81 | AI summit at MIT - I took notes for you at Imagination in Action this week and more AI news for week April 20

0:00Hello and welcome to Leveraging AI, the podcast that shares practical ethical ways to improve efficiency, grow your business, and advance your career with AI. This is Isar Maitis, your host, and this is a short weekend news episode of the podcast, or it is because this is a very unique episode. I'm recording this in a hotel room in Boston, Massachusetts. After attending the Imagination in Action Summit put together by MIT yesterday and the day before, the summit was absolutely amazing. The top leaders in AI across everything from investment to research to company leadership to technological advancements.

0:40Literally the smartest people in AI in the world were going back to back on stage, either as individuals or in panels in several different channels. And I really want to share with you what I've learned in that summit. So we're going to focus on that, but let's start with the biggest news of the week that I just cannot skip. And then we will dive into what happened at the summit.

1:07Let's start with the biggest news of the week or of the day. Yesterday, Meta has released Llama 3. It was highly anticipated and it was known that it's coming in the next few weeks. I spoke about this in the last couple of episodes. So the model was released yesterday on Thursday while we were actually speaking to Jan LeCun, who is the head of AI at Meta. And he shared with us some of the details. But the details are this. They released two models so far, one with 8 billion parameters. the other with 70 billion parameters. They are claiming that the 8 billion parameters model is better than the 70 billion model of LAMA 2, which is obviously a huge improvement in performance that only gets you to think what does the 70 billion parameter model can do.

1:54It offers improved capabilities in things such as reasoning and code generation, and it is available basically anywhere you want. They released it as open source as they have with all their previous models, and it's available on AWS and Databricks and Google Cloud and Hugging Face and Kaggle and IBM Watson and Microsoft Azure and NVIDIA NIME and basically anywhere that you can get access to models. It is available and it is supported on all the major chip manufacturers. Meta also announced that they're prioritizing responsible development as part of LAMA 3 and they're introducing new ways to keep those safeguards.

2:33And it includes tools such as LamaGuard 2, CodeShield, and CyberSecEval 2, multiple aspects of layers that is supposed to make the usage of Lama 3 safer for all of us. Now, if that's not enough, Meta shared that they're planning to introduce additional capabilities such as longer context window and enhanced performance in the next three months. And they're also planning to release a multimodal version with over 400 billion parameters. Again, different than the 8 billion and 70 billion that are released right now. And they are all coming in the next few months. So this is the biggest piece of news of this week.

3:13LAMA has been one of the best open source models so far. And there's no doubt in my mind that LAMA 3 will take the top of the list in at least the near future as far as the open source models. It will be very interesting to see it's benchmarked in the real live world against the top models right now from OpenAI and Cloud. And OpenAI just released GPT-4 Turbo into two all paid ChatGPT users. So they've announced that back in November, ChatGPT-4 Turbo allows to run a much larger context window. So 128 ,000 tokens versus 32 ,000 tokens of GPT-4. This is the biggest difference. The other differences are significantly lower need for compute and power to run these models.

4:02So basically everybody wins. We get a faster model that is also running a much larger context window. And Earth has to pay a smaller fee when it comes to the carbon footprint of this process. Adobe also made some interesting news this week. First of all, they are releasing an AI assistant for the Acrobat software, that is the PDF creation and viewing software. It costs$4.99 a month, and it can help you read, understand, and summarize PDF documents. I must admit, I don't see a huge benefit in using that because you can take PDF documents and upload them to Claude or ChatGPT and do all these things over there.

4:40But if you are an Adobe user and already using a lot of PDF stuff, that might be worth paying for. The more interesting thing that Adobe released that became viral across every place if you're consuming any kind of AI content is AI video capabilities for Adobe Premiere. You can now use these capabilities to extend the video. So you have a video that ends at a certain point and you want to add a few seconds in order to align it with the music or your needs and so on. You can literally generate frames out of thin air that will seamlessly stitch to the end of the segment you had before. You can add and remove objects from the video.

5:18You can change the type of clothes people wear in specific scenes, and you can generate video from text directly within Premiere Pro. They're also planning to integrate third-party generation tools into Premiere Pro, such as OpenAI Sora and movie generation tools like Pica Labs and Runway. So this entire ecosystem of Adobe Premiere that is probably the most commonly used post-production tool in the world today is getting a lot of AI attention. If you haven't watched the videos that show these demos, definitely worth watching. And the last piece of news I'm going to dive into, the Imagination in Action event, comes from Microsoft.

5:59Microsoft just released an AI model they call VASA1, V-A-S-A-1, which is a model that can turn a single image into deepfake videos in a really quick turnaround time, 170 milliseconds. And these are still not very high resolution videos. They are 512 by 512 resolution, but they're running at 40 frames per second. And you can upload a face of a person and upload a audio file of a person speaking, and it knows how to create the face motion and the lip syncing in a perfect way. It looks completely realistic. This is similar to tools like HeyGen, only it looks even more realistic than HeyGen. And it starts just with an image and not a video of you as the spokesperson.

6:46This is obviously a really impressive capability that has a lot of useful benefits if you are a content creator for any purpose, whether it's internal or external, but it also comes with very serious fears of generating deepfakes very easily and with literally just an image of a person. So now let's switch and talk about the summit at MIT. The summit was absolutely mind-blowing. First of all, in the level and quality of the people that participated, whether in the audience and on stage. Literally the smartest people on the planet and the most influential people on the planet when it comes to AI were there, either in person or through a Zoom call.

7:29I will name just a few. Yann LeCun, the chief AI scientist for Meta, was there. Aravind Srinivas, the CEO of Perplexity.ai, was there. Ramin Hassani, the CEO of Liquid AI, and his team were there. Dinesh Meheshwari, the CTO of Grok, was there. That's Grok with a Q, the cheap manufacturer, not Grok with a K, the AI model by X.AI. Steve Case was there. The founder of AOL was speaking there about the internet revolution and its similarities and differences than what happens now with AI. Vinod Khosla was there, one of the most successful venture capitalists in the tech world today. And the first external investor in open AI was speaking there.

8:13and many other influential people in the AI space and beyond. When I say beyond, people from the financial world was there speaking about the impacts on the financial world. Michael Wise, the CTO of Universal Pictures, was there talking about the impact of AI on entertainment and so on and so forth. Again, the most influential and smartest people on AI on the planet. The amount of AI and computer science PhDs that were in each room were probably around 50 % of the people at any given point. Literally incredible. So now let me share with you, first of all, my biggest stakes, and then I'm going to share some of the specific things that some of the specific individuals have shared.

8:55So the first thing I want to jump into, which is maybe counterintuitive when we're coming to talk about AI implications, is human relationships or the human aspect of the AI revolution. And I shared that in many cases before, and I talk about this a lot in the courses that I teach when I do consulting, but human relationships, human collaboration, and human innovation are still the most important aspect of everything we do, including business, including the AI revolution. It was very clear that the achievements that are achieved today in AI are all built around collaborations between really smart people and that anything you want to do in the future, the human relationship is going to be a very important factor of it because a lot of the other stuff, everybody will be able to do pretty damn well because they will be using AI tools to do these things.

9:48So if most of what business can do, everybody can do pretty well, the differentiator between one business and the other is going to be the human relationships that these people have and their ability to collaborate with other people. Another very important aspect of this is what skills are going to be the most relevant in this new future with AI. And in this particular case, I'm going to quote Vinod Khosla that said that the most important skill in this future is going to be adaptability or the ability to learn how to learn. Because the world is going to change significantly faster than we're used to, The only way to stay updated is to know how to continuously learn new skills and new capabilities so you can adapt to what the new norm is.

10:37And I agree with him 100%. This is obviously a very tough question for anybody who has kids on what do you push your kids to learn, whether it's at school or after school activities or in their college or post grad degrees. And there isn't a clear answer. But again, the general idea is we will need to learn how to learn and adapt a lot faster as humans. The next thing I want to talk about is the concern about the future of society and humanity with AI and its implications. The good news is that every one of the large companies have their own internal concerns about the potential risks of the things that they're developing, and they're investing a lot of resources and putting a lot of effort and try to reduce those risks.

11:26That includes relevant teams that are reviewing everything that they're doing. That includes developing technology that can safeguard what can and cannot be done with some of these tools, etc. There's obviously a very big debate between the open source approach of let's open this to everybody so everybody can see what's going on. So when people are making changes to the code, we can see what they're doing and it's open to the public versus the company that are saying this is too dangerous and we should allow few companies to control this. I'm not going to state what my opinion on this is. All I'm going to say is that really smart people are on both sides of that fence.

12:03But as I mentioned, the good news is whether it's open source like Llama 3 that just came out when we were speaking to Yann LeCun, he specifically said that they're putting a lot of efforts into making sure that it's used safely and pushing his belief that if it's open source, and everybody can see what's happening and that keeps us safe. Obviously, there are people there from OpenAI and Microsoft and so on who are doing it exactly the opposite way. And they're putting a lot of efforts in keeping those systems safe. We had an entire session with Alexander Alex Madry, the head of preparedness in OpenAI.

12:34And he was speaking about the amount of efforts that goes in OpenAI into keeping these models as safe as possible. I must admit that while I'm very happy that these organizations care about these things, The flip side of this is that they're moving very quickly to develop and deploy better and more capable models. And I'm sure that within that race, safety is not the top priority, even if it is a priority. So I truly hope that we're going to see more collaboration between these companies on the safety aspect of this. And I truly hope that the government and governments all around the world will jump in order to enforce some rules and guidelines and regulations on what AI can and cannot do in order to try to really allow us to enjoy the amazing benefits, which I'm going to talk about shortly, that AI can bring while reducing and maybe even eliminating the risks.

13:29There was a fascinating session about required breakthroughs in AI in order to make the next step in that. The people on the panel were senior people in various aspects of this field, including Raj Agarwal, the general manager of Gen AI at AWS. Dina Khatabi, a professor for that in MIT. Alan Chabra, the executive vice president, worldwide partners at Mongo, and others. They shared some of the limitations and hence the breakthroughs that they're expecting to happen in the near future to overcome these issues. So one of the issues that they mentioned was that RAG, which is the ability to extract information from documents and or information that you have, still has limitations.

14:14And it's not good enough even at the enterprise level, not to mention beyond that. There's still hallucinations, even when you're referencing specific data, and that obviously hinders the deployment of that in larger enterprises and across more safeguarded data where accuracy is important. And he's expecting those problems to be resolved in the near future. They spoke about the breakthrough in code generation, whether using co-pilots or by using natural language in order to code computers. This is something that was mentioned in several different aspects and in several different conversations.

14:50The concept that computers will adapt to humans versus the other way around, meaning instead of us having to learn how to write code, computers will understand our universe through natural language, through voice, through vision capabilities that will be installed on wearable devices. And we will be able to program and interact with computers naturally, just like we interact with other people. and the computers will adapt to that. One of the benefits of this is obviously an explosion in innovation because if any person on the planet can quote unquote create new software that does something specific, it will allow more people to be creative from a co-generation perspective way beyond the amount of people that are doing this today.

15:35There are several different companies who are working on solutions like that on basically allowing each and every one of us, any person in the world that has access to the internet to be able to write code and create software. Ash Keller, who is the CEO of Neurocity, mentioned that he's looking forward to a breakthrough that will allow foundation models to understand human brain activity in order to connect with us on a very basic level and in order to help us to create cures for existing brain damage, whether through age or through various injuries. Raj from AWS mentioned something that, again, if you've been in this field for a while, you would already know that data cleaning and data access is the biggest problem of the AI adoption today.

16:21Data in most companies is messy and it's in multiple places, in multiple formats, and bringing it all together and cleaning it up in a way that AI can actually use it effectively is 80 % of the work in most large AI projects. And he's looking for the breakthrough that will allow generative AI or other AI capabilities to do the AI data cleaning on their own, which will save a huge amount of work and will allow us to harness and mine significantly more information a lot faster. This is true, obviously, for society as a whole, but also for each individual company or enterprise that wants to implement AI solutions.

17:02They also mentioned a very important point when it comes to the struggle between the wish to run forward faster while keeping everything safe. And when I say everything's safe from the model creation perspective, it's obviously keeping us safe from what the model can do. And from a company perspective, that's obviously safeguarding the data that they own, whether it's proprietary data or private data or customer data and aligning with regulations of their industry while they're doing so. And almost everybody in almost every conversation, there was a discussion about the huge demand for compute on this path, the amount of energy that it requires and the amount of money that is spent on computer chips in order to achieve the next level of models and in order to get to AGI.

17:45Yalne Kuhn mentioned that Meta is going to invest$30 billion in buying additional NVIDIA chips in order to train their next huge 400 billion parameter model. And the implications is obviously not just money, it's also the amount of energy that these computers require and hence the amount of carbon footprint that they leave on our planet. There has to be a breakthrough in that. And two different companies that were represented in the event are working in that direction. One of these companies is Grok, again, G-R-O-Q, and they have developed what they're called the LPU, Language Processing Unit, which is a completely new innovative chip design that can run between 10 to 1 ,000 times more efficient than NVIDIA GPUs, which means they are running with significantly less compute to get to similar results.

18:36Now, in addition to running significantly faster, Grok also has some huge benefits from a performance perspective. Grok chips currently generate significantly more tokens per second than any other platform right now by a very big spread. They're also generating the first token significantly faster than any other platform. So you're getting very, very short latencies, which will allow to have much more natural communications with large language models than other AI models that will support what I just said earlier, which is that seamless interaction with computers and AI systems. The other company that is at the tip of the spear of promoting a new kind of AI that will dramatically reduce the need for compute and will be a lot more efficient is Liquid AI.

19:24Their entire founding and leadership team was there in one of the panels. And what they are doing is they're developing a new kind of AI infrastructure. So they're not using the transformer model like everybody else is using. They've actually developed a new kind of model that they are calling Liquid AI. And Liquid AI is a new architecture that supports explainability and higher efficiency. What do I mean by explainability? You can actually understand what the model is doing and why versus the black box of the current architecture, which allows faster development of new models because you can actually understand what the model is doing.

20:03So it provides transparency and significantly higher efficiency. It requires 10 to 20 times less cost and energy to train a model and 10 to 1 ,000 times higher efficiency when at inference, which is the generation of tokens, basically the responses you're getting out of these models. What they've developed is a developer package and the core models that they're giving to companies in order to develop their own in-house AI solutions. Because these models run significantly more efficiently, companies can run them in their own environment, even on-premise computers. And that means that their data and the models themselves never have to leave the company's organization closed loops as it's running today, which obviously provides a significant benefit when it comes to data safety and security.

20:52And because it's a lighter model, it can also run on device in the future, including much smaller devices, which will allow AI to be significantly more distributed. This concept was raised by several different people on various panels on the need for balance between on device AI capabilities to cloud AI capabilities, and that the balance in the future is going to be a lot more towards on device capabilities versus cloud capabilities for obvious reasons. A, data security, B, latency, like a lot of the assistants that are coming are going to be running on our local devices, whether they're going to be wearables or our own home or company computers or our phones, etc.

21:35One of the more inspiring aspects of this summit was that there's a huge number of really smart people who are working to use AI to solve some of the biggest problems the human race and our planet has, such as curing cancer and other diseases, fighting viruses, supporting the elderly against diseases that they have, such as dementia and others. And they're doing it in incredible ways. To give you an example, Jeffrey von Maltzen, who is a brilliant biology scientist who is betting on AI to solve a lot of problems, him and his team are working on understanding how the actual structure of proteins is building various viruses and so on so they can build drugs down to the molecular level so they can attack viruses in more advanced ways that we can do today.

22:26He gave some astonishing numbers. As an example, he mentioned that 90 to 95 % of the time, we fail to predict which molecules are going to be a good drug that solves the problem it's trying to solve and does not generate really high risks. With the help of AI, we should be able to drive the failure number by something. He's saying, let's say we drive it down to 50 % instead of 90 to 95%. We are changing medicine in the world completely. One of the interesting exercises that they did, and I'm sure I cannot explain it as good as he did, but they took a huge variety of existing drugs and used AI to try to help target the same locations and the same methods that these drugs are addressing now, but not in a way that will infringe on the IP of the companies that have created these existing drugs.

23:16And they were able to do this across all the drugs they tried it for. And that process would have taken them a decade to do in traditional methods and took them three months using AI. What he believes, which to me is the most inspiring thing that was said all day, is that in many cases, the AI-based research will allow us not to only find new treatments, but to actually find cures to some major diseases that we have today. Surprisingly, there were not a lot of discussions about AGI. It was mentioned several times, and actually two of the leading scientists and experts on this field actually agree that the current models that we have probably cannot achieve AGI.

24:00One of them is Stephen Wolfram, who has been in this field of computer communication and computer systems for the last 40 years. And he thinks we are not there yet. He believes that these models, while they're very good at guessing next words, are literally doing just that. And they don't really have an understanding of our world. and that language is actually easier than what we thought it's going to be and not as easy as understanding how the world around us works. So he thinks that there's a limitation to how large language model can progress and hence he believes that in their current shape we cannot achieve AGI.

24:42Another person who has been loud about the inability of large language models to achieve AGI is Jan LeCun. I've shared that with you several times before. In the conversation with Jan One, he's saying that the current large language models cannot achieve AGI because they do not understand how our world operates. They cannot reason. They cannot plan. And they do not have long-term memory. And hence, at Meta, they are taking a new approach to how to teach the next version of models. They are trying to teach models the way babies learn about the world. They're calling this new approach VJEPA stands for Joint Embedding Predictive Architecture.

25:19doesn't matter what it stands for, but it's a way to allow these models to learn just like babies and animals learn about the world. And in their initial testing, it's achieving very positive results as far as actually understanding how the world around us works versus just having a perception that they understand it because they can guess words. And he's saying that because these models will understand how the world works, they will be able to make predictions about the world and hence we'll be able to plan based on real world limitations and capabilities, which will allow us to achieve AGI and do a lot more with these models than what we can do right now.

25:59And because we're speaking of Yann LeCun, I want to mention the fact that more or less everybody on every single panel there, even the people working for the closed source companies like OpenAI, Microsoft and so on, have all shared that the open source world of AI is critical to the success of this world, partially because it's allowing significantly wider access to the models and a lot more brainpower can go into develop and monitor what the next generation of models can do, but also because of cost. Several different people, including Arvind Srinivas, who is the CEO and co-founder of Perplexity said that currently the cost aspect of large language models leans very favorably for businesses to run open source models as the backend of their platform versus the closed source ones.

26:51And what he's suggesting is something I'm doing with several of my clients is to test things out across multiple models and then pick the one that gives you the best cost efficiency combined with sufficient results, right? So they don't have to be the best results. They have to be results that are good enough for your needs and then optimized for cost efficiency, which in most cases will end up being an open source model. Another good piece of news is that the entrepreneurial spirit is alive and well, including in the AI world. This particular event had over 100 startup founders, 80 of them started pitching their businesses at 7 a.m.

27:33in the morning in front of an audience of a few hundreds of people, and each and every one of them had 90 seconds to present their pitch. So people flew in from all around the country and some of them from overseas in order to give a 90-second pitch about their AI business. It was really inspiring to see and be in that room and learn about the huge variety of ideas on how AI can be implemented across industries and across different aspects of our day-to-day life and business. In the conversation with Michael Wise, who is the Senior VP and CTO of Universal Pictures, there was a very interesting conversation about the impact of AI tools on the future of entertainment and specifically the movie industry.

28:14And Michael shared that they are using more and more of these AI tools in their work. He does not think it's going to have a significant impact on the future of movies as a whole. He thinks that there's very little risk that these tools can create movies in the quality of Hitchcock or Christopher Nolan, and that the key of the movie is the aspect of storytelling and not just the production of the movie. I must admit that while he obviously knows this industry a lot more than I do. I think he is either not sharing what he really thinks because he's in charge of a lot of people in that industry, or he is somewhat wrong.