AI for lonely people, GPT-3 is toxic, Tesla investigation develops, Kiwibot
29:04 Share

Hello and welcome to Skynet Today's Last Week in AI podcast, where you can hear AI researchers chat about what's going on with AI. As usual in this episode, we'll provide summaries and discussion of some of last week's most interesting AI news. I'm Dr. Sharon Zhou.

And I am Andrey Kornikov, and this week we'll discuss an AI chatbot making people less lonely, GPT-3 mimicking Reddit's offensive comments, Tesla being investigated by the federal government, and some cute robots, and some other things as well. So let's dive straight in.

Our first story is titled "Always There: An AI Chatbot Comforting China's Lonely Millions." So this is about this thing titled Xiao Ice, I guess, which was originally made by Microsoft and is now a really huge phenomenon, really. It has 150 million users in China.

And the main thing is that it's a chatbot that you can talk to as a sort of quasi friend. And what's interesting is this is kind of huge. A lot of people talk to it. And yeah, it's now something like a really highly valued spinoff from Microsoft.

I think, so I remember when CIS was first announced by Microsoft years ago, and it already had, you know, a quick buildup of users and it was really impressive. And I think now what's really impressive now is that the stickiness of it is that these users have stayed, you know,

150 million users in just China. And people are interacting with this bot as if it's their virtual, you know, in some cases, girlfriend or boyfriend and getting really close to them and also fulfilling that or filling that gap of loneliness since peak hours of usage are during the evening between 11 p.m. and 1 a.m.

And so that's when people are really just craving a companion. And Xiao Ice is always there, you know? Yeah, and I guess what's interesting is that, I don't know what other functionalities it has, but I guess the main idea is to, you know, mimic social interaction. In that sense, I think we've discussed this before, it's hard to know whether...

This is a good thing. You know, we know statistically that people are lonely or whenever and teenagers are more online whenever. But somehow we are more disconnected than before. And in some ways I could see this exacerbating that where people can use this as a crutch instead of making efforts to connect with their friends and, you know, do that sort of thing.

So, yeah, I feel a bit weird about this being this big and people being so into it. But perhaps it's also positive, you know, to help people out.

I think it's hard to say because I think we could say, you know, it's positive that, you know, it's helping people out since otherwise they would become extra depressed at night. But then on the flip side, it's like, oh, it's negative because maybe they will never want to interact with other human beings anymore. And so, yeah, I don't get the sense that people will never want to interact with human beings right now, given where Xiao Ice is and the peak hours of usage. But it's possible that

it discourages kind of a lot of human interaction because Xiaoice might be nicer. Easier. Less judgmental, you know, all this stuff. Yeah, yeah. But I mean, I guess it might be better than scrolling Reddit or Twitter. So perhaps using less social media is actually good. Yeah.

Yeah, I guess that's what you're comparing it to. If you're comparing it to the best possible thing, then we'll never really be able to make something that great as utopia, you know, but compared to the status quo, it's possible that this is filling a great need that we have nowadays.

So on to our next article, AI startups claim to detect depression from speech, but the jury's out on their accuracy. So there was a 2012 study published in the journal Biological Psychiatry and it was a team of scientists at the Center of Psychological Consultation, CPC, in Wisconsin. And they basically had

basically hypothesized that the characteristics of a depressed person's voice might reveal a lot about how depressed they are, you know, the severity of their depression. And the co-author said, you know, there were several quote unquote viable biomarkers. And this whole study was partially funded by Pfizer.

And since then, you know, a host of different startups have claimed to automate this detection of depression using AI trained on people's voices and being able to detect, oh, from your voice, I can tell how severe your depression is or whether you have depression at all.

And one of those efforts is Ellipsis Health, and they can generate an assessment of how depressed you are from just 90 seconds of you speaking. And they've raised $26 million in Series A funding.

What do you think of all this? Yeah, yeah. This was an interesting read. Certainly, I think it's interesting that this is already being piloted and is being supported by insurance provider Cigna.

And this article, as the title implies, that says the jury is still out on the accuracy, it also covers some of the concerns that people have. So, for instance, there's Mike Cook, an AI researcher who's quoted about feeling that this is unlikely to work.

And, you know, citing some things we know, like AI for emotion recognition, for instance, was a popular thing as well, where you look at someone's facial expressions. But in fact, you know, a lot of that research is not a reliable and has been shown to be problematic. So, yeah.

I don't know. I think certainly this would be great if it works because right now it's still very iffy to diagnose depression. You basically have to go to a person or self, basically rate yourself and just talk about it. So it would be great if this works.

It did work, but I guess it's too early to tell. But I think it's plausible and it would be cool if it actually turned out to be useful.

I think false positives might be a little bit concerning where people are giving self-diagnoses and then it becomes a self-fulfilling prophecy where you're like, I'm not depressed, but this says I'm depressed. So maybe I'm depressed, you know, and I don't know if that would actually happen. I could see that happening, just leading with the psychology of things. I'm also a little bit concerned about, you know,

what the downstream use case and intervention looks like, especially with insurance companies coming in. So they probably would...

amp up how much it costs if you seem depressed, you know, your actual insurance or something like that. And that would be a bit concerning. And I wonder if, you know, there are other easy variables to take into account, like some slight, you know, really small conversation with the person. Yeah. Yeah, I would imagine this would have to be used in a human-in-the-loop approach where maybe this is a first step to just, you know...

let you know that you should talk to a psychiatrist or a therapist, but I wouldn't imagine that this by itself would be enough for a diagnosis. So it is interesting to think about where this would be of use in that sense. Maybe it could be a tool that psychiatrists and therapists use in their practice to help diagnose. That would be interesting. Right. Absolutely. Absolutely.

And on to our research articles. The first is GPT-3 mimics human love for, quote, offensive Reddit comments study finds. So love for offensive Reddit comments. Do note that. So basically, a new study has found that chatbots,

very much are inclined to agree with toxic language and offensive comments over safe comments. So they're into the spicy stuff as much as we are, in a sense. So what the study was done at Georgia Tech and

and at University of Washington. And they basically looked at 2000 different Reddit threads and they paid Amazon Mechanical Turk workers to annotate them as safe or offensive. And then they had different GPT models, big language models agree with offensive comments or the safe one. And they found that these models were twice as likely to agree with the offensive comments than the safe one.

And one interesting detail is that the chatbots actually tended to respond with more personal attacks directed towards the individual, as opposed to humans were more likely to target or, you know, offend a specific demographic or an entire group as opposed to an individual. Yeah. So this is a pretty interesting study.

I wasn't entirely clear if this were pre-trained and not fine-tuned language models. I do believe they say that they looked at some existing models like GPT-3 and some other ones. And certainly this is useful in the sense of understanding what these sort of neural dialogue generation engines would do when you put them out in the wild.

And this is some pretty good evidence, as we've seen also in other cases, that

this they could lean towards more spiciness than you'd like and one thing that's neat about this paper is that they also analyzed existing controllable text generation methods to mitigate the contextual offensiveness of these dialogue models so in fact there's been work on you know making sure that language models do what you want them to do and

And they looked into this work and showed that it could improve things. So perhaps not too surprising. I mean, they also show that in general, Reddit users also display this behavior of preferring offensive stuff. So the language model just does the same. And yeah, I guess just something good to be aware of and just empirically interesting.

Yeah, it's very interesting that, you know, the spicy comment pique our interest just as much as language models. And I guess I guess a huge reason for that might be because statistically, you know, we can actually pick up on those things, even if they we can pick up on the fact that they're spicy and that they can pick up on that, too. Yeah.

I just find that really interesting because it's not necessarily intuitive, right? It's just that we think that this is spicier language because we've learned that it is. And it turns out you can learn that statistically from just the linguistic context. Yeah, so nice to see more work examining sort of downstream behavior of models and not just building more models that we don't really understand. Yeah.

But moving on to our second research article, we have AI researchers introduce a graph neural network estimator for ETA predictions at Google Maps. So ETA is estimated time of arrival. And that's what you get when you plot a path on Google Maps. It tells you roughly you'll be there in 25 minutes, 30 minutes, etc.,

And so a team of researchers from DeepMind, Waymo, Google, Amazon, Facebook AI and CII Lab, quite a team, proposed a new graph neural network for ETA predictions. And the model is already deployed in production at Google Maps. So basically, the model is meant to estimate the estimated time of arrival and

it led to reductions in wrong estimates by quite a bit. So certainly kind of an exciting result and, you know, a cool collaboration. I think it makes complete sense that graph neural networks are good at modeling the road and congestion since the road very much does seem to resemble kind of a graph like structure. And yeah, I mean, we know that Google Maps very much

It gives you an updated ETA as you drive. Sometimes I wonder if it was the true ETA from the very beginning. That's fine. That's fine. So more accurate one is always welcome, especially for, you know, really intense route planning or or diverting a certain kind of big congestion there.

And this is a huge collaboration across entities that are very much seen as competitors, right? Google, Amazon, Facebook, especially. Yeah, and it's cool to see that, you know, this is deployed on Google Maps, which is, you know, going out to millions of people and is already proving its benefits. And the paper also cites some other prior research that, you know,

led to being deployed on web scale recommender systems. So it's neat that, you know, for once, which is pretty rare, we have a paper on something that is deployed in a product versus usually I think for products, we just don't really have any idea of how they work. So it's nice to see a little more transparency for this one.

And onto our articles in Ethics and Society. The first is Tesla is ordered to turn over autopilot data to a federal safety agency.

All right. So as we've been tracking this Tesla story, Tesla has now been asked to hand over all of their autopilot data by October 22nd, or they may get fines up to $115 million if they refuse to comply or if they fail to achieve that.

And this is just part of this long investigation into Tesla cars crashing into all sorts of different vehicles, especially emergency vehicles and fire trucks. And NHTSA and the National Highway Traffic Safety Administration is the one requesting this information and is in charge of making sure things are safe on the road.

Yeah, so interesting to see this story developing so fast. I think we just discussed this investigation being a thing a few weeks ago, a couple of weeks ago.

And so, yeah, it seems like this is pretty serious. This request for documentation is asking for information on how Adapalette works, how insured drivers are paying attention, whether there are any limits on where it can be turned on, arbitration laws, lawsuits, complaints test that is achieved, and more.

And if Tesla doesn't deliver the information by October 22, it says that they could impose a fine of up to $115 million if the company doesn't comply. So, yeah, pretty, pretty, I don't know, kind of seems like a big deal. And like this investigation is really quite serious.

And a related article titled Amid Tesla's Autopilot Probe, nearly half the public thinks autonomous vehicles are less safe than normal cars. This is just...

how does the public look to Tesla now? Um, and I found some of these statistics pretty interesting in the article. So 37% of us adults said, uh, they might ride in an autonomous vehicle in the future. And, uh, 34% said they would not. So approximately the same number, uh, said they would or would not in the future ride in, uh,

self-driving car. 17% believe that autonomous vehicles are as safe as cars driven by humans. That is actually up from 8% in 2018. So that means a lot of people don't think it's very safe. That said, a lot of people, actually more than half of Americans, have not

heard much or anything about the crashes involving Tesla vehicles using autopilot, um, or the federal government's investigation in, into this. So this whole saga. And I found those statistics pretty interesting around, you know, uh, at least here in Silicon Valley, we're very much living in a bubble and everyone's heard of, you know, obviously Tesla autopilot, all this stuff, but, um, a lot of people haven't, uh, in the U S and, um, like, I guess faith in autonomous vehicles isn't as strong as it is, uh,

here in the heart of the valley.

Yeah, I agree with that as well. I found these statistics really interesting and I found it as a reminder that both of us and a lot of people in AI are pretty aware of where autonomous driving is and we know that it's not fully reliable, but it can handle a lot of things now. But outside of the AI crowd, it's fair to assume that a lot of the public is less aware of it

right now and yeah it's interesting to see this confirmed in these statistics.

Now onto our second FX and Society article, we have bias persists in face detection systems from Amazon, Microsoft, and Google. So there's been some studies in recent years showing bias, in particular with Gender Shades Project in 2018. And these companies said that they'd work to fix these biases in their commercial products for face detection.

But a new study by researchers at the University of Maryland found that these services are still flawed and pretty easy to detect waste. So all three are more likely to fail with older, darker-skinned people compared to their younger, whiter counterparts.

And this is in particular focusing on the robustness of these systems. So we don't just have normal images. We have images with different types of noise. So, you know, kind of a blur, motion blur, or maybe the image is bright or it's pixelated, things like that. So, yeah, I'm a bit surprised that there's still this bias given that

There's been years to fix it. Then again, maybe this kind of research on robustness for images that have some noise, that is a bit new. And I'm not sure if I would have expected things to work without any flaws. But of course, it's still bad that there's a bias in these systems when there is noise.

I will only be surprised when this is fixed or moderately fixed or just moderately improved because we're just not really improving on this despite all the press around it. And I hope we do. I think people are working on it, but a lot of these companies aren't prioritizing it. Yeah, we hear in the paper they said that

Amazon's face detection API was 145% more likely to make a face detection error for the oldest people as opposed to younger people. And the overall error rate for lighter and darkened skin types was 8.5 and 9.7. So still some pretty big disparities. And as you said, hopefully they are improving and I would imagine they have improved, but

It's good to see these sort of studies that keep confirming the case and kind of pushing them in the right direction. Right. And a related article is Facebook apologizes after AI puts, quote, primates label on video of black men. And so essentially, as you can probably guess from the headline, Facebook users say,

We're watching this video that showed black men in it. It was this British tabloid. But then there was an automated prompt from Facebook asking

asking whether they would like to quote, keep seeing videos about primates. And this caused, uh, the, uh, company to investigate, um, disable their AI powered feature that pushed that message. And it was definitely unacceptable. Facebook apologized. Uh, and this is very similar to the, uh, gorillas tagging, uh, that we had seen before, I think with Google. Uh, and I guess, um,

Yeah, it's unfortunate, super, super unfortunate. And I'm glad that Facebook did apologize for it. But it's like now years later, it is still still very much just not solved the same problem. Yeah. And I found this story also kind of interesting because I think it doesn't only point at bias, but also kind of.

Sometimes AI systems make mistakes just due to different inputs or something like that. And so it takes being mindful and really being aware of the kinds of errors that can happen and to design your system in such a way as to avoid these errors. And so you'd hope that...

that given the Google era that was similar in 2015, product designers would be aware that you want to avoid labels of people that might seem racist. And here that exact thing happened and hopefully that does kind of start making people more aware to think about these sorts of things.

And onto our fun articles. The first one is the time a human driven car ran over an autonomous robot. All right. So this is shown in a Tik TOK video that one of these little self-driving robots could not escape a human driving a car and actually got ran over. And in the

little autonomous robots defense, it actually appears in the video to have the right of way. And this video was captured at the University of Kentucky.

Yeah. So it's kind of a fun clip. It's very short, like five seconds. And it has this little tiny delivery robot. One of these ones that, you know, are like a little cart that drives itself crossing the street, which is already kind of funny. And then this car turns and clips it. And then you can see the robot, you know, having this wheel that kind of hangs by a thread and,

So, yeah, I don't know. It's kind of a cute, fun video. And I felt a little bad for the tiny robot. I don't know. Funny news story. I think we always feel bad for the underdog.

And the tables have turned, I think, a little bit. Yeah, because we hear and we discussed that autonomous driving Tesla is hitting stuff, but now it's this tiny, tiny autonomous little robot being hit by a way bigger car. It's kind of a fun reversal. Yeah.

Right, right. And related to that, our second article is titled, These Cute Electric Robots May Soon Deliver Your Dinner. So this is about the KiwiBot. And yeah, it's kind of exactly the same type of robot as in the last story. It delivers food and

And there's a lot of these that we've seen and discussed already a couple of months ago, I think, that this is an emerging trend. And as the article title implies, one of the selling points of this one is that it's cute. It has a little face display and can make different expressions. But there's also some interesting statistics and other stuff in the article.

Yeah, it's super cute. And they had a fantastic, I would have to say, promo video. And what is impressive, I do find the 150,000 food deliveries really, really impressive using these electric semi-autonomous robots. Yeah, yeah, exactly. I think it's smart that they're taking the semi-autonomous route. I don't know if it's a standard, but...

Definitely seems like the right way to start. And now we are expanding to San Jose, Pittsburgh, Detroit, and deploy to different universities. And now they have 400 robots since 2017 and are running pilot programs. So yeah, it seems like we'll actually have these little QWD robots driving around our city soon enough, which...

I certainly don't mind. I think that'll make things more fun. I think it makes complete sense that they've piloted on campuses, university campuses, because that is where you want, you know, a late night snack or something like that. And it's impressive that they've already deployed 400 robots. That's a lot of robots in production. Yeah. And it's interesting where this would be useful. I mean, imagine for like ultra local delivery of a shop that's like five, 10 minutes away.

this could be useful. I'm not sure what would make sense, but yeah, it appears that there's quite a few companies in this space. So interesting development. And I guess it's nice to see that not just this one, but a lot of these robots are kind of cute. And that's nice. Yeah, they're adorable.

And with that, that's it for us this episode. If you've enjoyed our discussion of these stories, be sure to share and review the podcast. We'd appreciate it a ton. Be sure to tune in next week.