GPT-3 NPCs, Rebel AI Startups, A Cute AI Book
24:41 Share

Hello and welcome to Scanitoday's Let's Talk AI podcast, where you can hear from AI researchers about what's actually going on with AI and what is just clickbait headlines. This is our latest Last Week in AI episode, in which you get a quick digest of last week's AI news, as well as a bit of discussion between two AI researchers as to what we think about this news.

To start things off, we'll hand it off to Daniel Bashir to summarize what happened in AI last week. We'll be back in just a few minutes to dive deeper into these stories and give our takes. Hello, this is Daniel Bashir here with our weekly news summary. This week, we'll discuss the future of games, Google AI, government innovation, and the US's lead over China in AI.

One of the most interesting potential areas for future game development is the idea of artificially intelligent NPCs, or non-player characters that a player interacts with during a video game.

According to UploadVR, the developer of Modbox linked together Windows Speech Recognition, GPT-3, and Replica's Natural Speech Synthesis to create a demo of an artificially intelligent virtual character. Since Microsoft has exclusive rights to the source code and commercial use of GPT-3, this feature is unlikely to be added to Modbox itself.

But the video demo offers a glimpse into the future of interactive characters and the possibilities language models offer game design. Next, Google AI has continued to be mired in controversy after the firing of Timnit Gebru. According to Engadget, Gebru's Ethical AI co-lead Margaret Mitchell was recently locked out of her Google email account and investigated for downloading and sharing internal documents.

Mitchell had penned a public letter criticizing Google and its treatment of Gebru. On February 19th, Mitchell was officially fired. A Google representative told Engadget that after a review, Google had found multiple violations of their code of conduct and security policies. Alongside the firing, Google announced that Marion Croak would take over the reins of the ethical AI team, despite not having direct experience with AI development.

Google has pledged to make changes in its diversity efforts and has shuffled its AI teams, but the handling of Croak's placement inspired further controversy because it failed to communicate with the ethical AI team.

As a pandemic rages on, so does the US's unemployment crisis. As Brent Mitchell writes for State Tech, AI technologies might help states that continue to field thousands of phone calls and emails from their constituents. In one example, Illinois leveraged AI technologies to respond to and manage a flood of unemployment claims at the beginning of the pandemic.

The Illinois Department of Employment Security, or IDES, received over 500,000 unemployment claims in the span of one month. IDES augmented its contact center operations with an AI component that alleviated the stress on human agents by answering questions about applying for benefits and claim statuses.

Mitchell believes that states can effectively use the cloud and AI technologies to augment their existing systems without having to implement massive, costly changes. Furthermore, using AI technologies can help citizens get the information they need more quickly. Perhaps this crisis will also act as a forcing function for innovations in government operations.

Finally, as the Biden administration continues working on its agenda, one key foreign policy question will be about our relationship with China. Testifying on February 23rd before the Senate Armed Services Committee, Eric Schmidt, former chairman of Google, warned that the US is only one to two years ahead of China in developing AI, but needs to maintain a five to ten year advantage in both AI and other high technology fields like quantum computing.

Schmidt stated that the Department of Defense treats software as a low priority, but that China represents a competitive threat. Schmidt said even routine operations like minesweeping in the Navy demonstrate the value of AI, while noting China's extensive use of facial recognition and aggregation of health data.

Brad Smith, president of Microsoft, also expressed concerns about the weaknesses and vulnerabilities of the commercial and non-defense department digital infrastructure. That's all for this week's news roundup. Stay tuned for a more in-depth discussion of recent events. Thanks, Daniel, and welcome back, listeners. Now that you've had the summary of last week's news, feel free to stick around for a more laid-back discussion about this news by two AI researchers.

I am Andrey Karankov, a third-year PhD student at the Stanford Vision and Learning Lab. I focus mostly on learning algorithms for robotic manipulation in my research. And with me is my co-host. I'm Sharon, a fourth-year PhD student in the Machine Learning Group working with Andrew Ng. I do research on generative models as well as applying them to tackling the climate crisis and to medicine.

And Sharon, I guess we're just going to dive straight in to something we were just checking out before starting this recording. Our first article from Upload VR, a little hint as to what it's about, titled This OpenAI GPT-3 Powered Demo is a glimpse of NPCs in the future.

So basically, some developer of this software Modbox, which is software for multiplayer VR linked together Windows speech recognition, OpenAI GPT-3, which generates text based on the inputs and replicas natural speech synthesis for a very cool demo.

of basically allowing you to talk to virtual characters in a video game and have those characters reply powered by AI, powered by the very powerful capabilities of GPT-3 and modern speech synthesis.

And yeah, so the idea is usually in video games like, you know, Grand Theft Auto or Skyrim, you have a lot of these non-player characters around the world. But if you talk to them, they usually don't have much to say. Right. They have their prescriptive lines and not much more than that. Whereas in this demo, because they're using these tools,

You can start talking to any non-playable character and they can improvise a conversation and basically reply to anything you have to say. And here, because we're using all these tools, you could actually speak and have it be recognized and hear back for response, which is also not always the case in video games.

So yeah, we would recommend, again, to probably go to this link, this OpenAI GPT-3 Power Demo is a glimpse of NPCs in the future, and check out the video for a taste of what that's all about, because it is pretty cool to see. What do you think, Sharon? I know you just watched it. Aside from the latency, I think we're there. It seems very legit, and it seems like a very good place to start with something like this, because...

I can imagine that there might be sometimes a lack of consistency in terms of what the characters might say, but that's fine in a sense if that's just some kind of randomness that comes from the models. Obviously, there are still issues with things like bias, which I can imagine seeping in if, let's say, you just ask the question a little bit more differently, but it happens that it's...

I don't know, like women tend to use a certain type of word more. I don't know. Like, and that then makes the hot dog man not give you a hot dog or something like that. I'm not sure, you know, but there are still areas where things could become problematic. And so being able to monitor those is important.

But I think aside from latency, it seemed like you could have a conversation with these virtual characters and they seemed they definitely seemed real. I mean, you couldn't tell otherwise. I don't think at least I couldn't. So I think this is kind of exciting and kind of a good a good use case, to be honest. Yeah.

Yeah, and it follows up on one of the, let's say, seminal uses of AI for video games. If people don't know, there's a thing called the AI Dungeon, also powered by GPT. First GPT-2, now GPT-3, I believe, where...

You can be in a story, essentially. You can make up a story. Initially, it was sort of themed around fantasy, you know, Vastu name, AI dungeon. But really, it can be anything. It's a sort of improvisational thing where you enter text, there's kind of a narrator. So the AI makes up a story that weaves together your inputs. And yeah, that also is very...

curious to play with just because you can have any sort of story and often it's really ridiculous. But I think this demo is a pretty exciting step from that to actually being able to converse with

video game characters. And yeah, as you say, it seems like we're almost there. There's just a lot of latency due to all of these cloud based tools. But I could see this being in video games within, you know, a year or two, really, which is pretty revolutionary and kind of, yeah, very exciting to imagine. It will make games a lot more, I don't know, open ended and interesting. Yeah, definitely.

Have you had a chance to play around with AI Dungeon at all? Or have you heard of it? I definitely have. And I think I heard of it from you, actually. Yeah, I played with AI Dungeon. It was pretty fun. And I think GPT-3 is pretty good in terms of creative stuff and creative content.

Yeah, I think with GPT-3 and creative stuff, it's nice because it doesn't have to make sense necessarily. It can go off rails, but that can be entertaining when you're doing something like AI dungeon or talking to random NPCs. So in some sense, actually, I think this application is much better suited than many others people have been looking into. I think for generating websites, for...

doing legal stuff, a lot of possible applications of GPT-3, you would have issues with the spurious outputs and, I don't know, unreliability of it. But for games, you know, you don't care necessarily too much if it bugs out sometimes because the idea is to have fun and wacky stuff is also fun.

So, yeah, nice to start off with something a bit less serious this week. But on to our next story, which is back to our more usual flavor. We have from VentureBeat, Bend of AI Startups launch Rebel Alliance for inter-pairability.

And so the idea here is that more than 20 AI startups have banded together to create the AI Infrastructure Alliance in order to build a software and hardware stack for machine learning and adopt common standards. It brings together companies like Algorithmia, Determined AI, Y Labs, Pockyterm, various sort of data science, AI, data monitoring startups.

And apparently they've raised about 200 million from investors. So these are really significant startups.

And the idea here is that these startups are hoping to create this set of standards to compete against Facebook and Google and Amazon, which all have their own tool sets and cloud tools for various AI things. And this...

set of startups is basically saying that this is bad, that the big players are going to just lock people in and are arguing that this is needed. So interesting development. We've discussed Microsoft, Google, Amazon a lot in this podcast. It seems like they really dominate the news. So kind of nice to talk about some other players. What is your take here, Sharon?

Yeah, I think it's exciting because I think we need to find... We definitely need to find standards that the AI community can agree upon and that don't serve only one source to continue on their monopoly. So I think...

I think that's really important. Um, and I, uh, I'm excited to see this and it's interesting that it's coming from the startup world. I find that really interesting. It makes sense. It makes sense. Cause I think it's like the cloud people, they feel it the most in the sense of, you know, the cost of cloud compute, uh, and everything. And so I think, um,

I think this completely makes sense in terms of trying to make the package offerings much more standardized and streamlined for the end user and not necessarily for one company to just monopolize. Exactly. Yeah. And so it'll be interesting to see how this develops. Of course, this is just kind of the initial announcement of this. So, yeah.

Yeah, it'll be, I guess, a journey. And the alliance initially plans to focus on small partnerships between developers working on tools and frameworks, facilitating joint documentation, creating test software for integration, and

Yeah, so these sort of initial steps. And so I imagine it's going to take a while to catch up to these big players where it's been, you know, at this point, probably a decade for most of them to build up what they have. But it will be interesting to see this developing. Right, definitely.

Oh, yeah, yeah, yeah. Back to something more fun. Our next article is titled This Cute Book Explains AI to Children Without Scary Terminators. And so this is a new book written by author Matteo Loglio, who's the co-founder of the product and interaction studio OEO, O-I-O, that is, that aims to make AI accessible to children specifically.

The cute children's book is called Many Intelligences, and it's for kids six and up. And it's one that Loglio both wrote and illustrated. I kind of was hoping that GP3 might have helped out, but no, no, that might have ended in Terminators. And it walks through the, quote, gradient of human, animal, and robotic intelligence.

One interesting part of the book is that at the end of the book, at the conclusion,

the author kind of suggests that we'll give similar rights to robots that we give to human beings and that machines or robots might actually become smarter than us. And I think this is an interesting conclusion, which with this like kind of light aspect of the entire book and without those scary terminators, but just like coming to the realization of, you know, there are aspects of machines that are very intelligent and,

and might exceed us in some way. It's kind of hinting at AGI in a very light, subtle way. What are your thoughts on this, Andre? Yeah, this is pretty cute. The idea of a book aimed at children that explains the concept of intelligence and its different manifestations is kind of interesting.

But I suppose that there's many books on science for children that explain different scientific concepts. And so in that sense, it's maybe appropriate that there is one that covers intelligence, including artificial intelligence. And I don't know, I guess it makes me think that this speaks to ubiquity and the rise of AI. But at this point, we are getting so many tangential developments and applications and

talks and blog posts and books. And yeah, now there's a book for children. And I imagine it's not the first one or it's not the last one. We'll have more given how significant AI is now to the world. So yeah, pretty cute, pretty fun and does speak to some extent to the state of AI in general.

Yeah, definitely. It means that it's kind of gone more mainstream and that there's the fact that there is a children's book or something. Someone thinks that making a children's book is a good idea and pertinent to the future is yeah, this very telling of of where AI is today. Yep, exactly. And onto our last story from the technology of you, something a bit more serious.

NAI is training counselors to deal with teens in crisis.

So this is from the Trevor Project, which believes that 1.8 million LGBTQ youth in America seriously consider suicide each year. And they have 600 counselors for its chat-based services. And, you know, given that number, it's difficult for them to handle the need to help all these people.

And that's why this group, like an increasing number of mental health organizations, is turning to AI powered tools to help meet demand. It's, you know, the idea here is to use an AI system to help train people or help them be more efficient.

Here in particular, there's going to be AI powered training role play called the Crisis Contact Simulator and supported actually by Google. And there's also another project that organization developed. They also have a machine learning algorithm to help determine who is at the highest risk of danger. And this is on top of trialing other approaches.

And it turns out that AI was the most efficient for it. And the reason for this crisis contact simulator is that doing role plays with actual people is very time consuming, right? And limited also to when people are working. Even though a lot of counselors volunteer during night and we can shifts.

So the idea is to really train more counselors faster and better accommodate volunteer schedules, but also just to help train people in a natural way. And again, we've got GPT-2 as a backbone. So you're seeing

GPT turns up many times and it's kind of funny. We started with GPT, we were ending with GPT and it shows the growing number of developments again, as we've said.

So anything that you think is interesting or any thoughts on this, Sharon? Yeah, this seems like a really cool project. I would be obviously a little bit concerned about what GPD2 or any GPD variant would output, given the fact that it wasn't trained with guardrails. And if this is for counselor training, I mean, I guess the bot will be in crisis and...

can technically say whatever it wants to say, even if it's inappropriate. So maybe it's okay in this case because the counselors being trained should be ready for anything. So maybe in that sense, this is more okay for the model not to have guardrails, which is kind of interesting. Yeah, it's interesting here. It's noted that

There are some guardrails. It's designed to limit the potential for weird outputs. It's not discussed here how it's done.

Which is in contrast to, we discussed, I think a few weeks ago, this South Korean chatbot called Lee Luda that had really problematic outputs that were out there. But as you say here, because of the application of helping people train, it's not as bad as something that's just put out there in a while and it was supposed to be a conversational AI tool.

Whereas here, because it's a tool to help train people, it's maybe a bit more thoughtful. And even if it has bugs overall, if it's able to help people role play, that's all it really needs to do. Right, exactly. So I think it I'm excited to see this be more useful here as we obviously do need more counselors, especially during the pandemic with a rise in mental health issues. And so

Yeah, let's see where this goes. This might be a good direction for GPT. Exactly. And kind of nice to cap off this week with mostly positive stories. Yes, definitely. We too often have just negative developments in AI. But yeah, this week it's nothing but good stuff.

So with that, thank you so much for listening to this week's episode of Scana Today's Let's Talk AI podcast. You can find the articles we discussed here today and subscribe to our weekly newsletter with similar ones at scanatoday.com. Subscribe to us wherever you get your podcasts and don't forget to leave us a rating and review if you like the show. Be sure to tune in next week.