Making AI Less Racist and Terrible, AI for Wildfires and Reading Lips, Fun AI Facts about Fun Guys
23:16 Share

Hi, and welcome to Skynet today's Let's Talk AI podcast, where you can hear AI researchers discuss what's actually going on with AI. This is our latest Last Week in AI episode in which you get summaries and discussion of some of last week's most interesting AI news. I am Dr. Sharon Zhou. And I am Andrey Krenkov. And let's dive straight in. So first off, we have some articles related to research.

starting with harnessing the wild power of AI image generation. There's been some new research on how to get more control over what gets generated. So in the past with GANs, people kind of

have tweaked let's say maybe global parameters so you can control you know what class of object gets generated some properties of it and here in this paper they introduced a lost GAN which enables you to have more control in that you can basically input sort of

Bounty boxes that specify put a person here, put a tree here, and it generates the image occurring on me. And we've seen some of this sort of style before, but the results are pretty nice. Of course, it beats prior work. So it's cool to see Gantz making more progress. As someone who is a bit of an expert on this, what do you think, Sharon?

I think controllability and understanding and being able to know what results you'll get when you change something and shift something as the input is really important for generative models. And so this work definitely falls into that category of improving controllability such that

we know it's not going to just randomly spit out unpredictable things. It's going to follow what we want it to do and maybe shift things here, shift things there. Based on some of the samples, it looks like you can get actually a person generated that looks similar from another person. That

That is interesting. I feel like if we could get there, that could be extremely cool to just be able to shift a person slightly as opposed to generating something completely different each time. Because I think a lot of work has been geared towards that, whereas these minor shifts and being able to maintain the same overall object is really, really important for at least us people to care about.

Yeah, exactly. Just browsing the paper a bit, the title is Learning Layout and Style Reconfigurable GANs for Controllable Image Synthesis. So exactly what you're talking about. And then, yeah, they have this idea of instance sensitive and layout aware normalization, which

That means that they do get some control over the style of what gets generated in addition to what gets generated. So yeah, the paper is pretty neat. Lots of images, as you would expect, but also pretty interesting looking network images and yeah, pretty detailed comparison to prior work. So yeah, check it out if you think that's interesting.

Yeah, and on to our next article in research, the efforts to make text-based AI less racist and terrible, this article from Wired. So GBD3 and these large language models are, you know, spitting out some pretty bad stuff that's not constrained. And so now open AI researchers who, you know, did put out GBD3 are finding ways to curtail GBD3 and these large language models to curtail

be a little bit more appropriate. And what they're doing is they're feeding this program roughly a hundred encyclopedia like samples of writing by human professionals on topics like history and technology, but also abuse, violence and injustice. And, uh,

Yeah, so what they're doing is that they've, you know, first they highlighted GBD3 is talking about supremacy and superiority, having racist jokes, condones terrorism. I actually was playing with GBD3 last night and was able to get a lot of these inappropriate outputs, uh, uh,

both intentionally and unintentionally. So that's obviously an issue and this is a step forward in trying to constrain these models to be much more appropriate and thus be able to be useful for society.

Yeah, and it's pretty neat that in this paper that is a process for adopting language models to society with values targeted data sets. Not only do they present this result of tuning the model with some additional training data,

But also they have this whole process for collecting data sets that aim to kind of change specific behaviors with regards to different things you don't want the model to do.

And this follows on some research before on some issues with GPT-3, not just people playing around with it, but there's been papers like persistent anti-Muslim bias in large language models, which showcased that it is really quite bad. So yeah, obviously something that is a positive step. And given that it's very trying to commercialize GPT-3, something that's

quite necessary. And yeah, good on OpenAI for this and the researchers who wrote this paper. Definitely. And on to our next article in the application section of AI from Scientific American. This article is titled AI could spot wildfires faster than humans. So this is a great...

As the title probably implies, AI can now spot wildfires from satellite imagery and if anything looks out of place, can alert the system. This is really important for, and here this is from the University of Nevada, but this is really important for California right now. As we know, the fire season is really bad this year, even worse from past years.

And Sonoma's Fire and EMS Dispatch Center are looking into improving their dispatcher to basically use this AI system and to have this coordination between AI and human operators.

Yeah, exactly. It's interesting. We have this alert wildfire system and it's neat that this isn't just a research paper, but we actually are having people use it. And they've had software developers that added a data dashboard, audible alerts, and other features that refine the coordination between AI and human operators. And

Yeah, it's saying it's still sort of a work in progress. Sometimes it's not faster than humans, but obviously for something so dire like wildfires, this is quite a good step. And it also mentions that this is being deployed in Sonoma County and also similar detection, similar technologies being tested in New Mexico.

So yeah, very cool and an example of AI where it's probably less hyped, but is something that's getting developed and being deployed. And onto another example of something that's being developed, not just research, but actual systems. We have our next article, "Tech companies are training AI to read your lips by vice."

So there's a whole kind of, of course, history on this. And this article is about SRAVI, which is speech recognition app for voice impaired. And the idea here is, like the title implies, that you can train AI to read lips. And this is from the Irish startup Leopa.

And apparently this will likely be the first lipreading AI app available for public purchase. And yeah, so it's obviously pretty cool. This article mentions that there is a wide array of possible commercial applications such as silent communication apps and improved virtual assistant apps.

And yeah, here also it mentions that other companies like Google, Huawei, Samsung and Sony are all researching the so-called visual speech recognition systems, AI for reading lips. So yeah, pretty cool. And just another example of how AI can be used, hopefully in a positive sense and to improve and make people's lives easier.

Yeah, I mean, it definitely can help improve and make people's lives easier. Of course, there's also the issue with, you know, surveillance and that. Being able to see what people are saying just by looking at a camera. And I know there are laws around, you know, you can't record audio and video at the same time. And so this might be able to get around something like that is definitely one concern.

And so on to our articles around societal impact. The first one is titled The False Comfort of Human Oversight as an Antidote to AI Harm.

And this is from Slate. All right. All right. This is just the whole human in the loop solution where everyone say, you know what I can't handle. We'll just give it to the humans in the loop. And I think what this article is really trying to highlight is we can't just defer everything to that human in the loop and say that every part of AI harm will just be mitigated by some really competent human. And that's that's it. So I think

Yeah, like definitely, I think, true in a lot of senses where I definitely see people being like, oh, and then a human will handle that. Don't worry about it. And I definitely see that, especially in research and in some companies as well. In other places where it is applied, I also do see the humans really being key to making sure the whole system works.

But it is a matter of, you know, what job did the AI replace to now create these new jobs of human in the loop AI systems? I find that kind of interesting, this flow of employment essentially. Yeah.

Yeah, I like this article quite a bit. It's a pretty deep dive into the topic and a pretty good one. Here they outline a few specific reasons why they have this argument. So for instance, that calling for human oversight creates kind of a shallow protection that can easily be avoided by companies.

And then it's very difficult to accomplish in practice. And there's kind of a blurring of responsibility where, you know, if you make bad AI and you don't do your due diligence, well, if you have a human, maybe they'll get blamed instead. So, yeah, a very convincing case that at least it shouldn't be the only thing.

And I was a bit disappointed they don't suggest an alternative, but definitely a good take on this topic. And on to our next article on societal problems, we have China's tech workers pushed to limits by surveillance software. So...

This article here is kind of providing examples of ways in which surveillance software is making life more difficult. They have the story of Andy Wang, an IT engineer at a Shanghai-based gaming company. And he developed this piece of surveillance software called InSight.

I can't pronounce it. And so the system was installed on the laptop of every colleague at his company to track their screens in real time, record their chats, their browsing activity, and every document edit they made and would automatically flag suspicious behaviors such as visiting job search sites or video streaming platforms

platforms and would generate these efficiency reports weekly, summarizing the time spent by website and application. And also there were high-definition surveillance cameras that were installed around the floor, including in this office, and there would be checking. So yeah, pretty concrete example of how this looks like.

And something that you could definitely see companies wanting to adopt for better or for worse. So kind of a bummer article, but I suppose not surprising. And maybe we in the West need to really get ahead of this and get norms in place to avoid this sort of situation. What do you think, Sharon? Does this sound like something you'd want in your workplace? Yeah.

Uh, obviously no. Uh, uh, no, probably not. Uh, ideally, you know, there are things like this that could maybe help employees in some way. Um, but yeah, this, this does feel like putting pressure on people in, in,

Not a great way. And hopefully we can find a better way to do management that is not, you know, all just about surveillance and punishing people, but also thinking about, you know, maybe not necessarily knowing which individual employee is doing something wrong, but finding issues wrong with the whole system of management, you know, finding out, oh, maybe we should...

I don't know. This is a really dumb example. Just like maybe we should move the water cooler over there. Maybe like something's wrong with the, you know, something over there. That's why the employees are all acting or avoiding that space. I don't know. You know, like there there could be things like that where I think could be beneficial for companies to try to improve the workplace. But yeah, but for the each individual that it's not it's not great.

Yeah, yeah. I have seen examples where software is used to sort of find ways to improve teamwork, right? Communication, you know, maybe make more incentives. So there are positive examples where these sorts of things can be helpful, but these sorts of things where they track every website you visit and send weekly reports definitely would add to stress levels. And so...

You know, having worked at a tech company, you do have a lot of leeway to sort of browse the Internet, go on Reddit, go on YouTube. And you'd hope that it would be more about what you deliver at the end of the day and not these sorts of sort of dystopian oversights that would encourage you to avoid that.

And onto our last laugh article, a fun fact about fun guys. This is a Tumblr post that has basically used GPT-3 to put together, you know, a few fun facts about, you know, potential dating profiles since GPT-3 has seen many dating profiles and there are

some very funny traits. So likes to revisit the classics, believes in leaving false modesty at the door, keeps his nails and hair trimmed, dreams of buying a small island, loves pumpkin lattes, love, love, loves champagne, believes in destiny, is a self-described daredevil, loves his blue sleeping mask. Okay.

Okay, so these are some really funny ones. I don't know if you have any other funny ones that you want to hit here. Yeah, yeah, for sure. Yeah, this is a fun thing. This is from AIweirdness.com. We see this on the Tumblr website. And, you know, we actually interviewed Janelle Shane, who runs AIweirdness.com. So if you want to hear more about this sort of thing, just look back in our podcast feed and check that out.

But yeah, this is kind of the latest thing from that site and it's pretty fun. And she also highlights that the best fun facts were generated by the smallest GPT variant because, you know, there you get some very weird stuff. The stuff you mentioned was

fairly, you know, reasonable. I could see this being on websites. But here, you know, these other ones include things like doesn't want any of his exes in the picture, does not wear a vest, is a type of schoolyard bully, loves to whistle, has no issues with using markers.

and has a healthy butt. So yeah, it's pretty fun. And you know, there's a lot of these, there's like dozens. Another one is like was advertised as having French accented skills and a blue bunny. So yeah, pretty surreal, pretty fun. So yeah.

Another one is if someone gives a music note to him, the note gets the first three cups of coffee, which is definitely attention getting. So yeah, fun. I would say Google this fun facts about fun guys for more of this. It's pretty entertaining.

And that's it for this episode. If you've enjoyed our discussion of these stories, be sure to share and review the podcast. We'd appreciate it a lot. And now be sure to stick around for a few more minutes to get a quick summary of some other cool news stories from our very own newscaster, Daniel Bashir. Thanks, Andrea and Sharon. Now I'll go through a few other interesting stories that haven't been touched on. First off, we've heard plenty about deepfakes over the past few years, how they're becoming more and more realistic,

the ways they can be used and are used for misinformation and blackmail, and how easy they are becoming to create. As CNBC reports, the fight against deepfakes might just have taken a step forward. Researchers from Facebook and Michigan State University say they have developed software that can reveal where deepfakes have come from.

They claim the software can establish if a piece of media is or isn't a deepfake from a single video frame and identify the AI used to create the deepfake in the first place. It's known that deepfake detection is a bit of a cat and mouse game, but if these reports are true, there's good reason to be optimistic. Our next two stories concern business and applications.

According to VentureBeat, 2021 has witnessed a significant year-over-year increase in AI budgets at companies of all sizes. Respondents to the Appen's State of AI Reports survey reported budgets ranging from $500,000 to $5 million per year, a 55% increase over 2020.

The report also notes that decision makers are moving towards using AI to support internal processes, and that enterprises are moving AI responsibility out of the C-suite and into lower levels of their organizations. Our next story takes place on another continent. Among AI-powered companies in Asia, you've probably heard of the likes of Alibaba and Tencent.

But there's a new unicorn AI startup on the horizon you might not be aware of. According to Next Platform, Japan's Preferred Networks has already developed the world's most efficient high-performance supercomputer. The company has stayed under the radar, but the level of its investing shows that Japan is aiming to develop a homegrown processor for deep learning. And finally, two stories about AI and society.

Our first concerns a company you might not think about when you hear AI, and that's McDonald's. As The Register reports, McDonald's has been accused of breaking Illinois' biometric privacy law by collecting and processing customer voice recordings without their consent. In 2019, McDonald's acquired voice recognition company Apprenti to build a voice-controlled chatbot for its drive-thrus.

The lawsuit doesn't just accuse McDonald's of not seeking customers' consent, but also of processing audio samples to determine a speaker's age, gender, accent, nationality, and national origin. Finally, we return to Facebook. Our conversations online frequently spiral out of control, and not being face-to-face with our interlocutors doesn't help things much.

As CNN Business reports, the social network is testing the use of AI systems to detect fights in Facebook groups so administrators can step in to calm things down. Thanks so much for listening to this week's episode of Skynet Today's Let's Talk AI podcast. You can find the articles we discussed today and subscribe to our weekly newsletter with even more content at skynetoday.com.

Don't forget to subscribe to us wherever you get your podcasts and leave us a review if you like the show. Be sure to tune in when we return next week.