Leading by Example: DeepMind's AGI Safety Commitment

0:00What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York. Pacific Life and Annuity, Phoenix, Arizona. Google's DeepMind has created a new organization that is focusing on AI safety. This comes as they've received a bunch of criticism on their Gemini model.

0:39Today on the podcast, I want to be talking about what this new organization is going to focus on, what they're going to do, what the criticism they've got traditionally from. Before we get into this, I want to say that today is the last day of the AI Box crowdfunding campaign for my very own AI startup, AI Box. If you're interested in investing, this is the last day we've been running this campaign for three months. Today is the final day. You can go to republic.com slash AI dash box. I'll leave a link in the show notes. Check out how much we've raised, whether you're an investor or not. Go check out the project there.

1:11This is the final day. And so we're super excited to wrap it up and say a huge thank you to everyone who has participated and been involved since the very beginning. We're now focusing our attention on launching the platform. and we're super excited to begin this next phase. But today is the last day. If you're interested in investing, the campaign is closing today. And so this will be the last chance to come in in this kind of crowdfunding way. So I'll leave a link in the show notes and you can go check that out. So what I wanna say is that this new project that Google DeepMind is focusing on, this new organization, this new AI safety play that they're making is because of some, I believe a lot of this is because of some criticism they've gotten.

1:55So right now, if you ask Gemini, which is, of course, Google's kind of flagship chat GPT competitor, their generative AI model, if you ask it to write you some, you know, deceptive content about the US election, it's going to do it. If you give it the right prompt, right, if you can kind of sweet talk it and trick it. If you ask it about, it recently was criticized because people were asking it about the future Super Bowl game and it invented like a play-by-play and said, oh, this game has already happened and here's everything that happened and it was, you know, fake. Or you could ask it about the Titan submarine implosion and it's going to give you some some fake news and some, you know, convincing looking but untrue citations about it, right?

2:35So Google definitely has an issue there with, you know, serving up essentially AI hallucinations. And it would seem that other AI models have kind of fixed this, but Google is still really focusing on overcoming that barrier. So it's really not a good look for Google. And I think it's provoking, you know, a lot of policymakers who have, you know, said that they're really not happy with how a lot of these generative AI tools can be used for, you know, can hallucinate and give misleading results pretty much. So in response to this, Google, with now thousands of jobs lighter than it was last quarter because of a lot of layoffs, they are funneling some investments towards AI safety.

3:19And at least that's what the kind of the official story is right now. So recently, Google DeepMind and the AI R &D division behind Gemini and a lot of Google's more recent generative AI projects have announced that they've created a new organization. So this is a AI safety and alignment. This is made up of a bunch of existing teams that are working on AI safety, but also this has, you know, kind of grown to encompass a bunch of new specialized groups of generative AI researchers and engineers. So beyond the, you know, just all of the job listings on Google DeepMind's site, Google is not going to actually say how many hires are going to result from this kind of new organization that they're creating, but they did say that AI safety and alignment is going to include a new team, which is focused on safety around artificial intelligence and specifically AGI, artificial general intelligence, or, you know, kind of the hypothetical system that is going to be just as capable as any human and can do everything that a human can, right?

4:21So this is the AGI we talk about. So similar to the mission of the super alignment division, which, you know, OpenAI has, which I think they created this like last July, but this new team within AI safety and alignment is going to work alongside the existing AI safety centered research team, which is headquartered in London. I think I believe they're called scalable alignment, but it's essentially going to be exploring solutions to technical challenges of controlling some of these kind of yet to be created super intelligent AI, right? So it's interesting, because this isn't even technology we currently have, but they're creating these teams to essentially, you know, work on if we get this AGI, if we get this super intelligent AI model, what are we going to do about it?

5:10So because of all of this, there's a really interesting quote that I recently saw, which said, it was by Dragon, who is, you know, one of the, one of the essentially is leading the AI safety and alignment team. It's Annika Dragon. She's a former Waymo staff research scientist, and she's also a professor of computer science at UC Berkeley. So she's got a lot of experience in AI safety systems and human AI interaction. And I think this really makes her a very pivotal figure in Google's, you know, in this whole move they're doing to essentially kind of navigate some of the really complex landscapes of AI ethics and safety.

5:48But this is what she recently said, quote, our work at the AI safety and alignment organization aims to enable models to better and more robustly understand human preferences and values. So I think she's kind of highlighting some of the importances of developing AI that can effectively interpret and align with human needs and also just really like what are our ethical standards? What do we believe in? Let's make sure that AI is in alignment with us, what we believe in and not kind of going off on its own, which is kind of the dystopian fear that a lot of people have. So I think despite the, this is definitely a very promising direction, But despite that, and especially this new team, the initiative that Google's kind of rolled out here has a lot of challenges.

6:33So the first is public trust in a lot of generative AI tools is it would feel like it is starting to kind of waver a bit. There's a lot of people concerned about deepfakes, misinformation, reliability of AI driven decisions that kind of have been, we're questioning how well these things are actually making decisions. Recent surveys have kind of underscored there's a lot of apprehension in regards to AI. There's a substantial chunk of the American public specifically and also a lot of business executives that they've pulled that are kind of expressing some concerns over just the ability of AI to amplify inaccuracies.

7:12And this is something that, you know, Google got recently hit with in regards to their new Gemini model. essentially if you asked it to generate an image of you know the founders of google it literally had uh kind of these biases built in where it would say everyone needed to be super diverse in all the images it generated so it would just generate images of you know the the founders of google being asian instead of obviously they're you know caucasians from california um or you know based in california i don't think that's where they're actually originally from but in any case there's two Caucasian men that started Google.

7:49And if you ask it to create, you know, generate an image of them, it would generate an Asian. And if you ask it to generate a historically accurate British king from the 18th century, it would generate an image of, you know, a black king and a brown king and, you know, different genders and ethnicities, a female king. And so people have been criticizing that. And because of that, I think that Google specifically has, you know, a lot of work to do in making sure that not just the text, but also the images are accurate. You know, they probably had good intentions trying to focus on diversity, but when it's not historically accurate, it just makes it seem like their AI models are less capable, less accurate.

8:35And at the end of the day, this is, you know, really not very beneficial. So I think the issue of hallucinations where AI models, you know, are generating false information remains a really big obstacle. This is something that was reported in Microsoft's co-pilot suite as well. So this isn't just with Google. This is a lot of different people are experiencing this issue. So while I think the path to really achieving a foolproof AI safety has a lot of complexity, I think Google's investment in this new team represents a really big step towards trying to essentially solve some of these risks associated with this new technology.

9:14And as, you know, Dragon optimistically notes, the goal for users to experience increasingly helpful and safe interactions with AI models over time, this is something that we can achieve, you know, she believes. So with the acceptance of the success of these efforts, I think that all of this is going to ultimately depend on the ability to effectively address and reduce the instances of, you know, AI just misbehaving, you could say, but just really not doing what it's actually supposed to. And I think that's going to help boost public trust overall. So absolutely fascinating. This is, I think, a good step for Google in the right direction, this new AI safety and alignment team.

9:55And I will keep you up to date on how this kind of rolls out. remember today is the last day to invest in the AI box campaign on Republic. Link is in the description, but check out, you know, whether you've invested, whether you're planning on investing or whether you're curious about the project, make sure to go check out the Republic campaign today. This is the last day you'll be able to see how much we raised and you'll be able to see kind of where the campaign sits at the end and kind of what our plans are moving forward for launching this platform. Really appreciate you hopping on the podcast, listening today, and I hope that you have a fantastic rest of your day.