46 | AI creating new AI on it's own, EU agrees on AI regulation, an AI generated news channel, and more AI news from this week

0:00Hello, and welcome to a short weekend news edition of the Leveraging AI podcast. Yes, this is Isar Maitis, your host. There are some really impactful and really important news for the progress and advancement of AI in general this week. So let's get started.

0:24I want to start with something that to me was somewhat alarming, even though to the people who are sharing it was very exciting. a company called AI Zip Inc, together with scientists from MIT, shared earlier this week that they were able to generate new small language models using a large model without human involvement, meaning they were able to use the computing power and the understanding of a large language model like ChatGPT4 in order to create smaller, very task-dedicated language models without human assistance. The researchers were obviously really excited about their finding. They're saying that such small models can do things such as identifying human voices among ambient noises, monitoring pipeline data to prevent integrity issues of those pipelines, and using different sensors in the wild to track wildlife.

1:19So positive things. But the reality is that this is a very scary, in my eyes, step in the ability of AI to create new AI. Now, they're saying the goal of these small AIs is to be supportive of the larger AI creating an ecosystem of something that can be more capable in these smaller tasks. And that to me raises huge fears. And I know a lot of people do not believe that an AI model can go rogue and start doing things that humans may not agree with, I think this might be one of the first seeds of something like this actually happening. If an AI system can create its own AI systems, and right now they're small and right now they're very specific in their capabilities, but if the concept is possible, then creating new large language models that will be better than the large one, or just the large one creating an unlimited amount of these small ones to do whatever it wants is becoming a possibility, which to me is not necessarily good news.

2:23And this makes the second piece of news even more important. The European Union finally agreed on the final language of the Artificial Intelligence Act that they have been putting together. This is a landmark set of rules that is probably from its size, the first of its kind. Their goal is to limit the usage of AI and define what is an ethical use of AI. It would ban things such as social scoring and inferring with political news and or religious issues. It bans facial recognition databases from being used across these platforms. It even limits what law enforcement can do with biometric IDs and stuff like that.

3:09And they require transparency from the creators of such models, companies like OpenAI and Google and so on will have to share their training data and be compliant with copyright rules. They've also defined the mechanism on how to judge these models based on their level of risk, and they monitor that based on the size of the model from a computing power perspective. The only one that is defined right now as the highest risk is ChatGPT4 that has passed what they've defined the systemic risk threshold. And such models that are defined as high risk must sign specific code of conduct that will allow additional oversight on what these models can do.

3:55Is that good news? Yes, I think regulation of AI is going to be extremely important in the future for multiple reasons, including potentially going back to my previous piece of news, the actual existence of mankind, at least in the way we know it. But that being said, there's a few issues with that law. One is that, first of all, it hasn't been voted on. So while there's a final agreement on the language, it's not a law yet. Two, the law is supposed to become effective only in 2025, which raises the question, what happens through this entire year of 2024 that's ahead of us? And the last thing is open source models are exempt from some of these rules.

4:35And the goal of that is obviously to drive innovation within the European Union to compete against some of the large language models from other places, as long as they are open source, which is not necessarily a good thing because that means you can release to the public models that can do things that are extremely powerful, like the Mistral models that are developed in France without the same oversight and regulations in place for the closed models. But to make this even more interesting, we see in the past few weeks more and more companies and organizations releasing very capable smaller models.

5:08So one of them is Microsoft. Microsoft just released their PHY 2. So they've been releasing these PHY models for a few months now. And this is version 2. They released version 1 and then 1.5 and now version 2 of it. It's a small model. It has only 2.7 billion parameters. It's only language, so it's not multimodal like all the new large language models. But it can run on laptops and phones, and it achieves on many of the benchmarks better results than Lama 2 and Mistral, which are the leading open source models right now. It also performs better than Google's newly released small model that is supposed to run on such devices called Nano.

5:49So far, they have released this model only for non-commercial research purposes. But how do you exactly monitor who is grabbing this open source model and how they're using it? And the reason it's relevant is it's showing that the advancement in AI enables us to do a lot more with a lot less. On one hand, that's very good. It means less computing power and less resources required to run these models, which is good for the world from a greenhouse and global warming perspective. But on the other hand, it means that some of these regulations that were just put in place or not even may not detect such models as high risk because they're much smaller and require smaller technological footprint, but they're highly capable.

6:33And again, combined with what we discussed in the first piece of news, sometime in the near future, these large language models will be able to create their own smaller models to go undetected by these regulations. I apologize that I'm starting with somewhat negative news this week, but I think it's important that you understand what is currently happening. And to stay on topic with the open source world, Meta just released what they called Audiobox. And Audiobox, as everything else with Meta, it's an open source model. And what Audiobox can do, it can replicate human voices as well as other noises and music.

7:13So it knows how to clone voices from recordings. We had this capability with several different models before. The most known one is 11labs. It's a service you can sign up for yourself and clone your own voice. You obviously sign that you know what you're doing, but the reality is it can clone anybody's voice and it does it extremely well, including replicating arms and changes in tone and so on to really sound like specific individuals to the point it's impossible to differentiate between them. And even companies like HeyGen knows how to do this, including videos of the people. But this is now an open source version that can do the same thing.

7:50The other thing it can do is replicate different sounds from nature, such as birds chirping and dog barking, and it can do it very convincingly. Same story from a licensing perspective. It is only released for non-commercial research use at this point, but we all understand how quickly that can go to not the right places. and not the right hands. But as I mentioned, this is the world we live in right now. So you need to be aware that cloning and replicating your voice is very easily doable. Another very interesting and very impactful piece of news that was released this week, there's been a video of Mark Zuckerberg from Meta testing real-time AI combined with their Ray-Ban glasses.

8:36So those of you who don't know, So Meta has a partnership with Ray-Ban where they release their glasses that has a video camera and some computing power in it that allows you to stream what you're doing into your social media. But now they're adding the capability to have a personal assistant that is built into the glasses that you can kick into action by saying, hey, Meta, so similar to, hey, Siri, or hey, Google, and all of those other assistants. But the biggest difference is that this is a wearable device, meaning you'll be able to wear glasses that see the world you see in real time and give you aid and information and assistance through voice and potentially probably in the future imaging as well that will be projected on the glasses to the world you see around you.

9:23It also knows how to do real-time language translation, meaning you can be in another country and have a very effective cross-cultural interaction with people who do not speak your language with real-time translation to both ways. Now, while this was just a demo, it gives us a glimpse into what's coming in the future. And these are capabilities that we've seen in the movies, such as Terminator and Iron Man, where the person that's wearing this device can interact with a computer that has a huge amount of data that can analyze a lot of information. And as we know from other things in AI, take actions as well, based on interactions with the real world.

10:00That is on one hand, extremely exciting, but like everything else with AI can be extremely troubling as well. So these are early steps in that direction, but with the speed this is moving, I'll be extremely surprised if within this coming year of 2024, we will not start seeing actual applications and actual devices that we can purchase and use in order to do these kinds of things. Meaning in the very near future, we will be able to use augmented reality devices that understands vision and speech and sounds of everything around them to provide us information and help guide our decisions across multiple things that we do.

10:42And you can decide for yourself whether you think that's good or bad. And continuing with some big releases and big news from the big players, Microsoft announced this week that they are adding new features to its co-pilot. The first one is that they're adding the multimodal capability into Bing and their chat interface, which means you'll be able to add images and not just text as part of your interaction with the model. They are adding GPT-4 Turbo, so it's the upgraded version of OpenAI that so far was only available on ChatGPT, is now going to be part of their search queries for long and complex queries to provide more accurate capabilities.

11:21and they're adding DALI 3 image generation to Copilot as well. So three capabilities that existed so far only on the OpenAI side will become available through the Microsoft ecosystem. It's not a surprise. It's just an upgrade to anybody who is using Microsoft and is watching what's happening in the OpenAI side and wondering when it's coming. Well, it's coming pretty fast. So I don't think it's been a month since the announcement of these capabilities in OpenAI. And now they're already available for Microsoft users and for free. Another big news came from Google. So Google released Gemini last week.

11:56We spoke about this in last week's news. Gemini is the new multimodal language model by Google. But this week, Google announced that Gemini Pro, which is their mid-tier model between nano and ultra, is going to have API access for free for developers who are using Google Cloud. This move, I am certain is done very quickly as soon as they announced the model in order to provide more attractiveness to the Google Cloud platform that was kind of left behind, behind AWS and Microsoft Cloud with the very aggressive moves that they have done with adding AI capabilities to this. So if you want to use a very powerful model that is defined around the level of GPT 3.5, You can now do this for free using Gemini Pro through Google Cloud Interface.

12:49It supports conversational capabilities in 38 languages, and it also provides SDK capabilities for developers in Android, Node.js, Swift, and JavaScript. So very powerful capabilities that are available for free in Google's attempt to start attracting people for AI capabilities. I must admit that like everything else that Google is doing in the AI world, they're very far behind of what's being offered on AWS by Amazon and Azure by Microsoft. But that being said, and I said that every single time when I talk about Google, they have more data and more computing power and more resources and an incredible research lab to catch up to everybody else and potentially surpass everybody else.

13:34So I'm not by any means stating that they are losing this game. I'm just saying that they're behind in the current round of the game. One interesting piece of news from the text to image generation world, Me Journey, the most advanced AI image generation model, just released their alpha access to their website that will eventually be Me Journey version 6. They have hinted that this is coming in Q4, and now they provided this alpha access. In order to get access to this, you need to be an existing user of MidJourney through their Discord interface, which was the only way to access it so far. But you need to be a heavy user that has generated at least 10 ,000 images in order to get access to this alpha.

14:22In this website, they offer significantly simpler user interface that does not require you obviously to go through Discord and get access to some of the capabilities in a much more user-friendly way, which will open MeJourney to probably a much larger range of users. That being said, it's missing some of the capabilities that advanced users loved to use using the Discord server. I love MeJourney and I use it almost daily and I'm in several different forums and I saw very interesting conversations on some people that are saying that despite the fact that this is really cool and nice, they will not transition over because the missing capabilities and the advanced capabilities that are available on Discord.

15:02The other benefit of Discord is obviously the community aspect of it that allows people to see what other people are doing and to share the results that they're achieving and to share different aspects of their work, which allows people to advance faster. Where do I think this is going? I don't know. They did not announce yet whether they're going to kill the Discord side of things. I'll be surprised just because the push from the community to keep that. I will also be surprised if they do not add these missing features once they actually release the version 6 to the public. So time will tell, but it's a very interesting and positive step forward from mid-journey to make the most advanced image generation model to the wider public.

15:44And the last piece of news is the news. A company called Channel1.ai just announced that they're releasing a new news website that will generate news using AI. So the actual people on the screen delivering the news are going to be AI generated. They are saying that the benefit of this, that they will be able to personalize the news and the newscasters to every specific individual. So you heard me talk about this several times before on the show, when we were talking about capabilities of AI to generate specific personas and how personalized the world is going to be when it comes to marketing. Well, now this company is going to implement the same exact concept for delivering news.

16:30So the newscaster can be female, male, whatever origin of the world, speak whatever language you want, and even provide localized news that will only be available to a very small number of people because they don't have to actually film anything. They're also saying that they'll be able to create video newscasts of events that actually happened, but that had no camera present at the point they can regenerate it or a variation of it, I guess, with AI. This raises some very, very big questions, because if you can do that, it's very obvious that you can create news that did not happen with AI. So the question becomes in, where do you draw the line between what's acceptable and not acceptable when it comes to delivering the news.

17:17Now, we already have a serious problem of fake news across the world where stuff that is completely fake is shared as news on sometimes major channels, but definitely news are sometimes reported in a very distorted way. And this technology will make this significantly worse. Now, if we look at the bigger problem that this represents is that we are enabling a technology that will spread a huge amount of junk, false content into the internet, meaning every content that we will consume, whether it's on social media or on Google search, whatever becomes of it, may not be based on actual facts. It could be completely fabricated.

18:01And more and more of such data is distributed to the world every single day with the advancement of AI, which means sometimes in the very near future, and I don't know the timeline, I'm not a data scientist, but sometime in the very near future, we will need to question every single piece of information we find, even if it's from a quote unquote reliable source that appears on a Google search. And the reason for that, that reliable source may be getting that information that looks highly realistic from a source that is not human and that may or may not have been questioned by editorial oversights that exist today, at least on the bigger channels.

18:41I hate to end on this negative note in the same note that I started, but a lot of things are happening in the AI world that is moving way faster than the regulation is moving right now. And I think the more of us are aware of this, we have higher chances of actually doing positive things with AI capabilities. That's it for this week. If you haven't done it so far, please subscribe to this podcast. Please share it with other people who can benefit from it. We will be back with another interview deep diving into a specific practical topic on Tuesday, as always. And for now, have an amazing day and keep exploring AI and testing new things and sharing what you find with me on LinkedIn and with the rest of the world.