44 | Google releases Gemini, and Meta, IBM, AMD, Intel and other giants form an AI alliance, and more AI news from this week

0:00Hello, and welcome to a short weekend news episode of the Leveraging AI podcast. This is Isar Maitis, your host, and we are going to dive into some of the big and interesting news that happened in the AI world this week, and I'm going to provide my thoughts on what I think about it.

0:22The first big piece of news is that Google finally announced Gemini. Now, those of you who've been listening to the show knows that last week, Google announced that they're delaying the release of Gemini at least until Q1 of next year, which is kind of surprising that they actually decided to release it a few days after that, but that's what they did. So Gemini, which is Google's new core model that is a complete multi-modal from the ground up, meaning it knows how to handle video and images and voice and text, et cetera, all built into one model. It has released Gemini in three different levels.

1:00One is called Gemini Nano, which is built in order to run on their mobile phones, and it's going to become available on Pixel 8 with a lot of really cool features running straight on the mobile device. Gemini Pro, which is the model that is available now on BARD, starting presumably right now, and Gemini Ultra, which is their highest end model that they're not really releasing yet. and it's going to be built and baked into more and more tools from Google as we get into 2024. Presumably, Gemini Pro is going to be roughly at the level of GPT 3.5, a Gemini Ultra that, as I said, is not available for testing yet, is going to be roughly on par with GPT 4 that is already available from OpenAI.

1:51Now, before we get a chance to actually play with it and test it and compare it to how OpenAI's GPT 3.5 and 4 works, I will say two things that are important to think about. Number one is that GPT-4 has been released in April, and it's been in testing since seven months before that. So GPT-4 technology has existed for more than a year. That means two things. A, it's more mature, which I'll touch on in a second. But B, it's very, very likely that OpenAI has been working very hard on GPT-5 since, which means the fact that Google is releasing this now and even not completely releasing, but releasing in stages because their ultra model is not really available yet is a little disappointing based on the fact that OpenAI probably has much more advanced technology available right now.

2:46And we don't know exactly when they will release GPT-5. We already know it's going to be multimodal just like Gemini. but I think they will open the gap from Google again as soon as GPT-5 comes out. So that's number one. Number two is the fact that Gemini Ultra is beating GPT-4 on some computer AI benchmarks does not mean it will actually work better in the real world because GPT-4 has been used by actual people for over six months now. Open AI has 100 million weekly users. I assume, let's say, 10 to 15 % of them pay the$20. So that's 10 to 15 million people that use GPT-4 weekly, which is a huge amount of data for Open AI to continue improving the actual day-to-day usage of GPT-4 versus what it performs on Benchmark.

3:41So I think that even once Gemini Ultra comes out, GPT-4 will still be better than it in the day-to-day usage until Google has enough actual user data to fine-tune it to be better on what people actually do with it versus what the benchmarks do with it. And speaking of the Gemini release, they released some really cool videos of a person drawing a duck on a piece of paper just by scribbling it and talking to the Gemini AI as he is scribbling it, which is incredibly impressive. but apparently it's a fake video, meaning that interaction happened over text with the bot and not in real time watching the video as the person is drawing the images.

4:26And Google quote unquote forgot to mention that when they shared the videos. I don't think that was a smart decision by Google doing it this way. I think it was very obvious that they will get caught and will just get some bad press. I think the capabilities are still really impressive even without faking the real-time video communication. That being said, Google is a huge company that has been doing AI related development for a very long time. They probably have the best AI lab as far as people and capabilities and has unlimited resources and unlimited access to data, including data from their search queries, all the websites that they're crawling, YouTube, phones, locations, and all of that.

5:11So their ability to create really incredible AI models definitely exist. And even if this one is a little disappointing, at least now and on paper, I have zero doubt that they will play a very major role in this new era of AI. And speaking about Google, this week, DeepMind, which is a research lab from Google, released two very interesting pieces of information. One is that they were able to develop a methodology to allow AI to learn by looking at quote-unquote experts, meaning instead of training it on specific data, it is looking at what another model is doing or looking at what people are doing and learning from that just like humans do learn by looking at other people.

5:57So expert quote-unquote agents showed the AI model the optimum path for performing specific tasks. And then the non-expert AI was able to copy and remember the optimal process and hence learn a lot faster. And they were able to show that this methodology allowed it to learn faster, multiple effective skills across different environments. What that means is it demonstrates the capability of AI to learn any skill by looking at others that are performing it, which will also include social skills and cultural aspects of communications with other people, which on one perspective makes it even more powerful.

6:38On the other hand, it's also scarier because it will be able to mimic everything that we do really well, really fast, with most likely a lot less investment in training models. And another interesting piece of news from DeepMind, just to show the capabilities of AI overall and not just generative AI, DeepMind came out with a new material discovery process called Genome. But what it does, it allows to predict the stability of different inorganic crystal materials. Now, that sounds very, very techie. But to put things in perspective, to develop new materials takes a huge amount of time and trial and error in order to evaluate whether these new materials that people are developing for whatever purposes are going to be stable and actually usable or not.

7:23And using this artificial intelligence model, this model was able to evaluate 2.2 million potential new materials, which is equal to 800 years of human research and discovery. Out of those, 380 ,000 materials show promise of actual performing better capabilities in the real world, like building better batteries and superconductors. Now, they gave that information into an actual third-party lab that was able to make 41 new materials guided by the AI, which improved the efficiency of development of these materials by 80%. So the opportunity of us, the human race, to develop new materials that can improve our lives, reduce carbon footprint, et cetera, et cetera, is at our fingertips right now using these AI technologies.

8:20Obviously, the same thing can be used for negative sides because you can now develop materials that do harm one way or another in a much faster speed than you could before in a lab. But this is the world we live in, and this is the promise of AI capabilities beyond just generative AI. Another huge piece of news from multiple giants comes from a partnership between Meta, IBM, AMD, Intel, Oracle, Stability AI, and several different universities and nonprofits to form an AI alliance. And the goal of this alliance is to share technology and boost open source models versus the proprietary ones currently coming out of OpenAI, Anthropic, Google, etc.

9:05Their idea is to focus on ethical, safe tools and share them with the world. And they believe that by sharing it as open source, it will drive additional innovation and will increase safety because everything will be out in the open and nobody will be able to hide anything. thing, this is exactly the opposite approach of what you hear from OpenAI and Anthropic and Google that are saying that releasing these powerful models to the public, allowing anybody who wants to use these tools to do whatever they want without anybody knowing. I cannot say which approach I agree with more. It's very hard to say, but there's really smart people on both sides of this argument.

9:42And speaking of Meta, they just released their standalone text-to-image generator. You can go and use it at imagine.meta.com. It is a free tool that generates pretty good results from my very initial testing. I'm planning to test it a lot more, but it's just another tool that you can now start using in order to create images out of thin air, similar to MidJourney that is a paid service. And you can also use other free tools such as the image generation capabilities within BARD that come from DALI 3 from OpenAI and tools like Leonardo that announced that they just raised $30 million this week. And as part of this raise, they've shared that they now have 7 million users and these 7 million users has created over 700 million images so far, which just tells us that there's going to be a lot of options for us, the users, to create images with AI.

10:41And speaking of the meta model, it currently includes a built-in visible watermark to tell people that it's been created by AI, and they're planning to replace this with invisible watermarks so they cannot be deleted by the users. So it will always be able to tell that these are AI generated. Another giant that is deeply involved in the whole AI craze right now is obviously Microsoft. And Microsoft has been exclusively providing the open AI tools across its capabilities. And with an interview with one of their executives, Eric Boyd, he suggested that they will soon offer other large language models beside open AI on their platforms.

11:23The reality is I don't think they have much of a choice because their largest competition when it comes to providing cloud services, Amazon on AWS, is providing their clients multiple options such as Anthropic and Coherent Stability AI and other open source models. And Microsoft on Azure is currently only offering the OpenAI model. And if they want to stay competitive and they want to keep clients that will be able to make choices, they will need to add additional options. I think the other fear is with all the madness that happened at the leadership of OpenAI, having that as the only option is a very big risk for Microsoft.

11:58So that's another reason for them to put other models on their platform. And now two very interesting pieces of research that were released this week. The first one comes from the UK Department of Education. They've analyzed 365 jobs based on different skills that they require, combining it with the 10 most common uses of artificial intelligence in different aspects of businesses. And what they found, while it's not surprising, it's now backed up by real research, which like top paying jobs are at the highest risks. Things like management consultants and financial advisors and accountants, salespeople, solicitors, and even psychologists ranked at the highest chances of displacement.

12:42Obviously, on the lowest risk of being replaced by AI are jobs that require a lot of hands-on and physical skills, such as sports players and roofers, which require a lot of physical accuracy in their actions, which means that white-collar, high-paying jobs of highly educated people are at a much higher risk than actually low-level, lower-paying, less-skill-related jobs that are at much less risk, at least at this point, until robots are developed in an economical scale. The last piece of research that is really interesting comes out of a spinoff out of MIT called Liquid AI. Liquid AI is building a new infrastructure for AI development that is called Liquid Neural Networks.

13:27This new model is different from GPTs, and the main difference is that it's capable of training on significantly less parameters and achieving the same results. And in addition, it can change and adapt as its parameter changes much better, much faster than GPTs do that are more or less stuck with the amount of data they were provided when they were trained. So as some of the news that I released last week, all this tells us is the acceleration of everything that we see in the AI world is not stopping. If anything, it's going to go even faster as companies and research groups develop capabilities to train and ride models faster and with smaller technological footprint, which means we'll be able to get a lot more of it for a lot less investment, which means we'll get higher capabilities faster than ever before.

14:18That's it for this week. We'll be back with another interview episode, as always, on Tuesday morning. And until then, explore AI, learn new things, share it with the world, share it with me on LinkedIn, and have an amazing weekend.