102 | Claude introduces Projects a better version of GPTs, Google Gemini 1.5 Pro with 2 million tokens is now publicly available, Suno and Udio are being sued, and other important AI news for the week ending on June 28

0:00Hello and welcome to Leveraging AI, the podcast that shares practical ethical ways to improve efficiency, grow your business, and advance your career. This is Isar, Maitis, your host, and this is another weekend news edition. I actually thought of maybe not doing the news edition this week because we released two episodes already, episode 100, part one and 100, part two, which is on your player episodes 100 and 101. If you miss those, they are amazing. You want to go back and listen to them because we had 18 different AI experts, each and every one of them sharing the number one tip that they have on how to leverage AI in a business context.

0:37So it's an amazing list of people providing an amazing wealth of AI knowledge that you don't want to miss. So if you miss those, go back and listen to those. But there were so many big things happening this week that I couldn't skip that and not let you know about it. So here we go.

0:57We're going to start with Anthropic. Anthropic made some really big announcements in the past two weeks. So last week, they released Claude 3.5 Sonnet, which is their middle scale model. And we talked a little bit about this. It's a very powerful model is twice as fast as the previous one that they released only three months ago. And it can do some cool stuff. One of the capabilities that released last week is called artifacts, which allows you to see the results of what you're doing on the right side of the screen, whether it's code generation or images or stuff like that. And it is an amazing capability that allows companies and individuals deploy pieces of code on the fly for any need.

1:34One of my clients showed me a really cool example where they took a matrix that they had for HR evaluation purposes and uploaded to Claude and ask it to code a web user interface that will allow people to give these reviews that they were doing manually for years on a web form. And within five minutes, they had the form ready that they can now copy the code and embed it on their internal employee website for them to evaluate each other. And that would have cost them X number of thousands of dollars previously to hire somebody to do this kind of work. So endless numbers of opportunities where you can code your own stuff, see the results right there, and then deploy it in any way you want.

2:13A lot of people are developing games and so on, but that was last week. I'm just sharing with you a lot of cool use cases. But this week they've announced projects. And projects is Claude's version of GPTs, which is the automation platform created by OpenAI a while back. So what you can do is you can create dedicated workspaces where users can chat with Claude, upload files, work together, collaborate on documents, code, images, and more, and build these recurring automations that will do the same thing consistently again and again. Similar to the things you can do with GPTs, you can upload files like PDFs and images and code snippets, and then use Claude to analyze the data and generate new outcome based on the data that you've uploaded.

2:58And as I mentioned, using the same instructions again and again. So you don't have to re-prompt to get the same results. As with anything anthropic, they put a big emphasis on privacy and security. So they are claiming that the data that's uploaded to projects is encrypted and is not used to train their models, which means you can, in theory, upload any data that you have, even if it's somewhat sensitive, without worrying where that data is going to go. Now, that being said, the only thing it's still missing that drives me freaking crazy is the fact it still does not have internet access, meaning you still are limited to the data that you're uploading and you cannot query anything from the web.

3:35The workaround for that is you can use both Cloud Sonnet 3.5 as well as Cloud 3 Opus, their largest model, within perplexity if you have the paid version. So I found a lot of success in doing things that were not possible before, just running Cloud within perplexity and that gives it internet access still with the power of Claude, which I absolutely love as a model to review, summarize, and create content. Staying on Anthropic, their CEO, Dario Amadei, has held an interview this past week, and he talked a lot about the potential negative income of AI on society as a whole. And he's claiming that UBI, universal basic income, is just not sufficient to address the potential economic inequality that AI is going to cause, meaning a lot of companies are going to make stupid amounts of money like we've never seen before.

4:29And many, many people are going to be out of jobs and will not be able to make money. So he's saying that UBI could provide a safety net, but as a society, we need to come up with proactive solutions that will help people get purpose in life and fulfillment once their job is displaced by the AI. I agree with him 100%. I think the only way to do this is some kind of an international collaboration that has to start now because this is right around the corner and that we'll find ways to help people find purpose in life and do things when a lot of jobs are going to disappear. He suggested stuff like national service programs, lifelong education and training programs, and incentives for people to pursue creative and care work that he claims will take AI longer to automate.

5:17So in a nutshell, I agree with him 100 % and that's one of my biggest fears as far as the negative impacts of AI on our society. I want to pause the news for just a second to share with you an exciting piece of information. We are opening the registration for the July cohort of our highly successful AI Business Transformation course. It's a course I've been personally teaching since April of last year. We're teaching two courses every single month. So hundreds of individuals and companies have went through that course and are literally transforming their businesses because the information that they have gathered.

5:49The course includes four sessions of two hours starting on July 8th, four weeks in a row at noon Eastern. So it fits anybody, whether you're in Europe or in the East Coast or on the West Coast or anywhere in between. These courses sell out every single time we run them. So if you want to change the trajectory of your career by becoming the AI expert of your company, or if you are in a leadership position and you want to transform your business, then go and check it out. The link is going to be in the show notes so you can open it right now on your phone and click through and see what is the information.

6:21In addition, if you're listening to this episode still on Saturday or Sunday, meaning in the month of June, you are in luck because to celebrate the episode 100 of the Leveraging AI podcast, we created a promo code for$200 off, which we have never done before and probably will never do again, at least not before episode 200. So maybe that will happen. But if you are listening to this, as I mentioned, before the end of June, which is when this promo code expires, you can use promo code leveragingai100, all uppercase, to get$200 off. So if you can make a decision quickly, you are in luck and you can save$200.

7:00If not a big deal, come and join us. I can guarantee you that the money you're going to invest is going to save you that tenfold in the first month during the course, probably. So it's worth every cent. And in the landing page, in the link that is shared in the notes, you can see what people who took the course are saying about it. I think that will help you make up your mind. And now back to the news. And from Anthropic and Claude to obviously their biggest competitor, OpenAI, an interesting piece of news surfaced this week that claims that OpenAI is now selling more of their own models directly to customers than through its partnership with Microsoft.

7:36Why is that interesting? Part of the partnership basically gave OpenAI a channel to get faster, bigger distribution because Microsoft is connected to most or a lot of the corporations, at least on the Western hemisphere. And through them, they were supposed to sell a lot more licenses. But apparently now the tables have turned and OpenAI is selling its licenses directly to corporations more than it's selling through Microsoft. That obviously builds on a lot of moves that we've seen from OpenAI by hiring a lot lot of senior people from the product side and distribution side and sales and trying to focus a lot more on distribution and productization of what started as a research project.

8:18And that definitely seems to be paying dividends. What does that mean for Microsoft and how they feel about this new thing? I don't know. I can imagine. But we've seen this frenemies relationship going on for a while where Microsoft announced that they're building their own models while in parallel, OpenAI are developing more and more capabilities for the enterprise level. So it will be interesting to see how this evolves in the future. Staying on OpenAI, OpenAI a few weeks ago gave us the most amazing demo of a voice-capable AI model that they call GPT-4.0 for Omni. And the demos were nothing short of incredible and allowed the users to have a normal human conversation across any topic while interrupting the model in the middle of the sentence and going back and forth and gaining benefits across business and education and personal life and so on.

9:10And then they released some of the features, but not the voice feature. And they said that the voice feature is going to be delayed through the fall. And yet a few users suddenly had access to it in this past week. These users already shared with excitement their capability to now use the tool and have a conversation with it and even create some really cool creative outcomes with role-playing with the models across the story. But OpenAI quickly turned this off and now apparently nobody has access to it. So was that a mistake? Maybe. Was that a test? We don't really know. There's no clear feedback on that from OpenAI themselves.

9:48But the reality is the models are not available right now and they are still claiming that they're going to be released around the fall. We've seen this pattern from a lot of the big players, including OpenAI, Microsoft, and Google, where they all in the recent announcement showed really amazing things. And most of them are not available to us in the near future. I really hope that they will release them as soon as possible, because some of these capabilities are game changers when it comes to business efficiency and our ability to engage with computers. Like this changes everything because we'll be able to literally have a conversation with any computer, whether it's our watch, our car, things that run our house, the computers at work, anything, we'll just won't need the keyboard anymore.

10:31And we'll be able to have conversations with computers and they will understand our needs and the background and the context and what we're trying to do. And we'll be able to complete the tasks and help us in being a lot more efficient and avoid tedious tasks. Another interesting report from OpenAI this week showed in a blog posts that they are now leveraging GPT-4 to critique GPT-4. So what they're doing is they're taking the outputs of the regular model and sending it to the same model, asking it to find errors and issues with the original content that was created. And it's actually doing a really good job.

11:05So the blog post shows examples of mistakes caught by GPT-4 from GPT-4, such as incorrect calculations and misinterpreted context and outdated information, as well as hallucinations and inaccuracies and biases in the original content that was created. This is probably going to be a built-in mechanism in most or all large language models to reduce hallucinations to a bare minimum. I don't think we can avoid them completely, but this will dramatically reduce the number of errors and issues that we have with the content generated by these large language models. Now, what does that mean as far as cost and energy and so on?

11:47That's a big question, because basically every prompt you run is running twice. So it will be interesting to see how they overcome that issue. Maybe a lot of it is going to be done through training or through some other mechanisms that will reduce the doubling of the cost of every prompt that we run. From OpenAI to another giant that we don't talk about a lot, and that's Databricks. Databricks has been around as a data and AI company, and they've been doing a very good and successful work in the corporate world, allowing companies to dump their data into their databases and to engage with it through different AI tools.

12:24And they're now announcing a new architecture. They're called Compound AI Systems that they're claiming can challenge the dominance of OpenAI and Anthropic and Gemini from Google. What compound AI systems is it combines multiple AI models and techniques to create more powerful and capable systems than just a single model. So think about how a large complex task is done in a business today. There are multiple people doing multiple things because they have different expertise and this works in exactly the same way. So basically there's multiple models and the task moves around from one model to the other to do more specialized aspects and type of tasks in order to complete the barrier task, they're claiming faster, better, and more efficiently.

13:10The other benefit of that is obviously cost flexibility because you can pick specific models for specific tasks that was going to be cheaper than running the same behemoth model if you're running, let's say, Claude III Opus. And this way you can save both time, money, electricity, and so on. So I think it's an interesting approach. I think in the future, that's what we're going to see from more or less from everyone. And obviously we cannot pass a week without talking about Google. Google made a big announcement this week. Google has released Gemini 1.5 Flash and Gemini 1.5 Pro models to the public.

13:42Both these models were released previously to a short waitlist of companies that were able to use it, but now they're opening it to anybody that are Google Cloud customers through their Google AI Studio in Vertex. And what these models are, the Flash 1.5 model has a relatively short context window of 8 ,000 tokens, but it's designed for fast, efficient processing of shortened text inputs. So anything that needs a quick response and is not too complicated, this is perfect for it. And the other one is 1.5 Pro model that has a 2 million tokens context window. So 1.5 Pro previous model had a million tokens context window, which was still the largest in the industry.

14:23The second one being Cloud3 with 200 ,000. So that was 5X. And now it's 10X, the next model. So it allows to run a huge amount of data into a single prompt and keep it consistent and coherent and get really good and detailed results. From that, I am certain that a lot of people will jump all over this for more complex analysis of company data. In addition, they have announced they're releasing Gemma 2. Gemma is their line of open source models. And Gemma 2 is a multimodal AI model that can generate and understand images, audio, and video, enabling the analysis and creation of sophisticated real world kind of scenarios.

15:04And as I mentioned, that's an open source model that anybody can manipulate and use. But maybe the biggest announcements from Google this week is that they are more or less forcing Gemini into workspace apps like Gmail, Drive, Docs, Sheets, Slides, et cetera, meaning previously you had to sign up for it and activate it in order to have it available. And now they're allowing anyone to get access to this. It's just about a matter of getting the licenses and activating them. And these tools will be available on a panel on the side of all the apps that we know in Google, as I mentioned, such as Slide and Drive and Docs, et cetera.

15:41And it will allow people to analyze the existing data, summarize it, create calculations, et cetera, et cetera, everything we know from large language models without leaving the native apps themselves. We saw the same exact thing done by Microsoft in their Office 365 environment. So I expect this competition to continue going and develop more and more capabilities and integrations that will not only allow you to work within one of these tools, but actually to share data and knowledge across all of these. So you can ask general questions about your business and find anything you want, regardless of which application you were in.

16:15And there's zero doubt in my mind. That's the direction it's all going. Now, speaking of new models and capabilities and who's leading and which one is better, Hugging Face, which is the biggest community for AI platforms, has launched a new leaderboard to track the performance of open source large language models. So they call it OpenLLM Leaderboard, and it evaluates open source models across multiple benchmarks, including language understanding and content generation and reasoning tasks, etc. And the interesting thing is that the leading models on the leaderboard right now are all from China.

16:49So models like Wudao 2 and Pangu Alpha and CPM2 have all achieved really high scores, outperforming the open source models that was released by OpenAI and Google. So the competition between the US and China or the West and China, if you want, is definitely burning. And I don't expect that to slow anytime soon. Now, speaking on interesting tools and releases, Figma, which is the world's most popular collaborative design platform, has announced a significant redesign that integrates AI to more or less everything that they do. So the AI-powered features allows designers to complete various tasks that previously were manual using only natural language.

17:32And it also can make a lot of suggestion and automate a lot of processes. It can do things like create images and icons and illustrations, but it can also create and suggest complete designs and color patterns and palettes and typography pairing for the relevant style that you're trying to get. and really helping designers be more creative a lot faster. Now, in addition, you can upload an existing design and Figma AI will give you insights about how accessible this website is going to be. Is it built well for responsiveness? What's the consistency across multiple pages and aspects of the website?

18:08And really help with creating a high quality, consistent design across everything that you're doing. This is not a surprise that we're talking about this for a while, but now it's released and available. and I'm expecting that it's going to drive a very dramatic change in the way websites are designed, mostly because Figma are a huge company that has a huge distribution with, as I mentioned, most of the web design is done on that platform today. And so we will see dramatic changes in that industry as well. Speaking of companies that has all the right components in order to deploy successful AI tools, McKinsey introduced this week a comprehensive suite of 20 plus AI tools within their environment developed by their in-house R &D team.

18:53These products combine advanced AI capabilities and machine learning techniques combined with a huge amount of data and domain expertise that McKinsey has gathered through the years by helping thousands of companies across multiple industries run their businesses more efficient. These tools are going to be available on McKinsey's Horizon platform, which incorporates best practices and risk guard rails for security and privacy. Now, going back to connecting this with Figma, both these companies have the three things that company needs in order to be successful in the AI era. They have a huge amount of data that is proprietary that almost nobody else has.

19:32So on the McKinsey side is all the data that they've collected from working, as I mentioned, with thousands of companies for generations. And because both these companies has the three things that are required in order to be successful in deploying AI tools on a large scale, they have the data, they have the resources to train the models and build the right technology, and they have the relationships and the distribution with their target audience. So in the Figma example, they have the data from millions of designs and the relationships with the designers themselves. And on the McKinsey side, they have the data from years of supporting thousands of companies across multiple industries, and obviously the current relationships that they have with their clients in order to embed this and benefit from it.

20:20So I expect both these tools to be highly successful and help both these companies grow faster and provide more value to their clients. And now two news about the creative world. So the first one, which is cool and interesting, is Toys R Us, apparently the company still exists somehow in some shape or form, has collaborated with a creative studio called Native Foreign to produce the first ever brand film using OpenAI's Sora tool. So those of you who don't know, Sora came as a complete surprise early this year, I think February, as an incredible tool to convert text to video. So short prompts turn into a minute long, high definition, highly consistent videos that is very hard to differentiate from real life.

21:06That being said, OpenAI never released this to the public. They only gave access to people from the industry in order to evaluate and learn and improve it. And now we're seeing the first example of what can be done with it. So the movie called The Origin of Toys R Us tells the story of their founder, Charles Lazarus, and his vision on how he thought he could transform and did for a while the toy store industry. And the video is actually a pretty cool, cute video. I suggest you watch it. It's a very short video and it's featuring the iconic mascot, Jeffrey the giraffe. Now this film was premiered on the 2024 Cannes Film Festival, showcasing the innovation in technology and the capabilities to do that.

21:48And the really crazy thing about it is that using Sora, the film was created in just a few weeks. So what they were able to do is to generate hundreds of iteration shots and then edit them together into one cohesive film in a timeframe that it was absolutely impossible with legacy systems. And I mentioned that before. I think that's the direction that we're going. I think we're going to see an explosion of creativity and a lot more video content than we've ever seen because anybody will be able to create really high-end videos with AI. I'm predicting that by the end of this year, but worst case scenario, Q1 of next year, these AI tools will be 100 % consistent, will create content that will be indifferentiable from real life or from traditional cartoons, and will allow the users to edit and control the camera views, the zooms, the transitions, and even edit stuff after it's created and change things in specific scenes, and hence replacing a lot of the manual work across the entire film and video generation industry.

22:51And that's going to be a huge revolution that has significant implications to people in that industry, but way beyond because as I mentioned, the amount of video content we'll be exposed to is going to be insane. And the other piece of news from the Creative World, Universal Music Group, Sony Music Entertainment and Warner Music Group and a bunch of other major record labels are filing a lawsuit against the AI generation companies like Suno and Udeo. Both these tools enable to create really high-end music from just a short prompt, and you can choose your style and genre and what the topic needs to be about, and it creates the lyrics and the music and the singer and everything else.

23:30And it's absolutely magical, and I love it, but these companies are arguing that both companies, Suno and Udio, are trained on copyright information, which they do not have the rights to train on. Now, they've shared several examples on how they can generate songs from these tools that sound very much like existing songs. The examples included Chuck Berry's Johnny B. Goode and Mariah Carey's All I Want for Christmas is You and other similar songs. But here's where we are. We've seen lawsuits like this against all the big players. And this is just another variation of this. Now, one thing that is similar and one thing that is highly different.

24:10The things that is similar is question number one, can they prove it? Can they prove that it was trained on this data? I don't think they can claim that based on assumptions. Oh, this sounds a lot like X. But a lot of songs that people create sound a lot like X, which leads to item number two. Can they prove that this is not fair use? So in all the previous copyright infringement cases with AI, what the AI companies claim is that training on the data is not a problem because this is what artists do. So one artist listens to music. he gets ideas to a new kind of music, and then they create the music.

24:43And you can't sue them for the new music that they created, despite the fact they were listening to another piece of music. So that's the claim that all these companies are making, and the courts will have to decide whether training on data is fair use or not. Once they decide that on one case, I assume this is going to be a domino effect, and that will trickle to all the other cases. Now, the third question is, can they prove damages? If you're suing somebody to get money, you need to prove that the fact that now there is a song that sounds like Johnny B. Goode is going to reduce the amount of times people listen to Johnny B.

25:16Goode. And that will be, I think, very hard to prove. That being said, I think as a society, we have to find ways to compensate creators in a different way than we did so far and come up with completely different mechanisms in order to allow people to continue being creative with or without AI and get compensated for it without having their work stolen in five seconds and reproduced in a million other variations because the song was successful. So where's this going? I don't know, but I hope we're going towards an era where there's going to be a fair solution for that that will enable us to experience the explosion in creativity while still allowing people to benefit from it.

25:53Be back on Tuesday with another fascinating interview that is going to show you seven powerful use cases you need to know on how to use ChatGPT in a business environment. It is an incredible episode with a huge amount of value. Again, seven very detailed use cases with how to across multiple aspects of the business from very basic to highly complex. And I'm sure you're going to absolutely love it. And that's going to be released on Tuesday. Until then, as I mentioned, if you haven't listened to episode 100, go and check it out. It's an amazing episode. and have an amazing weekend.