Everyday AI Podcast – An AI and ChatGPT Podcast - EP 167: When AI Outsmarts Humans - DeepMind's Historic Breakthrough

Episode Date: December 18, 2023

Some people think AI is as simple as autocomplete. But GenAI has gone from a working assistant to LLMs making new discoveries. We're talking about Google DeepMind's historic breakthrough and what it means for the future of GenAI and humans. Newsletter: Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan questions about Google DeepMindUpcoming Episodes: Check out the upcoming Everyday AI Livestream lineupWebsite: YourEverydayAI.comEmail The Show: [e-mail] with Jordan on LinkedInTimestamps:[00:01:30] Daily AI news[00:08:05] Google DeepMind's new discovery[00:12:30] First instance of AI outsmarting humans?[00:16:00] AI creating its own intelligence[00:22:00] Future of GenAITopics Covered in This Episode:1. DeepMind's Historic Breakthrough2. Evolution of AI Capabilities3. AI's Impact on Various Applications4. Future Implications of AIKeywords:AI, generative AI, large language models, GPT 4.5, virtual rally, AI technology, former prime minister, AI development, self-evolving AI technology, AI models, Everyday AI show, DeepMind, AI breakthrough, autocomplete, new discoveries, knowledge discovery,  Fun search, mathematical equation, outsmarting humans, accidental discoveries, head of content marketing, AI business strategy, higher education, DID, hot takes, AI in 2024, unsolved math problem, cap set, MIT, University of California, AI Zip IncorporatedSend Everyday AI and Jordan a text message. (We can't reply back unless you leave contact info) Start Here ▶️Not sure where to start when it comes to AI? Start with our Start Here Series. You can listen to the first drop -- Episode 691 -- or get free access to our Inner Cricle community and all episodes: StartHereSeries.com Also, here's a link to the entire series on a Spotify playlist. 

This is the Everyday AI Show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. Meet Firefly AI Assistant, now live in Adobe Firefly, the all-in-one creative AI studio. Just describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome. The assistant accelerates execution. People think that AI is just an auto-complete.

And that's kind of true. But AI is actually so much more than that, especially when we're talking about generative AI in large language models. So in just one short year and a half, we've gone from thinking of generative AI and large language models to this chat GPT craze to now large language models, making new discoveries that have plagued humans for decades. So we're going to be talking about Deep Mind's historic breakthrough

and talking about what does it mean now when AI starts to outsmart humans. We're going to talk about that and a lot more today on Everyday AI. Thanks for joining us. My name's Jordan Wilson. I am your host. And if you're new here to Everyday AI, welcome. It's a daily live stream podcast and free daily newsletter, helping everyday people like you and me,

not just learn what's going on in the world of generative AI, but how we can all actually understand it and use it, right? That's the most important thing is how can we actually use it to grow our companies, to grow our careers? So that's what we're going to talk about today. But before we get into this historic breakthrough at DeepMind, let's, as we always do, go, let's talk about what's going on in the world of AI news. So first, more AI on the campaign trail,

and this time it's coming from Pakistan. So artificial intelligence was used to create a virtual reality for Pakistan's former Prime Minister Imran Khan. So why does that matter? Well, Khan is currently in jail and is unable to campaign in person. So this rally was organized by Khan's party, PTI, and it garnered millions of views on social media, leveraging AI technology to replicate Khan's voice and speeches.

So this is a new one. right we've talked about at least here in the u.s with the uh the 2024 election cycle ramping up we've talked about it being used in ads but here's a new one use live in a campaign rally AI to replicate his voice and um likeness so interesting all right number two and this is a hot topic for today in AI news is gpt 4.5 actually out well maybe well let's talk about So we talked about this last week on the Everyday AI show and CEO Sam Altman essentially just denied it on Twitter after there were rumors and reports and leaked screenshots of OpenAI's next version of their large language model, GBT 4.5 being a thing, being a reality, right? So there was a screenshot. People were like, ah, it's not real.

The CEO, Sam Altman said, nope, it's not real. But there's some updates. So over the weekend, a lot of users, including myself, notice that if you ask, ChatGPT a certain way, what model it is specifically running, it will say GBT4.5 turbo. However, it's important to note that asking any large language model, the version of what it's using is not always very accurate at all. But this is kind of a widespread response that a lot of people are getting. So you're going to be hearing a lot of buzz this week. Okay, are we on GPD4 or are we on GBT4?

Also, what's important to see is there's been some under the hood updates that. that no one else has talked about or reported. So I talked about on the Everyday AI show, this is probably about two weeks ago, on how the plugins mode, the GBT4 version in plugins mode, had actually been rolled back to January 2020. It's knowledge cutoff, which is important

because now, as of since ChatGBT has been responding, it's on 4.5, that has been rolled up to April 20203. So the GPD4 that's being used in plugins mode was kind of, you know, caught off at the knees for about a month and it's been updated again. So is that a sign that yes, GPT 4.5 is out. I'm not sure. But it's worth taking a look at, especially after Google released its Gemini.

Its new update to Bard, kind of it's under the hood releases with Gemini Pro, now powering Google Bard. So all these big companies, they're obviously going back and forth trying to one-up each other and trying to make sure that their stocks continue to go up. So keep an eye on that. All right, last but not least in the world of AI news, AI can now create more AI with what's being called a self-evolving AI, right? All right, so researchers have developed a new technology

that allows artificial intelligence models to create smaller AI systems without human intervention, all right, showcasing the potential for self-evolving AI. So this breakthrough could have applications in improving things like hearing aids, monitoring pipelines, attacking endangered species. I mean, there's so many different potential use cases for this. So in short, larger AI models, as this report shows, can now design smaller, more specific AI applications for everyday use. So this study comes from several companies and universities that include the group AI Zip Inc, as well as researchers from MIT and multiple University

of California campuses. So a lot of different researchers and private company being involved in this. So this is something you're going to be hearing a lot about, especially as we talk about, you know, moving into 2024, what we should be expecting or could be expecting from generative AI and what's happening in the space or what could happen. Kind of wild, right? We've talked about this for years, right? Is this the first sign of, you know, superintelligence when AI models can create different versions of themselves? But it looks like, hey, it's 2023 and we're already there. So before we talk about this historic breakthrough from DeepMind, let's just go over what we have coming up this week in everyday AI. So tomorrow is going to be a good one.

So we're going to be talking about AI in higher education. Is it broken and how to fix it? So make sure to tune into that one tomorrow with Jason Gula, who's the chair of the artificial intelligence. Speaking of colleges in Berkeley, he's the chair of. the Artificial Intelligence Council at Berkeley. Wednesday, if you haven't heard of DID, it is one of the most impressive and unique generative AI tools. So we'll be talking with Rod Freeman, who's the head of content and creative marketing at DID. So that should be exciting, right? And hey, where else, aside from everyday AI, can you tune in live and talk to the actual

people who are shaping generative AI and the tools that we use? So make sure to tune in on Wednesday. on Thursday, we're going to be talking with Marcus Bernhardt about AI business strategy, and if you're ready for 2024. And then I'm going to follow that up the day after Friday solo show. So many hot takes. You're going to need oven mitts, all right, to touch the computer. I'm coming in hot and talking about generative AI in 2024, what's coming and what it means for you. All right.

So if you haven't already, please make sure to go to your everyday AI.com. Sign up for that free daily newsletter. But let's talk about this, right? Let's talk about the reason for, or the topic for today's show. And thank you to everyone who's joining us from all over the country and technically all over the world. So Tara's joining us from her family road trip. That's great.

Thanks for tuning in. Josh, what's going on from joining us from Dallas? Ellington. Brian, thank you, everyone, for tuning in. Also, Harvey from Texas. We have a lot of people from Texas listening. All right, but let's let's talk about this, this new deep mind discovery. Let me first talk about what it is and then we're going to talk about what it means.

And this is pretty historic. All right. So here's what's going on. So this just came out late last week. We talked about it on the show and in the newsletter, but I thought this was worth its own kind of deep dive to talk about this. So there's this, I think, very common misconception that all AI is is an autocomplete, right? So whether we're talking about generative AI, large language models, traditional AI, I think people assume that all AI does is it just, it's a very advanced auto complete.

It's not creating anything new. It's not creating new intelligence. So that's wrong because, I mean, as this new breakthrough from Google's, Google's AI arm, if you don't know, is called DeepMind, but they have a language model called FunSearch. And FunSearch and the DeepMind team have successfully solved a famous problem in peer mathematics using AI to make valuable in previously unknown discoveries. So here's what happened. They used their model FunSearch to solve what has been an unsolved math problem for decades

called Capset. All right. So I'm not a mathematician, and maybe someone tuning in today is and can enlighten this all, but from what I can read and what I can understand, Capset is essentially a famous math problem

or a math question, right? That no one has been able to solve for decades. And it's finding the largest set of points in a specific 3D space without three points forming a straight line. So if you can visualize that, I mean, hopefully that makes sense. But essentially, you've had many, many researchers, mathematicians, and even artificial intelligence, with the help of artificial intelligence, people have been trying to solve

this or understand this cap set problem for decades, right? And DeepMind did it by using generative AI, using their language model for. fund search. All right. So why is this important? Well, it is a new discovery, right? So this was not technically derived from existing data. So yes, this math problem and the data that it technically contains is the parameters. But this new discovery, this is a language model, right, creating a new math, like mathematical discovery. All right. So to do this, it combined the large language model with other systems.

And how it did is it used it to reject incorrect answers and generate the correct code. Okay. And this is one of those things that I think people misunderstand or maybe when you're looking at large language models or generate AI and what it can do for yourself or your company. I think people sell large language model short. Just because of the sheer volume in quantity that a large language model can work at, you know, much faster than any group of humans. So, you know, this problem, right, this cap set problem, it obviously has a solution, right?

But like I said, it has perplexed humans for decades. And I think this is one of the first definitive and somewhat easy to understand instances where a language model has created a new discovery, right? It has created something that was not there before. This is an autocomplete, right? So this problem, like I said, it has its own rules for solving. It has its own kind of confinement that the solution must. work within. However, Google DeepMind and their fun search model has done this. Right. So let me, let me illustrate. Let me illustrate this. And I'm also wondering from our,

from our live audience, if you've noticed this or what are your thoughts, right? And I mean, do you agree? Do you agree? Is this the first instance where AI is outsmarting humans? I mean, there's arguments that, you know, that it's happened before. But this is one of those things where you can look at it and say, this might be one of those, kind of like what Dr. Harvey Castro is mentioning here about the AlphaGo, right? The AlphaGo move from so long ago, right, when you had AI initially competing in different games, you know, like AlphaGo against the best human players in the world. So this is something, I think, even greater than that, right? Because in head-to-head human competitions, right, when AI, you know, there's the famous Watson, you know, winning on

jeopardy against human contestants, you know, so that was, I think, a lot of people's first interaction with artificial intelligence, seeing it live, you know, on their, on their television sets, you know, many years ago. However, that was always AI system versus human, right? So in theory, you could point at, okay, well, this is just the AI was just a little better than the human today or, okay, well, the human today was a little off, right? A lot of pressure for humans, right? People made those arguments, you know, when these, you know, kind of AI versus human competitions or, you know, events happened, you know, decades ago. But this is completely different, right? This cap set math problem and Google's

deep minds new discovery is completely different about that because it's not pitting, you know, deep mind against a single human. It's not doing something live. It's not, you know, setting down, you know, Google's large language model or their fun search model versus a human. This is against the history of humankind, right? Countless mathematicians, scientists, researchers have tried to solve this cap set problem for decades to no avail. And Google DeepMind did it. So I kind of want to talk a little bit, just a little analogy here, right? And hopefully this can kind of set our expectation or reset our mind.

When we look at generative AI, when we look at large sandwich models, when we look at what they mean and also what they're capable of. Because I think, you know, more about this Friday, so make sure you tune into that show. I think that we're going to see a completely different side of generative AI moving into 2024. I think a lot of companies, you know, enterprise and even tech giants, right, have barely put their toes into their collective toes into artificial intelligence, right? I think 2024 is going to be wild, but that's beside the point. But I want to illustrate what I think can hopefully help us better understand or get over what I think is a misconception that AI. I cannot create its own intelligence, right? I mean, one of the new stories today says otherwise, right?

So this company AI Zip incorporated and the researchers from MIT and University of California campuses have shown now that large AI models can create smaller AI models to solve problems. I mean, if that doesn't tell you right now that AI, AI has the capability to outperform humans and that AI is more than just an autocomplete. And AI can do a lot more than just work with data that it is given. If that alone doesn't tell you, right, just the fact that researchers are now showing that large language models can create their own AI models, right? This self-evolving AI system or self-solving AI system.

If that doesn't show you, let's take a little. exploration together, right? So think of space exploration, right? So for about, oh, yeah, it's been more than 60 years, right, since the 50s, since humans were physically exploring space, right, in manned aircrafts, unmanned aircrafts, it's been more than 60 years. So you would think, right, when there's billions of dollars of research going into this and humans have been tackling space exploration for decades, right? Adobe just introduced an entirely new way to create, bringing the power and precision of its creative suite into one conversational experience.

Meet Firefly AI Assistant, now live in the Adobe Firefly app, the all-in-one creative AI studio. Powered by Adobe's Creative Agent, Firefly AI Assistant lets you start with your vision, just describe what you want, and shape the outcome as it takes form with the assistant. The assistant orchestrates multi-step workflows, drawing on 60 plus pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premiere, Lightroom Express, and more to help bring your ideas to life. You can also get started with creative skills, a growing library of pre-built workflows for common creative tasks, like batch editing photos,

creating mood boards, portrait retouching, and creating social variations. Every step the assistant takes is visible so you can refine, redirect, or take over at any time. You stay in the driver's seat as the creative director. Adobe Firefly AI assistant now in public beta. See it today at firefly.adopi.com. That's what I want to set as the baseline for thinking of generative AI. And here's what I mean. I think generative AI in large language models are space explorers, right, in this scenario.

Humans have much, and I know this sounds weird to even say out loud. Humans are limited compared to large language models compared to generative AI, right? You need large teams of people. You need large teams of researchers. You need large teams of funding to explore space, right? Whereas, think of generative AI. Think of large language models. the level at which generative AI can compute is unfathorable to the human mind.

Right? Because there was the other study that we talked about. This was a couple of weeks ago. About just the, this was also Google and Deep Mind. So they said that with one of their new discoveries that they unlocked 800 years worth of knowledge after discovering 2.2 million new crystals, right? So 800 years of knowledge. So that is the difference, right? And if we take this analogy of space exploration, right, what large language models and generative AI can accomplish even just like by using this

this same methodology that Google DeepMind used with fun search, right, which was essentially testing all of, you know, creating a system to quickly at scale, reject incorrect answers. Even just by doing that, by being able to reject incorrect answers at thousands times the scale of what humans can do, that leads to new discoveries. Right. So in this example that I'm trying to illustrate where a large language model or generative AI is a space explorer. We know those stars, right?

We have pictures of the galaxy from satellites and NASA and all these things, right? But large language models more quickly than humans can explore things at the rate of hundreds of years at a time when humans cannot. That's the difference. So large language models in generative AI by using existing data can actually. create new intelligence. It can create new discoveries that can benefit humans that humans may otherwise never be able to come to on their own because it would take too long, right? That is, I think, one of the most powerful things about large language models, about generative AI, that people just look over. People think, oh, you know, large language models, they write,

You know, they make my C plus email to my boss into a minus or a B plus email, right? They think that's what a large language model is. And yes, obviously, they can do that. But when we talk about all of this new technology that we're going to be seeing in 2024, it is much more than that. I think we are going to be talking about on a weekly or maybe daily basis in a year or so about generative AI, about large language. models making new discoveries nonstop.

Right. So yeah, over the last year, we've heard about a handful of examples. I think we are literally going to be seeing this almost on a daily basis where large language models, generative AI systems are going to be outsmarting humans daily. Tackling societal issues, problems, equations, right, that have plagued humans for decades. I think this is going to become the new reality. Yes, Douglas. It's the elephant in the room. Yes. Are we moving toward this, this Terminator type world? I mean, kind of, right? Kind of. You know, you even have the company in Asia that is actually called Skynet and creating this technology, right? So yeah.

That's kind of where we're getting where large AIs are creating on their own smaller AIs. When we have artificial intelligence and large language models making new discoveries on their own. And then when they're let's let's think about that. What happens when you combine these two kind of instances, right? Where, oh, AI is now making discoveries on its own. But then what happens when AI can create an AI system or an AI. AI model to do something with this new intelligence that it discovered, right? Yeah, it's, it's hard for me to wrap my brain around. Yes, I talk about AI every single day, but that's

hard for me to even fathom. But that is the future of where we're heading, right? When AI is outsmarting humans pretty easily, right? Cecilia, I love this comment here and thanks for joining. And hey, if you listen on the podcast, check your show notes. You know, we do this live every day. You can come. You can ask questions when we have guests. But Cecilia, I love what Cecilia is saying here. I'm wondering whether sometimes discoveries against the history of humankind is complete history. Is it possible that someone or somebody who is or are not part of the circle of those trying to solve a problem may have solved it and their interactions on it? Yeah. That's great, right?

Because, yeah, I think we're also going to see a lot of accidental discoveries or new intelligence that is created when maybe that was not the outcome, right? So many of the greatest products that we use, services, we're not created ultimately for what we're using them for. So Cecilia, absolutely. I think that we are going to see researchers and generative AI and large language models solving, you know, not just problems that we didn't know existed, but also creating new intelligence in areas where it was not even intended to create intelligence on.

Absolutely. So I hope today's episode is helpful. I'm not going to continue to talk for 45 minutes, but let's just quickly, let's just quickly recap what we're talking about here with Google DeepMind's historic. breakthrough. All right. So by using their language model fund search, they successfully solved this mathematical equation that has been eluding humans for decades by solving this math problem cap set. And why this is significant. And I think this news came out Thursday or Friday, and I haven't seen a lot of people talking about this, which is interesting.

But this is significant because this is one of the first times where definitively we can say that an AI has outsmarted humans and has created a form of new intelligence, right? And this fun search and what Google DeepMind is doing is indicative of something so much more than just, yes, they crossed off. quicker or more quickly cross off all of the incorrect answers to this math problem to more quickly get the correct answer. Yes, that is in theory how they went about this process. But it signals something so much greater than that. It signals the enormous potential for large language models and for generative AI moving forward. And I hope that this also opens up your eyes to seeing that large language models are so much more than just writing some content. There's so much more than, you know, creating a marketing plan. Large language models and generative AI as we head into

24 are creating their own intelligence. I don't know what it means. But we're obviously going to find out together because we do this every day. And as more and more of these breakthroughs come, you better believe we're not only going to be talking about them. We're going to be bringing experts on. for you all to ask questions and to understand. But I will tell you this. If you have been on the fence, right, about the powers or capabilities of generative AI, if you think that generative AI and large language models are

nothing more than an auto-complete working off of a fixed data set, I think you have to open up your mind to the possibility that generative AI in large language models can be so much more than that, right? Yes, in theory, they are trained off of a data set, but they're creating new discoveries. They are creating new intelligence. And I think that's what's important to know. What's also important is for you to go to your everyday AI.com. Sign up for the free daily newsletter.

And like I said, we have some great shows, some great shows coming. up the rest of the week that I hope you'll join us. So as a reminder, tomorrow, we're going to be talking about AI and higher education. This is going to be good. I can't wait for this show. Wednesday, we're going to be talking with Ron Friedman, the head of content marketing from D.D. One of the largest generative AI tools out there, bar none. Then we're going to be talking about AI business strategy in 2024 and my hot takes for 2024 as well. So thank you for joining us. Appreciate you. I hope to see you back tomorrow and every day with more everyday AI. Thanks y'all. Meet Firefly AI assistant. Now live in Adobe Firefly, the Allman One Creative

AI studio. Just describe what you want to create in your own words and the assistant handles the rest, orchestrating multi-step workflows across Adobe Creative Cloud apps, including Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome while the assistant accelerates execution. Stand control with the ability to step in and refine at any time. See it today at firefly.adobie.com. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit Your EverydayAI.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.