Leveraging AI

117 | Deloitte Survey Reveals AI Challenges, US companies warn of growing AI risk, exciting new about Midjourney, Google Imagen 3, Adobe Drops Magic Fixup, Meta's Sapiens and more AI news for the week of August 24

• Isar Meitis • Season 1 • Episode 117

Share episode

Is AI more of a threat or an opportunity for your business?

In this AI news episode of Leveraging AI, dive into the latest data from industry-leading surveys, uncover the common pitfalls in AI implementation, and discuss practical strategies for mitigating risks while maximizing ROI. From Deloitte's quarterly insights to the rising concerns highlighted by the Financial Times, this episode arms you with the knowledge to navigate AI’s complex landscape with confidence.

- Why 68% of companies struggle to move AI projects from experiment to production.
- The top risks that are making AI a major concern for 56% of U.S. companies.
- How to overcome data management challenges to successfully deploy AI.
- The unexpected truths about AI model hallucinations and their impact on your business.
- Why continuous employee education is key to your AI strategy's success.

About Leveraging AI

If you’ve enjoyed or benefited from some of the insights of this episode, leave us a five-star review on your favorite podcast platform, and let us know what you learned, found helpful, or liked most about this show!

Hello, and welcome to a weekend news episode of the Leveraging AI podcast, a podcast that shows practical, ethical ways to leverage AI to improve efficiency, grow your business and advance your career. This is Isar Meitis, your host. And every week, there are a lot of news, but this one is going to be a little different. We're going to start with many surveys and research that was released this week that shows the impact and the confusion that AI has on multiple industries and different aspects of them. And then we'll dive into news from specific companies. Lots of new models that have been released. Also open source stuff that happened. Many news about image generation. So lots of good stuff. So let's get started. As I mentioned, we're going to start with a few surveys. The first survey I want to talk about is Deloitte's quarterly survey that touches on the status of AI. They surveyed over 2700 businesses and technology leaders across 14 countries and six different industries. And most of what they found wasn't too surprising. As an example, 67 percent of organizations are increasing their investment in generative AI. The interesting aspect is they're doing it because of strong early value, meaning they started using it and they're seeing actual results. We'll see that there's going to be a lot of mixed feelings around this whole aspect of, is there a positive ROI on generative AI right now across different companies? Now, that being said, despite early signs of value, 68 percent of companies surveyed say they moved 30 percent or fewer of their experiments into production, meaning 70 percent of what they tried either didn't work or were running into issues or did not provide the right ROI. But most companies said that only 30 percent of their experiments moved into production. Now there's a very obvious challenge in the preparation and deployment of AI solutions. And that is data management. So 75 percent of companies increase their investment in data lifecycle management in support of generative AI. And 55 percent of companies said that they're avoiding certain gen AI use cases because of data related issues. This may be data security. This may be data not being clean. This may be the cost of moving data from one place to the other may not produce positive ROI in the process. But data in general is a very big aspect obviously of preparation and deployment of AI solutions. On risk management side of AI deployment, only 23 percent feel highly prepared for Gen AI related risk management and governance challenges, meaning three out of every four companies are not feeling prepared. And that's just their feeling, right? Even those 23%, we don't know how much they're actually ready for that. And the risks they manage vary across data quality, bias, security, trust, privacy, regulatory compliance, and more. So the risks comes from multiple aspects that we are aware of or should be aware of. But as I mentioned, most companies don't really know how to deal with them yet. Now from a value measurement, and I'm going back to what I said before, 41 percent struggle to define how to measure or to actually measure the exact impact of their generative AI efforts. Only 16 percent of all these companies produce regular reports to the CFO detailing the value generated by implementing GenAI. So quick summary, high risks that nobody exactly knows how to deal with. Huge investment in data management and data cleaning and so on, and mixed feelings as far as results, if they can even be measured. Now the Financial Times released another interesting report this week that focuses on risk factors for major U. S. Companies. So they surveyed 500 companies on their current status on a I implementation. 56 percent of companies cited a I as a risk factor. Factor in their recent annual reports. So this is beyond the survey. It's the company saying we have risks to the future of our company or specific aspects of our company due to the implementation AI. Only 9 percent of companies cited that in 2022. So two years later, the number grew from 9 percent to 56%. And that's very obvious because AI is now everywhere. And in 2022, it was in very early stages. If you remember Chachapiti came out late October of 2022. So that was the moment that started changing everything. Now out of those companies, only 33 saw generative AI as an opportunity versus a huge amount, more than two thirds of them that specified generative AI as a risk to the company. Types of risks they mentioned include increased competition, failure to keep up with rivals, exploding AI. So that's another aspect of competition, reputational and operational issues of implementing AI and what might be the outcomes of that ethical concerns, financial risks, meaning untested predictable costs of implementing AI or using AI, legal and regulatory uncertainties, cyber security, and more. These were just the main ones. Now, as far as companies in specific industries, over 90 percent of media companies surveyed said that they see AI as a serious risk. 86 percent of software and technology group mentioned the same thing. And high percentage of telecommunication, healthcare, financial services, retail, consumer and aerospace sectors. So a lot of companies see AI more as a risk than as an opportunity. Now, diving a little deeper, as I mentioned, the IT world sees that as a very significant problem. A company called Pluralsight did a survey of 200 executives at IT professionals in the US and the UK about what they think the impact of AI is going to be about their industry. What is going to be the impact of AI on their industry and their career. So 74 percent worry that AI will make their day to day skills obsolete. 69 percent believe that they're at risk of being replaced by AI. 96 percent of those people prioritize staying up to date with AI skills for job security. So if you're listening to this podcast in order to help yourself stay on top of what's happening in the AI world. You're part of a very big group that sees this as a risk and sees education as a way forward. On the executive side, on the executive survey, 35 percent plan to invest in AI. in order to eliminate unnecessary positions, which kind of tells you that the fears of being replaced by AI are justified because one third of executives right now, and this may grow in the future, plan to replace people with AI capabilities. 90 percent don't completely understand their team's AI skills and proficiency, meaning even though people are using more and more AI, we saw that in other surveys. Most executives don't really know how AI is being implemented in their businesses. And that's obviously generates a very big problem. the next finding is self contradicting and I will try to explain it to you, but 81 percent of it professionals feels confident that they can integrate AI into their current roles, but only 12 percent have significant experience working with AI. The way I interpret these two parameters is that people understand that it's a Potentially serious risk to their career and hence they have to convince themselves that we'll be able to make those changes, but not enough of them are actually investing the time, whether because they don't have the right resources from their company or because they don't have the time because they're too swamped with their day to day in order to actually learn the materials. Now the one thing that was almost unanimous across all these surveys that we're looking at is that 94 percent of IT professionals and 95 percent of executives believe that AI initiatives will fail without staff that can effectively use the tools. So that goes back to something I've said in multiple shows here before the finding the right tools and buying them does nothing for the success of your business or for your personal growth. The only way to make this work is to have serious education, actual plans in place on how to develop, Processes around the tools that you implement and how to train the people on how to follow these processes and how to use the tools and what guardrails they need to put in place in order to avoid the risks that we talked about before. So it all comes to leadership. Putting together in place, training mechanisms and processes development. I do this with multiple companies around the world. And I can tell you, even whether, when I'm working with small companies or when I'm working with large international corporations, it's always the same. People do not have enough training before we get started. And then we develop the right systems. We develop the right processes. We develop the right training and we see. Immediate results. So if you are looking to do something like that, first of all, visit my website, multiply. ai, and you'll find multiple opportunities on how we can help you. But whether it's us or somebody else, you have to find the right way to train your people. You have to find, you have to set up a group within your company, an AI committee that will look at all the different opportunities. We'll evaluate the risks, we'll develop Processes around the systems that you choose and we'll develop a training plan, a continuous evergreen training plan that continuously updates and keeps people up to date with what are the tools, what are the risks, what are the capabilities so you can maximize the benefit while reducing the risk and hence you're going to become one of these companies that see AI as an opportunity versus a risk and all your competitors Most likely are going to see the risk as higher than the opportunity because companies like yours will be significantly ahead of them. This will dramatically change the competitive landscape in the next two or three years and the companies are going to be first movers to do this are going to gain big benefits. I want to pause the news just for one second to share something exciting from my company, Multiplai. We have been teaching the AI business transformation course to companies since April of last year. We've been teaching two courses every single month. Most of them are private courses. And we had hundreds of companies take the course and completely transform their businesses with AI based on what they've learned in the course. But this course is instructor led by me and it requires you to spend two hours a week, four weeks in a row at the same time of day, working with me and my team in order to learn these kind of things, which is may or may not be comfortable for your time schedule and other commitments. So I'm really excited to share that we now have an offline self paced version of the same course. You can log into our platform and there's a link in the show notes. So you don't have to look for it or try to figure out or remember what I say. Literally just open the platform, which you're listening right now. And there's a link in the show notes, which will take you to get to the course. The course is eight hours of video of me explaining multiple aspects on how to implement AI in your business, going from an initial introduction to AI. If nothing. All the way to hands-on experimentation and exercises across multiple aspects of the business from data analysis to decision making, to business strategy, to content creation. Literally every aspect you can use AI in your business today, including explanation. Tools and use cases for you to test as well as a full checklist in the end of how to implement AI successfully in your business. So if this is something that's interesting to you and you don't have the time to spend time with me in an instructor led environment, you can now do this on your own and drive significant efficiencies to your business using the step by step course that we now have available and now back to the news. Now, the last survey I want to talk about was actually a survey done by people in Cornell. And by several different universities, plus a company. So cornell university of Washington and Waterloo, as well as a 12, which is an AI open source company, which we're going to talk about more later in this show, have done a research to try to find out how much different AI models actually hallucinate and the way they've done this, they've checked models like. GPT 4. 0, Lama 3, Mixed Trial, Claude, Advanced Models, and so on. So the biggest and latest models that exist today, but also from different sizes, big ones and small ones, they specifically asked questions that are not easily answerable by Wikipedia, as an example, which is a data source that they all trained on. So they covered topics like law, health, history, geography, culture, and more, but they were trying to ask questions that the answers are not obvious to see what these models would do. The key findings were that All the models produce about 35 percent. What they find is actually pretty alarming. What they found is that even the best models produce hallucination free text about 35 percent of the time, meaning 65 percent of the time, two thirds of the time, some of the data that comes back is not real. Now, none of the models performed exceptionally well across all topics. Two topics that were very hard for the models is celebrities and information about them as well as financial questions, which tells you they're obviously not very good in the latest and greatest information. So if you're trying to find the late news or something that's happening right now, which is a lot of financial stuff as well as celebrity stuff, you're probably not going to get accurate information. And the type of questions that were easiest for the models and had the least amount of hallucinations is geography and computer science questions. Now, as far as the differences between the models, open AI models, GPT 3. 5 and GPT 4. 0 hallucinated less than any other model. And even web search capable models still struggled with non Wikipedia related questions. Now the model that was actually the most factual was Cloud3 Haiku, but it only answered 72 percent of the questions, meaning it did not answer about 30%, But out of those 72 percent of questions he did answer, it provided the most accurate information. Now, if you ask me, that's exactly the direction we need to go, meaning the models need to tell you they don't know when they don't know and not answer the questions or answer shorter when they don't have all the information and tell you what they don't know as far as the rest of your answer versus making stuff up. So I actually see that as a very good step in the right direction on specifically Haiku. Why is Haiku different than the other cloud models? I don't really know. So how do we connect that to the previous surveys we looked at? There is a clear craze in all companies, large and small to implement AI, whether they know what the ROI is, whether they don't know what the ROI is, whether they have the right training in place or the right guardrails, or they have solutions for problems and so on. There's a very strong push to implement AI. The problem is Most people are not aware on how bad hallucinations can be. Now, yes, this was a very specific use case of asking AI about general information that is not easily available in Wikipedia. So that does not represent the general solution. But let's say you, let's say you are implementing RAG solutions in your company, and you're actually giving it your company's information, and then your employees will become dependent on using it to find information about products, about services, about refund policies, about HR stuff, whatever the solution that you're putting in place. And let's say only 15 percent or 20 percent is incorrect and not 65%. That still generates a very big problem, whether internal or external when using these tools. And that's something you must be aware of. Ethan Mollick, who is a person I follow regularly, he's a professor and he shares a lot of AI related research and so on, shared two facts very clearly. One, these models hallucinate, period. There's no model that doesn't. You can get that number lower, but it still will. Number two, from their research, it's very clear that people become dependent on AI. On these AI chatbots, once they get implemented and they stop checking the facts as soon as they get used to using these tools. So the combination of these two facts tells you one thing that you have to continuously invest in education of your employees. So they know that they have to check the information that these tools provide. And yes, it's going to make it less efficient, but verifying information or collecting and summarizing information take very different kind of times. It's a lot faster to just verify information, especially if you build your solution with links and citations from the actual source content, which dramatically increases your ability to test and check whether the information is accurate. Now, all these models, developers are working to continuously reduce hallucinations. But as of right now, the only way to really be sure is to have a human in the loop, fact checking what these tools actually providing to you. Now, from different types of research, I want to talk about something else that impacts right now, specifically California, but then is going to impact All of us across everywhere in the world and specifically i'm talking about bill sb 1047 which is a legislation that was put in place by Democratic state senator in california scott weiner And the key and the top aspects of the bill SB 1047 talks about increasing security and reducing risks from the implementation of frontier AI models. It is very detailed and it's very harsh. It is Putting a lot of requirements or come on companies that are developing these larger models, any the way it measures, whether a model is a frontier model or not is any company that has spent over a hundred million dollar in training means it's going to be a large model that falls under the impact of this law. It also provides. whistleblower protection for employees who are going to come out and discuss some things are done, some say shady actions done by their companies when it comes to AI development. Now since the law has been introduced, it gotten both positive and negative responses from some very influential people. So among the opposition is Nancy Pelosi, the former house speaker who said, and I'm quoting, it's well intentioned, but ill informed, also industry group TechNet that represents Google, Meta, and OpenAI, obviously oppose this law. VC firm Anderson Hurwitz, one of the largest VCs in the tech world that is very active in Silicon Valley, and eight democratic members of Congress are opposing this law. There is a. extremely strong lobbying from big tech against this law that is claiming that it's going to stifle innovation, that it will likely become obsolete because it might be cheaper to train models that are still extremely highly dangerous. And so the arbitrary number of a hundred million may become irrelevant and that there is unclear public benefits Because right now, there aren't any clear facts that are showing that these models generate any risk. On the flip side, there's a letter by four very prominent AI professors, including Yoshua Bengio and Geoffrey Hinton, Both consider the godfathers of AI that are big supporters off the law. And they're saying that it's extremely necessary and that the risks are way more significant that the big tech companies are willing to admit. So they are supporting the law. Now, where am I on this? I'm probably in the middle. I'm trying to keep myself informed and try to have an opinion on this. I definitely feel that there are many risks to AI. The flip side of that is that there's fierce competition from international groups. So do we want China to be ahead of us in AI development? Do we want Russia to be ahead of us in AI development? Do you want Iran to be ahead of us in AI development? I think there's a very fine line that has to be walked. I think there has to be government involvement and hopefully in the international bodies that will monitor What's possible and what's allowed just like nuclear weapons that will monitor the development that the deployment of these AI models. There's a whole conversation about open source models and should they be covered and how much by this legislation. But the questions are real and they're very big questions. And as you can understand from leading people in the top of, from the top level of understanding AI, there's contradiction, contradicting opinions on where we should go with this. Now, where's the law right now? The law has passed the committee after multiple amendments that were driven mostly by the tech lobby. But as of right now, it's moving from the committee to a full vote on the California's assembly. Will it pass or not? I will keep you posted as soon as that happens, but I'm just want to remind you that a similar law was passed by the EU called the EU act. We talked about it many times before, and that has led to a multiple companies not releasing their models in the EU because they're afraid of the potential consequences of being caught on something they're not supposed to be doing or by doing stuff that they think is okay, but the legislator in the EU does not think is okay. And that definitely is harming the companies in the EU that want to use these tools. So there are definitely arguments Good arguments on both sides of that story. Whatever happens in California is going to impact not just California from two different aspects. One, it's the first state that's going to have a law like that. If it passes, that will obviously trickle and trigger other States to do the same thing. It will also probably impact the federal level that I shock that hasn't done anything similar to this so far. So there've been several attempts. There's been a presidential act that. moved in that direction, but nothing major that has been done yet and is definitely required. But also this law covers any AI tools that are used in California and not just developed in California. So that's going to impact probably the rest of the world. What I suggest to you, keep following the development and develop your own opinion on what do you think based on these multiple experts, what do you think is the right way to go? And if you have a way to impact that, go and use it in order to push it in the direction you believe will be most beneficial for us and for our kids. So with this, we're going to end the section about surveys and legislation and we're going to move into some other topics. As I mentioned, a lot of stuff happened this past week with AI image generation. So the first one is Midjourney, which I'm still considering the best image generator on the planet, even though the new Flux gives it a run for his money. But Midjourney just finally opened their website to everyone. so far, your only way to access Midjourney, unless you are a regular user, was to use it through a discord account on a discord server. It wasn't a big deal, but it definitely was cumbersome, especially for people who are not techies and the only people who could access the website initially was people who made a thousand images before that on discord. And then they got that number down to a hundred, but now you don't need that at all. You can just go to the mid journey website. And sign up for free, you can sign up either with your discord account, if you already have one, or with a Google account, and you also get 25 free images, which did not exist before, before you had to have a paid account. So a big step forward as far as making Me Journey available to everyone in an easy way. In addition, Me Journey released an image editor tool within the Me Journey website that actually works. extremely well and provides a lot of very useful capabilities. You can revise prompts, you can edit segments of the image by highlighting them with a brush and then telling it what to do with that specific section. You can change the aspect ratio of the image and zoom out and use your existing image as just one part of a bigger image. You And a lot of other cool stuff. I've been using it regularly since it came out a few days ago, and I absolutely love it. It's a very big improvement and a very useful tool. And it's now, as I mentioned, available to anyone for free for up to 25 images. And then you will have to buy a plan. If you are creating images regularly, it's worth every penny. Staying on the topic of image generation, Google just released their new AI image generator. It's called imagine. com. Three, and it's currently available to the public in the U S as a beta release by Google. It is very capable and powerful. I didn't get a chance to test it myself, but I saw a lot of examples that people shared online. It has much better text to image generation. It looks very realistic and it can generate almost any style of image you can imagine. It has improved prompt understanding for detail capture of what you're actually trying to create. It has in painting capabilities. So similar to what I mentioned about MidJourney, you can go and select a section of an image and change that, just that, while keeping everything else the same. And it's currently completely free to use. several different online publications like Petapixel, Tested and suggesting it's as good as mid journey in its image generation. As I mentioned, I did not test it myself. Once I will share the results with you, but it's definitely a very capable tool right now. And as I mentioned, it's available for free. You can access it through DeepMind's website if you want to test it out. Now, if you remember in the previous release, when Imagine 2 was released, there was a huge backlash against Google When people are trying to generate images of people, they got really weird results, such as people trying to generate images of the founding fathers, and they were from different races. There were one Asian and one black and so on. And there were other examples like that, which led a lot of people to assume, which Google obviously did not admit to, that they're over correcting for politically correctness. And hence these are the results, but either way, it led Google to. eliminate the ability to generate images of people at all from the platform. So Google is claiming that right now there are restrictions preventing people from creating specific content, which is a good idea in general. If you remember the episode from last week talking about what's happening right now on X with the image generation capabilities on X. AI and Grok, that is actually flux behind the scenes as far as the image generation model, there's definitely problematic data. Aspects of being able to generate anything. That being said from Google's history and the way Gemini is working at all, it's usually too restrictive to at least my personal taste, but you can definitely go and test it out and see if it will generate the stuff that you need. As I mentioned, very powerful model for free, at least for now. Now, in addition to the imagine capabilities. Last week, when Google introduced Pixel 9, they introduced a feature on the phone that's called re imagine and what re imagine tool allows you to do is it allows you to modify and edit actual images and pictures you're taking with your phone, meaning you can take a picture of something realistic that's actually happening. It's a picture taken with your phone, and then you can use AI to manipulate the image and add stuff to it in a highly realistic way. So the people testing this, that shared this article were able to add to it. Car wrecks, explosions, and even bloody corpses to specific images. Now, because it's AI driven and it's very well done, it's very hard to distinguish that these additions were added by AI and they match the lighting and the position and the perspective in the images to make them look highly realistic. There are no watermarks to the images. The only thing that indicates that AI changed it is the metadata of the images and you don't have to have the metadata because you can literally take a screenshot of the image that you created and now the metadata is gone and you can show the new image as a realistic image. So there's no way right now to detect that these images were manipulated by AI. Now, yes, this is doable with other tools that exist today, but not right there. And then on your phone, meaning your ability to manipulate the truth in real time, and then share it with the world on social media or text or chat or whatever is now becoming significantly more available, which raises the question. A lot of questions and concerns going back to the episode from last week and mentioning the episode that I've done episode 13 of this podcast that's called the truth is dead. We are walking with open eyes into an era where we don't have any clue if what we're saying is real, not now. Google has a technological solution they call Synth ID that allows them to tag synthetic images, basically changing the pixels on the image, creating an invisible watermark that can be detected. The only problem with Synth ID is That it only works on Google generated images that are generated from scratch by Google. And it doesn't add that to the manipulated images, and it's only for Google. So what happens when you do this with other tools? Right now, nothing. I really hope that the government will step in and will define this as a mandatory requirement that any image generation tool has either Synth ID or another technology that will enable identifying every image. image that is either created or manipulated by AI, that's going to make our lives a lot easier. Now, that being said, there are open source models out there, which anybody can take and then do whatever they want with them and remove whatever watermarking, even if it's this kind of watermarking from the images. But at least it's not going to exist as widely available to sing every single person on staying on the topic of image editing. Adobe just released a tool they call magic fix up, and it's really cool and again, also really troubling, depending your point of view. You can take an existing image that you took and then make changes to the image, very dramatic changes. You can move objects in the image. You can change their pose. You can move things from the background to the foreground or vice versa. And it knows how to adjust the rest of the image, the lighting, the shadows, the reflections, the contrast, the sharpness, everything in the object, when you move it or change it or manipulate it to fit its new position in the image and make it look completely realistic. So on one hand, for creators, this is very cool. You can use an image that you already have and make minor or major changes to it and keep it looking highly realistic. The problem becomes is that there are a lot of dark colors. Negative usages of this kind of technology that will allow people to easily manipulate existing images and change stuff in them, claiming that they are the original. This was possible before, right? So people use Photoshop again, an Adobe tool to manipulate images before, but it was very hard to do. It took a lot of time and to make it. Look realistic. It required very serious skills that not a lot of people had. And now literally anybody can do this. And in the test that Adobe did magic fix up was preferred over existing tools. 80 percent of the time, which. makes you understand that the tool is user friendly and gets better results than doing this manually, which means a lot more people will start using it to manipulate images. Going back to all the topics that we talked before, how do we know what's real and what's not? Now, the interesting thing about this particular, another interesting thing about this particular tool is that Adobe released it as an open source. So you can get the code for that. You On Adobe's GitHub channel, meaning instead of keeping this tool to themselves, which is what Adobe did so far, they're releasing it on open source, which represents a significant shift from their traditional strategy. And I think the goal behind this is to get more collaboration and make Adobe more accepted across the open source community, which will provide them additional benefits in the long run. Now, speaking of open source visual capabilities, Meta just released sapiens, like homo sapiens, just without the first half of the word. And it's a family of models that Allows to capture the human body across multiple aspects. One of them is to depose estimation. The other is body part segmentation. So even in a video, it knows what's, what are your arms? What are your feet? What are your fingers? What is your head? And so on for different types of analysis and manipulations and so on. It has depth estimation. So it does the depth of every pixel in a video as it's running and surface normal prediction. So it can predict where is the surface of everything that exists within an image or moving within a video. This Capability provides amazing functionality for developers. And as everything with meta, they're releasing it as open source. It was pre trained on 300 plus million real images of people, which makes you wonder, where did these people come from? And were these people agreeing for their images to be used for training? The answer is obviously not. And the answer is probably coming from the huge amount of database of images they have on Facebook and Instagram. They obviously did not admit to that, but that is my personal assumption. But it's currently maybe the most capable human analysis tool for images and videos. It will allow a lot of improvement across augmented reality and motion capture and human computer interfaces for multiple industries. And it reduces the risk, it reduces the need for task specific data collection. So previously, if you were trying to use these tools to understand when somebody is pulling a gun based on a data feed from your security cameras, and you had to train the models for that, now this general model can do a lot of these things. And let's switch gears. As always, there's almost every week there's news about OpenAI. So OpenAI shared that they found an Iranian group that was using the username of Storm2035 to generate misinformation and disinformation to run campaigns in the U. S. using ChatGPT. They block the account in order to prevent this specific group from using it further to do these kind of things. But that raises obviously the awareness to the fact That there are many negative aspects to AI generation. One of them is that bad actors, whether they're private or government funded, can use this technology, whether in this particular case, closed source or open source to generate and distribute misinformation and disinformation that will change, or at least try to change public opinion on major topics. In a year like this, where many countries in the world are going through an election cycle, including the U S that makes it even more dangerous. Staying on the topic of open AI wired magazine reported that their parent company Condé Nast have signed a multi year partnership with open AI to provide ChatGPT access and SearchGPT access to all the content generated by Condé Nast's publications, which include the New Yorker, Vogue, Wired, and a few others. So big name publications that are now going to be included in the chat. licensing agreements that OpenAI has signed. If you remember, OpenAI has had multiple of these agreements in the past with the Atlantic, Axel Springer, The Time, and other media outlets. The interesting thing about this is that Condé Nast's CEO, Roger Lynch, was very harsh about the way their data and other people's data was used to train these models. He even testified before Congress for favoring licensing of data instead of illegal training on data. And now I guess he got what he wanted. Now, in general, all these licensing deals represent, I think, a really big opportunity for everyone. For us as consumers, we will be able to find information a lot easier, and potentially not have to read full articles, but we'll be able to read just the segments that are relevant to us because the AI will allow us to do that. The second thing, it will allow these Publications to exist, because right now they're in a very serious struggle because the ad model kind of played its term and it's not working very well anymore. And many of them are running out of money. And the third is to me, that is the most important one is that it has the opportunity to actually increase and improve the quality of the content. Journalism, because right now all these publications have to chase clicks and ads because that's their only way to survive, meaning they will prefer to generate more articles versus better and deeper articles because they need new articles in order to generate more clicks in order to actually finance their operation. If these licensing deals change this trajectory, it will allow them to go back to better, deeper, more meaningful journalism, potentially releasing less articles that are well written and better research, which means we are all going to benefit from it. That's obviously a utopian direction that I hope we're going towards, but I really don't know what the outcome is going to be, but it's very clear that more and more of these licensing deal are being signed across all the leading AI companies together with the leading media companies. Now, the flip side of that Anthropic was just sued by a list of different authors for allegedly using copyright information to train their models without permission. What they are claiming is that Anthropic did, and I'm quoting, a large scale theft, end of quote, by using pirated copies of copyrighted books. Now, we have seen such lawsuits in the past, but not against Anthropic. So there were lawsuits against OpenAI and Microsoft and others. This is the first lawsuit against Anthropic themselves, but it's no different from all the other lawsuits that existed before. And as I mentioned in all the other episodes, it'll most likely go away by writing a big compensation check by these companies to whoever is suing them one way or another and or signing licensings deal with them moving forward. But I don't see that stopping the innovation and I don't see that stopping the stealing quote unquote of data in order to train future models. The stakes are too high, the competition is too fierce and they're going to keep on doing it. to show you how common this concept is Synchron Valley in a university appearance by Eric Schmidt this past week, he advised student to steal TikToks IP, but that was just an example to steal IP from other companies. in the development of their future startups. And then if it's not successful, nothing bad will happen because nobody will ever know. If it is successful, they'll have enough money to hire lawyers to clean it up. So he was not exactly aware that this was caught on camera. This interview is now available everywhere, even though it was taken down after it was put up. But as once you put something up on the internet, it's very hard to eliminate it. So you can go and watch the entire interview, but the summary is very clear. This is the mindset of Silicon Valley and how they drive innovation faster. And so all these companies are doing this. They're not even trying to hard it very hard. And they know that one day they will have to write a big check that the lawyers will have to clean up. But that's the way this game is currently played. Our next topic is going to be about models that have been released this week. So the first one that I want to talk about that I find very interesting is Microsoft just released by 3. 5. Microsoft has been releasing these open source five models for a while. So this is just their latest released. It has three different versions, one with 3. 8 billion parameters, one with 4. 15 billion and one with 40 billion. One billion parameters, which is obviously an order of magnitude bigger. And it's a multi modal model that has two different interesting capabilities. One, it knows vision and text, and the other, it has a mixture of expert model as well. It on several benchmarks, it actually outperforms the models from Google, OpenAI, Mistral, and Meta, not on all of them. And it's particularly strong on reasoning and math benchmarks. On the MOE mixture of experts models, it actually beats Gemini Flash 1. 5. So very capable model. It has 128, 000 tokens context window, which means you can put in about 100, 000 words into a single chat, which is the same as chat GPT. And it's a very small and fast model that can run locally on small devices and not send data to the cloud. Now that's on paper, people who actually tested it said that it struggled with some simple tests and phrasing issues and that it's not really as impressive as Microsoft is suggesting. But it's definitely a much better model. more capable model that people tested it also reported that it was actually very good at STEM and social science area and that it was also very good at maintaining efficiency across complex AI tasks. What does that mean? It means that there's another open source model that is available for absolutely free that people can download and use locally on their devices or host on their computers. Selected hosting services and run it without paying anybody, anything. And even if it's not as good as the latest smaller frontier models, it's very close to that and. Another company that released a interesting set of new models is Salesforce. So if you remember few weeks ago, I told you that Salesforce has released a huge data set for training that's called Mint 1T that will enable anybody to train models on the data. So if you want to Now they're sharing the models that they themselves trained on this data. They call these set of models XGen It's also known as BLEEP3. I don't know why there's two sets of names, but you can ask Salesforce. They will probably let you know. And these are also multi modal. They combine text, images, and other types of data. And they can even work with multiple aspects of data that are overlaid one on top of the other, meaning you can bring multiple images and different aspects of text, and the model will know how to combine them together, overlay the data one on top of the other, and understand the overall thing that you're trying to convey from the different data sources. The largest model has 4 billion parameters, and as I mentioned, it was trained on the 1 trillion dataset called Mint1T that they've released over the years. Previously. They are claiming that it's very good in complex tasks like answering questions about multiple images simultaneously and that it's very strong in medical, diagnostics, autonomous vehicles, and they said that it could be used potentially for medical diagnostics, autonomous vehicles, and more. Now, as I mentioned, they released these models as open source, and it's very similar to what meta is doing in two different aspects. One. It's a very capable model that they released as open source. But the second is they can do that. They can release very highly capable models as open source because they don't need the revenue from the model. So if you think about Yeah. Open AI when they're releasing a model or Anthropic when they're releasing a model, their livelihood of the company depends on the model making money, at least in the long run. And that's also questionable. We talked about this, in the past few weeks, but. Meta doesn't need this revenue and the same thing about Salesforce. So they can develop these models, benefit from the outcomes of the research in the things that they're doing in Salesforce software, in Meta it is their social media stuff, and still release them to potentially compete with the other models and put them in a disadvantage because the other companies actually need the revenue. So this model was already released and it's available on Salesforce GitHub repository, repository, and you can go and grab it and use it yourself. We're going to go back to the topic of what does that mean from a business model perspective in just a couple of minutes. Another company that we don't talk about a lot that released three interesting models this week is AI21. AI21 has been around for a while, also releasing open source models. And they have released Jamba 1. 5 family of models, they have Jamba 1. 5 mini and Jamba 1. 5 large, and it has long context. It has a long context window of 256, 000 tokens, which makes it right now the longest context window of any token. Open source model. The only two that compete with it is Claude three Opus that has 256, 000 tokens, but it's a closed source one and Gemini 1. 5 pro with it's 2 million tokens context window, but from an open source perspective, they are far ahead of everybody else, at least doubling the other top models when it comes to context windows. And they are claiming that it maintains the performance throughout the entire context window. So this is another claim that Google has for themselves, but the other models are known not to do it. So as an example, if you're using cloud three and you're getting close to the edge of the context window, it will start hallucinating more quality of the results is going to decline as you get closer to the context window. The same thing with open AI. So they're claiming that you can use the full context window while maintaining the performance, which is Very powerful, especially for an open source model that you can use for free. They're also claiming that it's two and a half times faster than the previous model, but also the fastest in their class of size across all the other tools in the same size. And it's multilingual. It currently supports English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, and Hebrew. So including two right to left languages, which is not common through all these tools. On the developer side, they have structured outputs like JSON output and function calling and document object digestion. So it also built for creating code and working with code as well. So very powerful, all around solid contender in the open source world. And it's available already on the A21 Studio platform, Google Cloud, Microsoft Azure and NVIDIA NIMH. And they're going to release it on Amazon Bedrock. And databricks and snowflakes and probably more in the future. So another powerful, very capable and fast open source model in this case with a very long context window. Speaking on long capabilities of models. The Tsinghua University in Beijing just released LongRider, which is an AI system that can produce coherent text of over 10, 000 words. even though the context window is very low, the output of each specific question is highly limited on these models. So while the context window of some of these models are very long, the output of each specific question is highly limited on most of these models, and they were able to train a model on 6, 000 writing samples that are between 32, 000 words long in order to create a model that specializes in writing coherent long text and they're claiming it outperforms larger proprietary model like OpenAI and so on that are more generalized tools versus tailored specifically to writing long content. This is obviously very helpful for publishing and writing first drafts of books or reports or marketing in writing papers and studies or education and research and so on. Very capable and it's just showing you that as we move forward, we're going to get more and more of these smaller models that are geared towards specific tasks, and they're going to do them very well, most likely with some kind of an agent that is going to control them, and he's going to pull the right models for the right tasks, getting the best results across multiple aspects of the bigger tasks that is required. Now, I told you we'll be back to talk about the business model, and I shared with you my thoughts on this. In the past few months, that as these models are getting better, faster and cheaper, it's going to represent a very big question mark on the viability of just selling access to these models. So in an interview with Cohere CEO, Aiden Gomez, he shared the same exact concern. He said that AI, that the current AI models cost of development is significantly higher than the potential revenue that they can generate. He's also claiming that the bigger company are basically financing their losses, which means what we are paying for these most advanced, even smaller, cheap models is less than it actually costs to operate them. And they're only doing this because of the fierce competition and they want to get more market share. And that means several different things. One is that these companies will have to focus on the application layer, basically developing tools that companies can use versus selling just access to the model themselves, and then there's more value. And then there's ways to charge money for them versus just. Come and use my model as you wish. The other is that smaller companies will find it harder and harder to compete, which I think is a big shame. But the reality is that if you don't have access to billions of dollars, which is what it costs to train and then run these models, your chances of survival are very low because your chances of generating revenue that will justify this expense. It is very low. So the future of this, while right now there are dozens of companies who are in this game, is going to be probably a handful of companies that are going to keep developing at least the frontier models that does leave room for companies to develop unique, smaller models who become very good at specific things that provide value on those specific topics. And hence they will be able to sell their services for the right companies and still make money. Speaking of interesting tools, HubSpot just released what they call the AI Search Grader tool. What it does is it can help brands and companies understand their presence in AI power search and large language models. So we are moving from an era of SEO to an era of large language models. Of language model optimization, LMO, because more and more traffic is going to go from traditional SEO to just getting answers or large language models. So if you want to understand how your company, how your brand is going to potentially appear in large language models, there was no tool available to the public so far until. The introduction of this tool by HubSpot and the key features of that tool is it's going to give you an overall grade for your company's ability to appear in large language models, brand sentiment score, share of voice score, and personalized analysis specific to your website and to your industry. Now, as I mentioned, this is the first tool of its kind, at least that I know of that is available to the public. It is going to help companies prepare for what's coming in the change between SEO and LMO. And it builds on the same exact trick that HubSpot did shortly after they launched HubSpot. So those of you remember, I was one of the earlier adopters of HubSpot, when they just came out and they had a tool that was called Website Grader. And it allowed you to put in your URL and it will give you a very detailed SEO analysis. And it was a lead magnet. So millions of people put their email addresses into that tool in order to get the report. Now, it wasn't just a lead magnet. It wasn't tricking people. The report was actually highly valuable and obviously suggested ideas on how you can use HubSpot in order to generate content, in order to drive people to your website. And I assume they're just replicating the play that was very successful to them back then. to this new era. HubSpot is known to be an advanced company. They have made several different announcements on how they're going to integrate AI into their platform. The announcements were amazing. Most of it was not materialized yet, but I assume they're moving in the right direction to adding a very powerful AI layer to their tools, but this is just another way for them to provide value. For free while capturing attention and getting leads in order to bring more people into HubSpot. But if you want to learn how your company may or may not appear in large language models, you now have a tool that you can at least try out. And the last thing I'm going to mention today is that the World Robot Conference Or 2024 kicked off this week in Beijing, it was absolutely nuts. if you want to be shocked with what's possible with robotics today, go and watch some videos from this show. It had robots doing anything from Playing tic tac toe opening and serving drink cans to complete humanoid robots that can do backflips and push ups and so on. And it's showing that the advancements in AI on the cognitive side are walking hand in hand with the advancement in humanoid robotics and other robotics in this particular case. And the combination of these is going to put at risk a lot more jobs than just AI. white collar ones because they're already robots today that can do more or less most tasks that we can do in the physical world as well. interesting mention is Unitrix G1 robot, which we talked about A couple of months ago when they announced it, but now they're actually revealing that they're going to mass produce this. And they didn't say exactly when yet, but they have a price for it. So the G1 robot is only four feet tall. It's significantly smaller than it's bigger brother, the H1 model. But the interesting thing is this G1 robot that again, they're going to mass produce probably into 2025 is going to cost only 16, It's not as sophisticated as some of the other robots. It only has three fingers instead of five, but it can do many tasks at home that do not require heavy lifting. It can climb stairs, it can go around objects, it can lift small stuff, and it can handle delicate and specific precision operations. So this is already putting it in a price range that is reasonable to many Households, and this is just the first version of that. as I mentioned, go and watch videos from this conference. It will blow your mind what robots can do today. And you combine it with everything else that we're talking about in this show, as far as artificial intelligence capabilities. And you will understand that we are going into a world that we don't know, like total sci fi with robots walking everywhere, doing tasks for us across multiple aspects of operations from manufacturing stuff, but all the way to serving coffee to us in coffee shops and cleaning our streets and mowing our yard and doing our dishes and laundry, hopefully sometime soon that is going to happen. And it's not going to happen a decade from now. It's going to happen three to four years from now when these robots will stop popping up more or less. That's it for this week. If you enjoy this podcast, please like it. Please give us a review on your favorite podcasting platform. Pull your phone up right now and give it whatever star review you think we deserve. Hopefully five and write a comment. And while you're at it, and you have your phone in your hand and the application open, share it with a few people who will benefit from it. This is your way to help us educate the world on AI, so hopefully we can all benefit from the pros of it and try to minimize the negative impacts. I will see you on Tuesday on another fascinating expert episode. And until then, have an amazing weekend.

People on this episode

Head Shot

Isar Meitis

Host