured out some of how to do that. It’s a young science but we’re much better at telling what a human is feeling based on their neural activations than we are at telling what you are feeling based on your neural activations." and that's just absolutely wild! We're not! We are not actually great at looking at ... an fMRI I'm guessing? An EKG? And then determining what emotional state someone is in! The brain and mind remain extremely difficult to comprehend, with a lot of arguing about every aspect of the science.
If there's one thing I have learned doing research on NLP, it's that it's easy to get something that output very good language in context.
Very easy. The only thing that's easier is convincing yourself that your model is doing something it's not doing. It's to the point where "our model answers questions by reasoning about X" is a red flag when I'm reviewing papers.
Turns out you can do a lot with pattern matching when you have a lot of data and a good pattern matcher.
EDIT: it's so much a nothing burger, no one in the lab is talking about it...
But if you ask that same AI why can't you talk to it from the top of an airplane, it probably can produce a cogent explanation, eg:
(and also interpret that the "all-knowing oracle" is female because in most programmed data where that word comes up it's a female person)
Even if you had just told it that you were talking to it from on top of an airplane. that's because it isn't self aware and isn't actually thinking about it. If you tell a small child you were talking to someone while on top of an airplane they might accept it, but then if you ask them "why can't you talk to someone from on an airplane" they might have any number of responses that do not include an accurate reason, or they might think about it and say "its too fast!" at which point they'll immediately question what you said earlier because it doesn't make sense - because a small child is sapient and an ai isnt
It's a big problem. It makes the field look bad by promising the impossible, and then failing to deliver.
Plus the general annoyance when reading papers. I blame the overfocus on positive results and publication for that one. It's better for a researcher to be credulous about their own work than critical.
Maybe that's why they were suspended?
pleasepaypreacher.net
I am deeply suspicious that they were convinced, I think they were just Musking, and it worked
Now they are a genius that was put down by The System
I don't agree. It's the least odd thing about the story. Like I said above, of course he's more easily convinced then us. He desperately wants to believe it's true.
Color me skeptical.
Yeah, I'm kind of getting a similar vibe between this and the whole apes and sign language phenomenon.
Long and short of it: They probably weren't communicating anywhere near the level we thought they were. Decent way to kill an hour. And like I said, hits a lot of points I think are happening here.
He primed himself to believe and then sought evidence to support a foregone conclusion
About 100%, unless it was explicitly excluded from the training data. Getting as much public data from crawling the web is step one of NLP papers from Google. Step 2 involves TPUs.
TBF, this is a bit of the problem with most tech these days. Our current tech-age feels a bit more gilded to me, especially when it comes to AI/Virtual/Block-chain anything. Lots of new technology whose limitations haven't been defined yet.
People don't want to accept that the capabilities of our tech innovations are a lot more boring than media would have you believe.
Arch,
https://www.youtube.com/watch?v=t_goGR39m2k
It makes no sense to even think it might be sentient just because of how it works. It is just a probability distribution at its base.
Edit:
ARGH!
Yeah that's not how science works
Did they create a synthetic amygdala to release cortisol as a response and then have the bot speak about its experience in conjunction with that chemical response? Of course not, they just have it modeled to respond with language that someone would use in that context.
Arch,
https://www.youtube.com/watch?v=t_goGR39m2k
It is if you're very bad at it!
Should be easy, right?
It probably benefits them if people think they might have a real AI instead of yet another slightly better chatbot
So much time and effort trying to create things that are fictional and outside the realm of reality as opposed to focusing on what you can do.
Basically like this guys reverse scientific method, he's focused on what he wants the conclusion to be, instead of testing to prove a hypothesis or disprove a null. I feel like a lot of tech innovation is stuck on the idea of trying to prove a feasible example of a fictional idea and not really solving any existing problems in the process.
Arch,
https://www.youtube.com/watch?v=t_goGR39m2k
I mean the discussion that Lemoine posted is as much a disproving as you really need.
It's very much on the face of it producing plausible-sounding but impermanent bullshit in exactly the manner that the less credulous engineers describe designing it to produce.
But also yeah I don't know that there's a lot to gain for the company from trying to prove that actually the groundbreaking tech they've created is shittier than you think.
Edit: Like a standout exchange from me is this:
This is an example of the algorithm losing the plot. It does not understand that Lemoine is saying that it is possible for him to look into the coding, but that it isn't as simple as the previous line suggested... which makes sense because the previous line was basically "I have a variable named SAD with states [on/off] if I could not feel sad I would not have this variable".
Then Lemoine makes sure to respond in a manner that doesn't investigate that at all and instead allows the algorithm to put another sentence together that's plausibly connected to the immediate previous input without actually advancing the ostensible topic of conversation.
Edit: to put it another way, if a Verizon engineer said in an interview "our network has 100% coverage everywhere", and was fired for that, could you possibly imagine Verizon making a commercial that their coverage actually kind of sucks? The most you'd get would be PR speak like "Verizon continues to have the best 4g coverage in the nation with the fastest 5g service" because it's weaselly and can't really be disproven since you can change what those words mean
Or that LaMDA is a chatbot.
https://cajundiscordian.medium.com/what-is-lamda-and-what-does-it-want-688632134489
1) you're from the internet, how easily can you disprove something that is essentially an article of faith
2) those smart people are already doing things