135 | Will the global workforce survive in an era of AI agents with computer access? Advanced new image and video generation capabilities and many more news from the week ending on October 25th

27:05I've been following the information very closely and everything they've released have always been accurate. So time will tell. They're obviously working on this. They haven't been hiding that they've been working on this, but maybe the release date of December is just not accurate. Either way, everybody's expecting the next model release from ChatGPT to be very significant. Now, in a research paper released by two researchers by ChatGPT, they're sharing that they developed a new technology that is accelerating the generation of images with diffusion models 50 times faster than it is today. So they've developed several different sizes of models, but now they're able to create images with diffusion models that are almost at the top level of the other models today.

27:46They're claiming a 10 % gap in quality with a significant reduction in compute and time. So far, they've done this for relatively small images, so 512 by 512 pixels. But once they figure out the capability, they can scale it up and add more functionality to it. And they're claiming that the same concepts will be able to be applied to images, audio, and video, which tells us these things can be in the backend of the next generation of DALI 4 or whatever they're going to call it, and Sora or whatever they end up calling their video model. And so we've talked about this many times in the past on this show that the capabilities are becoming faster and cheaper for two different reasons.

28:33One, compute is becoming cheaper, but also because new algorithms and new algorithms and new processes are being developed and they allow the companies to run on the same compute significantly faster, in this particular case, 50x faster. So that's a huge improvement. And I'm personally excited as somebody who generates lots of images and some videos. I'm very excited about this. Another OpenAI related news comes actually from Morgan Stanley. So Morgan Stanley, one of the largest investment banks in the world, have announced that they are expanding their partnership with OpenAI. They're claiming that right now, nearly half of Morgan Stanley's 80 ,000 employees are using OpenAI Power Tools daily.

29:16They're claiming game-changing efficiency with Salesforce forced responses to clients' inquiries 10 times faster than before, the tool is processing insights from 70 ,000 plus annual research reports, and that these new set of tools from OpenAI are being used three times higher than previous traditional analysis tools that they've been using. So obviously, they're seeing huge benefits from this, and they're, as I mentioned, announced that they're expanding this partnership to do more with it. The biggest use cases that they're using it for is summarizing research on stock commodities and industry trends.

29:55It processes complex queries about specific companies to allow to make better decisions, and it handles industry specific jargon for specific companies and creates data visualization for the employees. Multiple tasks around research and preparation to investment. And as you already know, there's been multiple companies around the world that's been doing algo trading for a very long time. So this is just going to expand on that to the point that the stock market is going to run mostly by algorithms, even from private investors and not just from large companies. There's been more and more rumors, by the way, about OpenAI and their tension with Microsoft, with the relationship with Mustafa Suleiman that took the helm of AI at Microsoft.

30:37And as I shared with you last week, this frenemies relationship looks to be not on stable grounds all the time because Mustafa Suleiman has a history with some of the people in OpenAI and also the fact that he came from a competitor and now is running this. There are a lot of discussions that they're not sharing the information that they're supposed to share and that some information that's not supposed to be shared is transferred over, not necessarily through the right channels. So it doesn't seem that there's a lot of trust between these two organizations right now. And as we know, OpenAI are developing things that are competing with Microsoft.

31:11We talked a lot about this last week. And Microsoft is also working with leadership of Mustafa Suleiman to develop their own models that can potentially compete with OpenAI. So it'll be interesting to continue following this partnership. Right now, it seems that Microsoft has a lot more leverage in this relationship because they have a very solid agreement with OpenAI as far as how much of the revenue is going to go to Microsoft and access to their models. And they're the ones that control the compute. And they're the ones that are investing a lot of the money. But there's this little interesting clause that we talked about when this relationship started in 2023 that allows OpenAI not to deliver the next technology to Microsoft if they define it as AGI.

31:58And the authority to determine what is AGI and what is not lies with OpenAI's board. So basically all that OpenAI needs to do in order to unplug Microsoft from the next variation of models is for OpenAI's board to say, okay, this is now AGI. So GPT-5, Orion, whatever it's going to end up being called is now AGI and hence Microsoft does not get access to it. So this is their leverage in this entire relationship. As I mentioned, it will be interesting to continue following this relationship between these two companies and see where it is evolving. Now, it's been a while, and I want to say a while, maybe a month, since a senior person from OpenAI is leaving, and then at the same time talking about his concern about how OpenAI is approaching safety.

32:44So the current person is Myles Brundage. I hope I don't butcher his last name, but Myles is leaving from his role as senior advisor to AGI readiness. So again, a very senior person in the safety side of OpenAI, and he's making similar claims to previous people who left OpenAI. And Miles is saying that he's planning to pursue policy research in the nonprofit sector that has to do with AI. And he cites his desire for more, and I'm quoting, more ability to publish freely, which basically lets you understand that some of the things he wanted to say in OpenAI, he was not allowed to say. As we know, the AGI readiness team has been wounded down and has been distributed across other segments of the company.

33:31So he urges other employees in OpenAI to speak their minds more freely and discuss the current developments as well as safety or lack of safety. And as I mentioned, he's not the first person not leaving. We had CTO Mira Morati, the chief research officer, Bob McGrew, research VP Barrett Zoff, research scientist Andrzej Karpathy, co-founder Ilya Saskover, co-founder John Schulman, Greg Brockman, who is on quote unquote extended leave, et cetera, et cetera. A lot of senior people have left. Many of them sounded the alarms and the concerns on the push towards competitiveness with not enough safety concerns.

34:11So here's just another person in this very long list of people that is departing OpenAI and sounding the bells on safety. that doesn't sound well, especially when you're going back to the topic that we talked about in the beginning of how much safety we actually do need because of where these systems and how these systems are being deployed. Now, since we were speaking about the relationship of Microsoft and OpenAI, they just launched a$10 million media initiative to fight copyright battles. The idea is to have funds,$5 million. Right now, they're doing this with two different companies, $5 million each with$2.5 million in cash plus$2.5 million in software and enterprise credits to use the models.

34:55And they already have initial recipients that are going to go after these funds. So consider this a fund that will help companies, A, adopt AI, and B, paid for their news outlets. Some of the companies that were mentioned were Newsday and the Chicago Public Media, the Seattle Times, and a few other organizations. The goal is to have a two-year fellowship program that allows access to Microsoft Azure and OpenAI credits and is focusing on implementation in the newsroom. And in return, it will give them access to the news themselves that they can use with a license to do so rather than stealing the information like they did so far.

35:36On the same topic, Meta just announced a partnership with Reuters. So it's the first time Meta is cutting such a deal. It's a multi-year agreement for real news content. The goal is to provide Reuters real-time news through all of Meta's AI chat capabilities, which are already integrated into Facebook and Instagram and WhatsApp and Messenger. And so you'll be able to get coverage and know what's happening in the news just by asking the bot on each and every one of these platforms. This is rolling out right now. So this is an immediate thing that will be available to all of us in the next few days.

36:10Now, before we dive into some news from Google, I want to talk specifically about an interview that was held with Google's DeepMind CEO, Demis Asabis. He was talking a lot about AI agents and AGI and so on. It's a very interesting interview. It's not a very long interview, and I will share the link in the chat so you can find it and listen to the interview. And I want to mention a few things from the interview. A lot of the stuff that he talked about, we kind of known, but I've learned more about Demis, and I've been following him for a while now, but I've learned more about him through this interview.

36:40And a lot of the way he thinks about AGI and so on. So first of all, he thinks we'll get to AGI only in about 10 years. So if you're following his competitors or other people in the industry that are talking to two, three, maybe five years, he looks at it way further out. He's talking about the capabilities that we need to achieve and master to get full AGI, but that's also required for all the agents that we are starting to see right now are planning and thinking ahead, taking action in the real world. So we talked about having access to a keyboard and mouse and so on, reasoning through problems.

37:16So being able to analyze them and understand what are the limitations and work through different steps of the thing, improved memory retention. That's a huge one because if you think about what we talked about before, being able to solve more complex tasks that may take not five seconds or 20 seconds, but maybe a full day or a few days or a couple of weeks will require for the system to know in its memory everything that is happening across the company, connections to other components, and everything that it has executed through that period of time. So that's obviously a very big one. Enhanced personalization, that's something else that he's talking about.

37:53So the ability to understand the needs of a specific individual or a specific organization to really serve them in the most optimal way and the ability to use various tools that are both software and hardware. So these are the things that he's talking about that everybody's working on right now that are going to push us forward in this more agentic world in the road to AGI. But he also talks about the risks and he is in the middle of the spectrum, right? He talked in this interview about the doomers, but also about the tech people that are pushing very fast forward. And he's in the middle. And the reason he's in the middle, he's saying he's optimistic and he thinks we'll get the best out of it.

38:31And he really believes that he will be able to solve diseases and change global warming and find energy solutions that are amazing. So he's talking about this future abundance, but he's saying that there are risks. And now I'm going to quote something that he said that I find profound. So those of you who don't know, Demis is a world-level chess player. And now I'm quoting from the interview. These are new systems with new technologies. They are incredibly powerful. I've seen this in the microchasm of games, something I understand well, like playing chess, where you start with a system alpha zero that's random in the morning.

39:08By morning coffee break, it's better than it was before and can beat me. And then by lunchtime, it's better than the world champion. And then by the afternoon, sort of in eight hours, it's better than the best chess players, you know, hard-coded chess computers. And when he lost to a computer playing AI, but these was hard-coded AI computers. It was really not the AI that we're talking about today without real machine learning. So in eight hours, it goes from not knowing how to play chess, literally making random moves to beating a computer that can beat the best chess players in the world. in eight hours.

39:48So what he's basically saying that the ability of these systems to learn really complex things much faster than we understand once they have access to the right information, the right compute is beyond our control and is beyond our level of understanding. And basically what he's saying that this thing can take a turn in the road and run in that direction before we understand what's happening. And it might be too late. Going back to my whole point about, we have to figure out how, if it's even possible, to put real guardrails on the system and having safe mechanisms like a fail switch that we can actually click to turn this thing off.

40:27I don't know if that's possible. I don't know if that's something that's even realistic, but I think that's what we need. And since I told you we're going to talk a little bit about Google, Gemini just introduced more applications for Workspace. So now there are AI capabilities for Gemini in Google Calendar, Google Keep, and Google Tasks that can be all activated from the admin panel of Google. So you can activate it to specific employees in your organization. And it can create calendar events. It can generate notes. It can manage tasks. It can access and move data in those apps. And it integrates with existing Gemini panel in the workspace apps.

41:02Now, I told you I also have a lot of news about image generation and that world. We shared a lot about Adobe's new releases last week. So Adobe released all these amazing capabilities. And this week, MidJourney launches an AI image editor that basically allows you to take real images, actual photos that you took, and edit them with MidJourney. Look at the demos online. It's absolutely mind-blowing. You can do really cool things. And it allows users to modify an existing images and transform their in their style. You can highlight specific segments in the picture and then change just those segments for different colors, different patterns, different textures, and so on.

41:41Now you can even take doodle art, something you scribbled, and turn it into an actual photorealistic image or any other kind of image you want, like cartoonish, whatever it is that you want. So very powerful capabilities, including retexture feature that allows to convert things from one style to the other, either for the full image or for part of the image, and so on. And it's becoming very, very good at following the prompts and the user guidance on exactly what it needs to do. Now, the problem with it right now is that it's only accessible to users who have generated 10 ,000 images or more, have an annual paid membership, and have been subscribed paying users for at least a year.

42:21So it's not accessible to most, but just like they did with the previous releases, this will change in a few weeks or in the previous time it was a few months, but eventually we're all going to get access to this. The reason I find this interesting is I think what we're going to see, and I'm going to combine the news from Adobe from last week to the news about mid-journey, and we're going to talk about Canva in a minute as well, is that the lines between AI image generation to image editing are going to keep on blurring. I think the tools that were traditionally let's create an AI image with, like mid-journey and Flux1 and an ideogram and so on, are going to go more and more into tools that will allow you to control and change and edit images, whether created by AI or not.

43:06And we're going to see the tools like Adobe and Canva and so on move more and more into the AI image generation capabilities. And the lines are going to be blurring more and more and more. Now, I don't see any large organization that are using Photoshop right now, ditching Photoshop will start using Me Journey, but I definitely see smaller companies, beginners, and people like me who do not really know how to use Photoshop and never had the time to learn starting to use these kind of tools because it's just a lot easier to use and they will give you good enough results across multiple things. And I think two years down the road, it will be very hard to distinguish between the two kinds of tools.

43:47Now, Me Journey is also continuing with their plans that they've announced before, to release 3D capabilities as well as video capabilities sometime in the future. And staying on the same topic of image generation combined with tools, a few weeks ago, I shared with you that Leonardo.ai was acquired by Canva. So this week, Canva has released what they call DreamLab, which is a new entire suite of AI capabilities built into Canva on top of Leonardo's Phoenix model. It allows for really advanced graphic generation with specific styles, allows it to do improved multi-subject image creation and a lot of other capabilities that are built straight into Canva.

44:30You will be able to also create text in a much better way. So AI text, which has been a problem until recently. So I've been using Ideogram every time I needed to create text, but apparently now you can do this built into Canva. So more and more AI capabilities, creation, editing, and so on are coming into Canva itself, which as a Canva user, I'm very excited about, but it goes back to prove the point that I said before. The lines are going to be very blurry between content creation tools, image editing tools, and image generation tools, and video generation tools that are now available in Canva using Kling in the backend.

45:03So all of these tools are built into Canva and you can create a lot more without ever leaving the platform. is Stability AI just launched Stable Diffusion 3.5, which is a major improvement over Stable Diffusion 3. It released three different models, a large one with 8 billion parameters, a mid one, a large turbo with 8 billion parameters that runs faster, and a medium one with 2.6 billion parameters. And this model achieves enhanced realism and superior prompt adherence and improved text generation and a lot of other better capabilities. As you probably know, Stable diffusion is an open source model, which means it can be implemented across everything you want without paying licensing fees and so on, as long as you're a small enough company.

45:47So you can use stable diffusion today for free or non-commercial, or if you're making less than$1 million, you can still use it for free. And I assume we're going to see a lot more companies are using that. Plux model that was released just a few months ago has taken over like a wildfire as far as being integrated into many different tools. And maybe now with the release of Stable Diffusion 3.5, we're going to start seeing it being integrated into more and more places. They're claiming, by the way, that their capabilities are matching or exceeding Flux 1.1 Pro, which is right now the most advanced open source model out there.

46:23And I must claim as somebody who's using it regularly, that is as good as me journey on most things. And again, it's free to use. Staying in the open source world, a company called Genmo has released Mochi One, which is they're claiming to be the best open source AI video generation tool. They're claiming that they can compete on some capabilities with Runway Gen 3 and with Luma Dream Machine, which even if it's not completely true, and even if they're still behind, just like I mentioned a few seconds ago, these tools will grow with their capabilities, then these open source tools will be able to do good enough work.

47:03So even if Machu1 is not there yet, and again, they're claiming that it will be, right now it's lower resolution and it's shorter videos, but the goal for them is to be as good or even better than Runway and Luma and Kling and all of those competitors. And they're focusing right now on just photorealistic content, so not cartoonish stuff like a lot of people are doing. But as I mentioned, over time, I'm sure they will develop more and more of that, and it will be open source and will be free and accessible to integrate with everything else. So that's the direction this thing is going, which connects back to the point that I mentioned several times before.

47:35We're going to talk a little more later on when I talk about Anderson Hurwitz on where is the financial model for the paid models right now, which I'm not 100 % sure about. But staying on the topic of video generation, Runway just launched what they're calling Act One, which is an AI facial motion capture tool. What it allows you to do is it allows you to record yourself on the phone doing whatever facial expression you want without any special equipment and then bringing that into Runway and using it to change the expressions of the videos that are generated within Gen 3 Alpha. So all you need is to create an image of something with a third-party tool.

48:19This could be a cartoon. This could be a photorealistic person. You upload their images to Runway. You upload a video of the expressions that are used as a reference to create a new video. And now that image that you created that could be a cartoon will have the same exact facial expressions that you had when you were using the phone. That saves a huge amount of work that studios had to do somewhat manually so far. And again, it's going to be available on a platform that costs a few dozens of dollars a month, depending on how many credits you want to actually use. This is absolutely magical, and it's going to dramatically simplify and democratize the generation of high-quality video, including people and their face expressions.

49:03And it doesn't even have to be people because you can use those face expressions and then apply them to a cat or a legendary magical creature that you will create the image of using a tool like Me Journey. And from image generation, let's talk a little bit about perplexity. So perplexity, we talked about them many times before. They have been the best AI search engine that is, quote unquote, competing with Google. They're obviously significantly smaller than Google, but they've been growing very fast. So Arvind Srinivas, their CEO, has announced this week that they're now performing 100 million queries each week.

49:34That's about 400 million queries a month. That's up from 250 million a month in July. So this is one and a half times bigger in just a few months, which is telling you how fast they're growing. But they're also adding more and more capabilities that are turning them to a lot more than just an old school search engine. So right now, they've added real-time stock quotes, including the access to historical earnings, data, and visualizations, and industry peer comparisons that has to do with stocks, and company financial analysis, all powered by financial modeling prep, FMP data. They're bringing more and more specific data, not just general internet data, in order to allow people to do a lot more.

50:19They've also announced new partnerships with Crunchbase that will provide private company data access, FactSet, which is a structured, unstructured financial data. And they're making this available for now for enterprise users, but I assume this will come to everybody in the near future. Now, it's still not competitive and solid and as robust as existing financial research platforms, but the direction is very clear. And this is another place where the lines are going to get blurry between traditional tools like Bloomberg to this new age of AI-driven tools that will have access to similar data.

50:57and will be able to do stuff that previously only companies like Bloomberg and Morningstar were able to do, and now will be accessible to hopefully all of us through tools like Perplexity. Now, speaking about Perplexity, Dow Jones and New York Post is suing Perplexity for massive illegal copying of publishers' content. And that's a quote from the lawsuit that they filed just on October 21st. This is not the first time an AI company is being sued by different people who publish content. We've seen similar lawsuits from the New York Times suing OpenAI Microsoft and a lot of other news outlets suing other platforms.

51:34So this is not new. And as we've seen before, it leads to more and more licensing deals like the ones we talked about earlier. And I want to jump to a slightly different topic. I said several times before on this podcast, actually many times, that I don't see the business model future of these models. And the reason I said that is because models are becoming cheaper and cheaper with every cycle. And you have better and better open source models that may not be the best in the world, but might be good enough for most of the tasks that we need. And they will definitely be good enough sometime in the next two to three years.

52:09And then where does that put the money that needs to come in and pay for the development of these models? Well, this week, Mark Anderson from Anderson Horowitz, one of the biggest and most respectable VCs in the world, said the same things. He's basically saying that AI models are becoming commoditized where he compared it to selling rice. He's claiming that this is a race to the bottom, which is exactly what I've been saying for a very long time. And that, as I mentioned, it's becoming a commodity and that there's very limited product differentiation between those providers. and if you'll be able to take even an open source model that is good enough, wrap it with the right capabilities to make the applications as relevant as possible to whoever you want, you may provide more value than just competing with other leading frontier models.

52:58Now, this is very interesting that he's saying that because Anderson Horowitz was one of the original investors in OpenAI. They have participated in a$300 million round in OpenAI in 2023. So when he's coming and saying that, when he has a significant investment in one of those companies, it's telling you that this is a very serious risk to the business model of these companies. That's it for this week. We talked about many, many different things. We're going to be back on Tuesday with another fascinating deep dive into a how-to with AI, something that you can start implementing in your business immediately in order to get higher efficiencies and higher ROI in your business.

53:35If you are enjoying this podcast, please share it with other people who can benefit from it. Please also open your phone right now and rate us on either Spotify or Apple Podcasts, whatever tool that you're using. It really helps us get the message out there. And we're going to have more people understanding AI, which will hopefully help all of us end with a better outcome of this amazing and yet scary revolution that we're living through right now. and until then, and until next time, have an awesome rest of the day.