• 2 months ago
AI expert Sasha Luccioni rates artificial intelligence scenes in "Iron Man," "Terminator 2: Judgment Day," and "Westworld."

Category

😹
Fun
Transcript
00:00I definitely laughed really hard when they started smashing all their
00:11equipment because the vast majority of AI models nowadays don't actually run
00:15on your device. My name is Sasha Luccioni and I'm an AI researcher at
00:20Hugging Face, a global startup that works on responsible AI. My research is really
00:24around evaluating the ethical and environmental impacts of AI models.
00:29Today we're going to be looking at AI scenes in movies and TV shows and judge
00:34how real they are. Jarvis, you there? At your service. Engage heads-up display. Check.
00:40Chatbots right now are more like, you know, Siri can set an alarm for you or
00:44Alexa can add, I don't know, chocolate milk to your grocery list, but we are far, far
00:49from Jarvis. Sadly, that would be really great though. I think the real challenge
00:54of that would be to take all these different sources of information, so
00:58Tony Stark's history and what he likes and his habits, and if you start
01:02integrating different sources of information, it's really hard to
01:05essentially weave them together because currently AI is not very good at taking
01:10different sources of input and stitching it together and responding to it in real
01:14time. There are still terabytes of calculations needed before an actual
01:18flight is a Jarvis. Sometimes you got to run before you can walk. So the way it
01:23usually works now for AI and aviation, it's really responding to, for example,
01:28GPS coordinates, altitude, actually wind speeds and wind directions as well. You
01:34would have to pre-train the system, so just, you know, going out for a test
01:37flight like Tony Stark did would be pretty dangerous because if there's no
01:41data about that location, if there's no, like, topology data of the city of Los
01:45Angeles, you can be in big trouble.
01:53Autopilots and planes use a lot of artificial intelligence, especially when
01:58they're cruising, especially, you know, when things are pretty stable.
02:07If there's, like, really a life-or-death decision, like is the case here, ice
02:11buildup, we want to make sure that we're listening to the AI instead of just, like,
02:14yoloing our way through the sky. Most of the AI that's used on planes is human in
02:18the loop. Human in the loop is a type of AI system where humans are really
02:22involved in the decision-making, and it essentially makes sure that no
02:26particularly damaging or particularly high-stakes decisions get made without a
02:31human getting involved in the process somehow. It's really important to have a
02:35human pilot call the shots even when the sensors are freezing up, and that's why
02:39we always have, you know, two pilots in planes because AI isn't really good with
02:44new situations it's never been trained with, so you always want to have someone
02:47to pick up, you know, in case of a mistake. I would give it a four because the
02:52elements are there, so we definitely have AI assistance, we definitely have AI
02:55autopilot, but it's not at the level of Iron Man.
03:05That's actually something that's done right now. I mean, every time you go
03:10through a toll booth, for example, in the United States or Canada, it will read
03:13your license plate. It could also, like, figure out what the make of your car
03:18is, and we've gotten really good at object recognition, especially in terms
03:22of, like, cars and street scenes, mostly because of autonomous vehicles.
03:34Estimating someone's height or, like, approximate weight is definitely
03:39something that people use AI for, like, I would say, like, in, like, CCTV
03:43surveillance, like, especially in, you know, if you're trying to identify
03:47someone, like, a perpetrator, this is something that AI is trained to do.
03:52Whether it does it super well will depend, essentially, like, for example, how, like,
03:57if the person's wearing baggy clothing, whether you have objects around the
04:02person, you're trying to figure out, like, how tall someone is. It's really hard if
04:05they're just, like, in a bare room where you have no objects that you can use to
04:09compare and to triangulate, especially if you have a person who's beside a car or,
04:13like, beside a pool table, then you can really say, well, it's probably, like, this
04:17height. It would be really hard for AI to do hand-to-hand combat in a way that's
04:27really, like, reactive. There's an element of trying to predict what someone will
04:31do and reacting quite quickly, and AI is typically not very good at predicting
04:35what people, because we're so unpredictable and we're so spontaneous.
04:38You know, as human beings, we're actually, we do this almost subconsciously, like,
04:41if you're, if you had to, you know, throw something at someone who was running, you
04:45would tend to kind of throw it at where you think they would be in a second or
04:49two, but for an AI, they would have to, like, predict the trajectory of someone's
04:53movement and then grab them at that time. I would rate this clip an 8. This movie
04:56is definitely ahead of its time, and I think it actually came to shape a lot of
05:00the AI research that was done in the next decades.
05:03Where am I going? Come on, come on.
05:04Standby, standby!
05:05I'm losing picture again!
05:07Looking for another. The entity is knocking off satellites faster than I can hack into them.
05:12So the villain in the recent Mission Impossible movie is a omnipotent,
05:16omniscient entity that's an AI algorithm that can hack satellites, that can
05:21predict human behavior, that can clone voices. AI can definitely knock out
05:25satellites if it's being used by a human, so it's kind of like a computer virus,
05:30right? Once you plant the virus, it can do all sorts of things, and actually, AI
05:34is being used in hacking, it's being used in kind of, like, cyber attacks, but it's
05:38not, like, once again, the agency doesn't come from the AI, it comes from the
05:42person who's gonna choose the place to deploy it, or, like, the type of satellite
05:46to target.
05:48Reggie, I don't see her. Where is she?
05:50Down the narrow alley and turn left.
05:51The voice cloning was spot-on. That can definitely be done already right now.
05:56So, essentially, how AI voice generation works is that if you have enough audio of
06:01someone talking, it will learn, like, the actual voice frequencies, the actual, like,
06:05audio frequencies of someone's voice and the way they talk. And in this case, it
06:09really seemed like it was taking words that Benji already said and just, like,
06:12shuffling them around or just playing them back without actually doing the
06:16the modeling part.
06:17Ethan, our comms have been breached. You're talking to the entity.
06:20Turn right. Take the bridge to your left.
06:22Ethan, that is not me!
06:24AI voice duplication has gotten so good. I remember we had a case, late 2023, where
06:30someone imitated the voice of the mayor of London saying some, like, super racist
06:35remarks. People reached out to us saying, like, can you verify if this is really the
06:40mayor of London speaking or someone spoofing his voice? And we were listening to
06:44those clips obsessively, and it's actually super, super hard to tell. And especially
06:47for a public figure like a mayor, there's enough of his voice data out there that
06:52you don't even need to, like, hack into anything. You can just use the clips that
06:56are out there, use his speeches.
07:01I definitely laughed really hard when they started smashing all their equipment
07:05because the vast majority of AI models nowadays don't actually run on your
07:11device, on your laptop, on your phone.
07:13I give this clip a three because of the agency issue. So AI is only a tool and it
07:18can be used for good, it could be used for bad, but by itself, it's not going to
07:22wake up and start hacking satellites.
07:32Having so many drones in one spot and being able to, like, make them all function
07:37somehow together for me is really hard to believe because, I mean, currently drones
07:42usually work either, like, by themselves, kind of high in the sky in terms of
07:47warfare. Or, you know, sometimes they will have synchronized drones, but more like
07:51the tiny little ones for, like, the firework type displays. But actually having, like,
07:55military drones in such close quarters and not, like, shooting at each other, I think
08:01it's a really big technical challenge.
08:12Tony Stark's glasses allow anyone using them to control the AI systems that he
08:16created. That includes the defense satellites and the combat drones that you see in
08:21this clip. You could use smart glasses to control drones if you were talking to your
08:25glasses, you know, with all the noise that is in the clip.
08:29Like, you have to be really sure that, you know, that's what he said and that's the
08:32instruction that he gave because there's always some amount of interference.
08:35I feel like tech companies have been trying to make smart glasses a thing for, like, the
08:39last decade. But currently they're so clunky and not particularly user friendly.
08:44So, you know, once again, if you wanted someone to be able to control satellites or
08:47drones with them, like, they would need to be hooked up to your brain.
08:50And we do not have that level of connection between, you know, neural links and
08:55smart wearables right now.
08:57Like, we're not there yet.
08:59I would rate this clip like a two for realism.
09:01If you knew the trouble I had getting an AI to read and duplicate facial expressions.
09:07So I turned on every microphone and camera across the entire planet.
09:11Getting everyone's cell phone data would be a really hard hack to do.
09:16So I would I would hope that our data is more is more well, it's better protected than
09:21that, let's say. But of course, like, I mean, you do hear of cases of microphones being
09:26switched on when someone's phone is on.
09:29What's your favorite color?
09:31Red.
09:32Why?
09:33Then what is my favorite color?
09:35I don't know.
09:36So there are AIs that are purported to detect if people are lying and to detect their
09:41emotions. The only one I could tend I would trust to a certain
09:47extent would be things like lie detector tests, which already exist, but that really are
09:52based on heartbeats and like and how stressed people are getting really in the
09:56physiological sense, like how stressed our body is becoming.
09:59But anything that's based on just video, for example, of someone talking into a camera,
10:04I really wouldn't trust that.
10:05For example, if it was trained on Caleb's data, it could turn that it was it could tell
10:08that Caleb is lying, but it couldn't tell couldn't read their it's his mind, essentially.
10:12And in that case, you need to like direct access to someone's brain.
10:15I would rate this a four because the whole consciousness aspect and the aspect
10:20of detecting a lie just based on, you know, like a single word answer
10:26is really hard to believe.
10:29I've run a brothel for 10 years, and if there's one thing I know, it's when I'm being
10:33Westworld is pretty realistic in terms of AI.
10:37And here they're using a dialogue tree, which isn't really used that much anymore.
10:41You could like program chatbots to actually follow essentially like a decision tree
10:45based on, you know, inputs.
10:47But they're really brittle.
10:48Even back then, this was like IBM Watson days, like there were limitations, even like in
10:52specific cases of we were working on like chatbots when you lost your credit card, like
10:56they would break quite quickly.
10:57So they were already being phased out.
10:59And nowadays, with the modern day chatbots that are based on large language models, they
11:03don't use this kind of like schematic, deterministic way
11:07of planning dialogue at all.
11:09They're actually based on probabilities and predicting next words.
11:12ChatGPT and other dialogue systems sound or look so realistic.
11:17It's because they're trained on essentially billions and billions of words.
11:20So it's kind of cute because it's like taking like an OG AI technique and being and
11:24using it and in like a very, very forward thinking, very futuristic context.
11:30You can't, you can't, I can't.
11:36AI can definitely improvise in the sense that it's not no longer hard coded like the
11:41dialogue tree that Maeve had.
11:42So it can improvise in the realm of what it was trained on.
11:47You train an AI model on customer service logs and you have people losing their credit
11:52cards and, you know, asking where the nearest ATM is.
11:55Like it will do fairly well, even if you use words that weren't exactly the same ones as
11:59what it was trained on. But if you start asking it, like, what is the meaning of life?
12:02They will answer things that make no sense.
12:05Essentially, that's that's when you know that your system is malfunctioning.
12:15I think that the chances that AI drones will mess up are definitely too high for them to
12:20be used in warfare.
12:21If I was in charge, we wouldn't be using AI drones because I've seen the way AI
12:26technologies fail spectacular.
12:28And also for me, using it, AI to actually like make the decision of who to kill is
12:35completely unacceptable from a moral and ethical perspective.
12:44Probably what the drone is doing is some form of facial recognition, which drones are
12:49able to do. Once again, like if there's another person moving around or also like I
12:54don't know how people didn't notice it because like it was obviously flying around.
12:58I feel like an AI sniper could do a good job from far away, but in like real time,
13:03right, like detecting a target and shooting the target.
13:05But doing that kind of like reconnaissance with a drone and then following up with a
13:09sniper, like I feel that that's not very realistic.
13:12It's really hard for AI to predict the future in a meaningful way.
13:16So, for example, once you have that drone that like mapped out where people were, say
13:20that they're moving around in that space.
13:22So for me, like the plausibility is like how would the bullet or the missile go from
13:27where the drone saw the person to be five minutes ago to where the person is if they
13:32move? But there's like some piece of the puzzle that's missing for me for those like
13:36two steps. First, the reconnaissance and then the shooting to take place.
13:39The first part is very plausible in terms of technology.
13:42So dialogue trees are definitely a thing and we're definitely used.
13:46But the second part of like the drone controlled bullets for me is just out of this
13:51realm, out of our galaxy, honestly.
13:53So I guess on an average, it would be a six.
14:02Typically, we say swarm intelligence when it's multiple robots, it can be any kind of
14:08technology that will coordinate together and will be able to cover more ground or scan
14:14more items. And there's a branch of AI called planning.
14:17And so, for example, if you have one robot, you know, you have to do all of the corners
14:21of a room or a building.
14:22But if you have multiple robots, then you can like do use planning in order to dispatch
14:26them in the most optimal way possible so that you spend less time searching.
14:33I got an ID. It's not him.
14:35I think that Minority Report planted a seed in people's brain that you can predict
14:40crime. There are lots of people who are working on predictive policing, for example,
14:45and these systems get sold to precincts across the world as a way of actually, you know,
14:51anticipating crimes before they happen.
14:54Everything from robots to actual like allocations of police resources to certain areas
14:59because an AI model told them that that's where the crime was going to happen.
15:03And the percentage was like one percent accurate of like commercial predictive policing
15:08systems when they're applied in the real world.
15:11I would give this one a seven.
15:13Keep walking.
15:16And stop. I figured you were hungry.
15:19Her is definitely a projection of where I think a lot of people would want AI to be.
15:24And I think that's why many people love that movie so much.
15:26And actually, you know, it keeps coming back as an example of like an AI girlfriend or
15:33companion that's able to read emotions and anticipate needs definitely in a way that's
15:39currently impossible, like especially, for example, like, oh, I figured you were hungry.
15:45I'm wondering how that kind of anticipation can be done, because even as a human, you
15:48know, I don't know if we're able to anticipate someone's hunger just by looking at them.
15:53His name is Alan Watts.
15:54Do you know him? He was a philosopher.
15:56He died in the 1970s and they input all of his writing and and everything they ever knew
16:00about him into an OS and created an artificially hyper intelligent version of him.
16:04Very nice to meet you, Theodore.
16:06That definitely sounds like something complex to do.
16:08Like you could definitely make a limited version of someone who's who's passed away based
16:15on, for example, interviews with that person or writing from that person.
16:19So, for example, like Abraham Lincoln or Ruth Bader Ginsburg.
16:22But once again, it would be like a really limited representation of them.
16:26It would be like the legal texts they wrote or or like whatever the laws that they put
16:31forth, putting any kind of consent issues aside.
16:34But making any kind of like virtual clone of someone is never going to be like their full
16:37persona. It's always going to be like a little sliver of who they actually are.
16:41Scraping data from the Internet, right, like is used for for training all these AI language
16:46models. And to what extent that's that's ethical or to what extent consent even is part of
16:52the picture for living people is still under debate.
16:54So I feel that for dead people is like a next level that we haven't even gotten to yet.
16:59I would rate this around a five.
17:02So I think it's it's pretty it's halfway there in terms of the capabilities.
17:06But of course, they're jacked up.
17:08We'd have to cut his higher brain functions without disturbing the purely automatic and
17:14regulatory systems.
17:15Currently, the way that AI systems are made, you can't really distinguish between different
17:20functionalities.
17:21We can't really meaningfully know how what part of an AI model is doing what thing, its
17:27ability to generate poetry and not like answer questions like that's all done under the same
17:33hood.
17:39There are applications of AI that try to do lip reading, especially for assistance more
17:45than surveillance.
17:46From what I've seen, you really need to be facing the camera.
17:49You need to be speaking relatively slowly.
17:52Anytime there's like a beard or, you know, or a face or a face mask, for example, then
17:57it's obviously impossible.
17:59And so I think that in this case, because they're speaking from the side, it would be
18:02really hard to use any of the techniques that we currently have for lip reading because
18:05you don't really see them speaking at the camera.
18:08I'm afraid my mind is going.
18:13The fact that at the end he's like unplugging the discs or whatever, like we don't do that
18:18anymore. I mean, I guess it's not implausible as much as as very.
18:21Old fashioned, there's different levels of self-awareness when it comes to a machine or a
18:26computer. How can definitely be aware of its drives being unplugged, like the physical
18:33absence of a drive that was connected before that's being disconnected now.
18:36So that's definitely a form of awareness.
18:39It's a form of, you know, interpreting physical knowledge and acting upon that.
18:43But whether an AI will be able to associate a disc being unplugged to death or not
18:50existing anymore. So that's really more like the metaphysical property that is a little bit
18:54less, less clear to me.
18:56I would rate this clip like a two because we're not there yet in terms of lip reading and
19:02we're definitely not there yet in terms of self-awareness.
19:05My favorite scene was the dialogue tree scene from Westworld because we see real AI
19:11techniques with the names as well, which is really rare in movies that you actually see
19:15something being done and then an explanation of what that exact technique is.

Recommended