113 | OpenAI leadership Exodus continues, AI cost competition intensifies, Google search dominance at risk? And other important AI news for the week ending on August 9 2024

0:00Hello, and welcome to a weekend news edition of the Leveraging AI podcast, the podcast that shares practical ethical ways to improve efficiency, grow your business, grow your business, and advance your career. This is Isar Maitis, your host, and we have a jam-packed week of news today, including a significant decline in the cost of using advanced AI models, release of a few new open source models from some interesting sources, and the departure of three key figures from OpenAI. Two of them are founders, so lots to talk about. So let's get started.

0:37it. And we'll start with the big news from OpenAI. As I mentioned, three key figures left the company. So it seems that every time we thought we've seen everything that we can, as far as the term oil from a personnel perspective in OpenAI, every time we think it's over, something new happens. And it's another big one. So two of the co-founders, one of them is John Shulman, which has been with OpenAI from the beginning, one of the researchers behind a lot of the stuff and the systems and the capabilities that we know, is leaving to Anthropic. So OpenAI's largest competitor from the close-source world other than Google.

1:18And he's leaving to Anthropic saying that he wants to focus more on super alignment, which means on AI safety, which is the area that got maybe the biggest hit in OpenAI with Ilya Saskover and Yaliki, which has left after running that division and that division more or less dissolved. So another hint on maybe the lack of focus on safety within OpenAI. So John Shulman leaving to do exactly that, the same thing that he claims that he's passionate about working for Anthropic. The other big departure is Greg Brockman. So Greg Brockman is one of the co-founders and the president of OpenAI. And he's not leaving.

1:58He's taking an extended leave of absence, a sabbatical through the end of 2024. But I think that's a very big hint where the wind is blowing. I want to remind you, roll back to the past year in senior departures and changes in OpenAI. So at the end of 2023, the board decided to axe Sam Altman and get him out of the company. He was let go of the company. And then in a big turmoil that involved, obviously, Microsoft as a big investor that pulled a lot of strings, as well as employees within OpenAI, he was brought back as the CEO again. One of the people that supported him very strongly was Greg Brockman, who was the president before that and resigned from his position following the oust of Sam Altman.

2:49And he was brought back and reinstated as president when Sam got back as well. So his departure is very interesting from a politics perspective, because he was a very strong ally for Sam Altman. No information was released from OpenAI itself. The third big figure that left is Peter Dank. Now that's a less known name, but he's a very interesting position. So Peter was brought in in November of 2023 to lead product development in OpenAI. He has deep experience in running product leadership at Meta, Uber, and Airtable. So he was brought in basically to help OpenAI commercialize and build better products.

3:33And he's leaving less than a year after coming over. He's also not a good sign for top leadership and the direction that OpenAI is taking. So out of the original 11 founders of OpenAI, only three are left in the company. One of them, and the most important one is obviously Sam Altman. Everybody else left many of them in this past year. Another big change that happened from a safety perspective in OpenAI recently that we talked about was Alexander Madri that was leading the safety team was reassigned to another role. So overall, lots of big senior changes in OpenAI. This is never good news when the core leadership of a team is either leaving or changing places and especially moving to the competition.

4:21And a few of the people moved to Anthropic. It's never a good sign. Now, I don't know what that means to OpenAI. We talked last week about their very, very significant and deep losses that they're expected to experience this year. The assumption is about$5 billion in loss. It will be very interesting to see where that leads the company and how much that term oil is going to impact its performance. But let's dive into some additional details about OpenAI, and that's going to show you where they are, maybe more both from a technical perspective, as well as from a business perspective. If you remember a couple of weeks ago, there was a rain of new models and new faster models and cheaper models.

5:02And I told you that one of the things that I'm very curious about is where does that lead from a business model perspective? When models are getting better and better, and at the same time, cheaper and cheaper to operate, A, because smaller models can outperform very well, and B, because there's fierce competition, and a lot of open source models are coming into the market that are highly capable, where does that leave companies like OpenAI and Anthropic that are trying to run a closed-sourced enterprise-level solution for this universe of AI solutions? And I don't have a good answer, and I don't know if there is a good answer, but as of right now, there are serious signs that the competition is only getting stronger.

5:43So OpenAI, without announcing it, just cut the costs of using GPT-4.0 by 50%. So there's a new model that was just released. It's called GPT-4.0 2024-8-06. That basically states the dates when it was released. It is supposedly slightly better than the previous GPT-4.0, and it's 50 % cheaper to use through the API. This comes obviously from significant pressure by the new models by Metalama 3.1, which does very well, and it was cheaper than that, and Gemini 1.5 Pro, that is cheaper than that, and from Anthropic 3.5 Sonnet, which is a very capable model that is roughly priced the same way. So this new model comes to make OpenAI GPT-4.0 more competitive to developers to keep on using.

6:36They also cut the cost for fine-tuning GPT-4.0 Mini. So I mentioned last week that they've now enabling companies to fine-tune GPT-4-0-mini, again, to compete with open-source model that are becoming better and closer to its capabilities. They've cut the cost for that dramatically as well and increased the capabilities of the model. So now the data for training can be up to 65 ,000 tokens, which is significantly more than it was before. That said, the fine-tuning functionality currently only exists for the text functionality and not for images. So they're dropping their prices dramatically. Now, in addition to that, there's some other interesting news that are coming from the OpenAI direction.

7:21As we all know, they've been training and working on GPT-5 for a very long time. Nobody's really telling us what's in GPT-5 or when exactly it's going to be released, other than a lot of rumors and hints from Sam Altman saying that the current models that they have is, quote unquote, is, quote, embarrassing at best. And so they're working on something much bigger, but they have been releasing more and more capabilities on the side. So the voice capability that has been rolling out to a few test users, and we're going to talk more about that in a minute, as well as demoing of video capabilities and so on.

7:57This could be our keyhole to look into the room of GPT-5 and starting to understand some of the capabilities that it's going to bring to us. So maybe that's part of the strategy, and that's what OpenAI has always done, is just releasing bits and pieces of things and iterating through its releases versus one big release of everything. I still think there's going to be a release of GPT-5, but we might be looking at some of its capabilities already through these iterated things. But there's a rumored Project Strawberry that is brewing within OpenAI. Nobody knows exactly what it is. If you remember, when we go back, there were a lot of rumors about Project Q-Star that may have led to the axing of Sam Altman that I mentioned earlier back in 2023.

8:40So now there's this new secretive project called Project Strawberry. And Sam Altman tweeted this week an image of strawberries. And all he wrote was, I love summer in the garden, but with strawberries, everybody obviously jumped into conclusions that I has to do something with Project Strawberry that might be really soon. So we'll keep you posted once we learn what that is. In parallel, a new model appeared on LMC's chatbot arena, which is a platform we talked about many times before that allows users to rank models based on their performance at a blank test. So you basically give it a score running two side by side, picking the one you think is working better.

9:20And this is how they rank AI models based on actual people's usage without people knowing what they're using. In my perspective, this is the most relevant leaderboard right now. So a new model called anonymous chatbot showed up there and it's actually doing pretty well. And the model itself, when asked what model it is, it's claiming it's based on GPT-4 architecture. OpenAI has done stuff like that before, basically released models, not taking responsibility for them or claiming them as theirs, but testing them on the chatbot arena to see how they perform. So multiple things are brewing in the backend, despite everything that's going on the personnel level and the price wars with the competition, which we're going to talk more about in a minute.

9:59Now, two interesting things that have been in the backend that OpenAI has been talking about. One is they've been working on a invisible text watermarking for everything that they're generating. That's basically a way for them to allow other organizations to detect any text that was generated by any chat GPT models. All he does is it makes slightly small changes to the pick of words from the algorithm that they can be redetected by their own detection tool. So from one perspective, that's interesting because it will allow us to detect what ChatGPT creates. But there are apparently several different problems with that.

10:36One is it's not the only model. It can only detect what ChatGPT has created. Two is that they're afraid that bad actors can circumvent that change in the algorithm and hence still allowing the wrong people to create text that nobody knows was created with ChatGPT. And they have an issue with the fact that it's disproportionately impacting non-English speakers. So people who speak English can write English very well. So their need to use a tool like ChatGPT to help them write in English is significantly smaller than people who are non-native English speakers. And so they're afraid from that impact as well.

11:15So that tool, while it exists right now, is not going to be released, at least not for now. Now, we talked about last week that OpenAI started rolling out their voice functionality of GPT-4.0 to remind everybody. They demoed that capability earlier this year, a day before Google's big announcement, and they showed a model that can have a human-like conversation across multiple use cases, including having multiple kinds of conversations. And the biggest difference between that and the current voice capability is that it's actually multimodal, meaning the previous model would actually do three steps to respond to you.

11:51Step one, it would transcribe what you say. Step two, it will analyze and provide an answer, just like a regular chat. And step three, it will translate the words in the regular chat into voice. So that took about five seconds on average to get a response. And this new model just talks like a regular human. So we're talking about milliseconds in the delay of its response. It literally, just like you and I, you don't take steps when you're having a conversation with somebody, you just know what to say and you say it. And this works the same way, including the capability to interrupt it in the middle of a sentence and have it continue with its logic.

12:23So very powerful capabilities. And they just started rolling it out, but they also just shared what they call a system card for the voice capability of GPT-4-0, stating some of the risks and why they're still working with it and haven't released it yet. I'll share a link to that card. It's pretty long and it details a lot of reasons why they're not releasing it yet and what might be the issues. But the two very interesting things that they mentioned there, one is that they're acknowledging that users may become emotionally attached to a human-like voice interface. The researchers that evaluated multiple people using it has observed language that's indicating emotional connection during the testing.

13:03So that's a very short period amount of time. Don't mention the fact that this might be your day-to-day assistant and you may find yourself talking to it more than you talk to other people, including your spouse and other employees in your companies and so on, because that might become, and most likely will become, our interface to computers. We won't use keyboard. Keyboard doesn't make any sense. It's not an efficient way of communication. And part of the fear is that the persuasive capabilities that chats already have over us will grow significantly with voice capabilities, especially that this voice capability can mimic and manipulate human emotion.

13:39That may lead to a lot of bad things, including obviously hurting real human relationships and social interactions in general, as well as people will start building trust in these models and being dependent on them for multiple things, which obviously has a lot of other negative impacts. Another really weird, scary, and interesting thing that happened is that the AI started mimicking the voices of the people it was speaking with. So basically duplicating the user and using his nuances of the voice, the way he speaks, his diction, as well as the voice itself, without being asked to do that. That's again, very scary on multiple levels and very interesting on why and how the AI is doing it, but it's not something they definitely want to release into the wild.

14:24Another thing that they're afraid of is all these models have various guardrails to keep us safe from other things that they might be able to do, and we don't necessarily want people to do with them. And they fear that the ability to have a conversation will enable new jailbreaking capabilities that will allow people and bad actors to put these models to a bad use. So overall, lots of good reasons why not release the model yet. But as I mentioned, they already started releasing it to a small group of people to test it out. Now, what does that small group of people mean? I don't know if it's 3 ,000 or 30 ,000.

14:57Either way, they have over 100 million users active and probably 200 million people registered. So whatever that number is, it's probably not three people, but it's probably not very large. And once they figure out all these kinks, I assume they're going to roll it out to everybody, including any user of ChatGPT. That is going to change everything. As I mentioned before, once this can be connected to other software as the interface, nobody will want to use a keyboard and mouse anymore because you can just have a conversation, explain what you want. and the AI will perform the tasks for you in the software that you're operating.

15:32I don't see that happening this year, but I definitely see that happening in 2025. And the last small but exciting piece of news from OpenAI is they're now allowing free ChatGPT users to generate two images per day with DAL E3. So DAL E3 is the text-to-image generation model from OpenAI, and it's been available to paid users for a very long time. I'm using it regularly. And now they're going to allow free users to use the model for, as I mentioned, two images per day. Now, why is that important? I use DALI for image generation for multiple purposes. When I need higher quality, I use other models.

16:14When I need text in images, I use other models. So I use mid-journey for the higher resolution images. And we're going to talk about an option for that. That just came out really exciting this week. And I use ideogram every time I want to create images with text in them because it's just amazing at it. So if you've never tried Ideogram, go and test it out. If you need any kind of graphics that involves text in it or just different styles of text that you can then copy and paste on whatever graphics that you have. But I use DALI 3 every time I need context and it's actually the best tool for that reason.

16:48So it's the best tool when that is your need. So let me explain two specific use cases. Let's say you're developing a presentation for whatever business need. So I'm working on my presentations and brainstorming what needs to be on different slides and researching things using ChatGPT. But then at the end of this, because it has a very deep understanding of my target audience and who I am and what I'm trying to achieve and what's the goal of the presentation and what is the flow of the presentation because I've worked on the presentation together with it, I can ask it to recommend what kind of images should go with each slide.

17:21I ask it to explain to me in text what those might be. Then I pick the one I like. I make whatever modifications I want by explaining what modifications I want. And then I ask ChatGPT to create the images for me. I do the same thing when I create some blog posts. And I do the same thing when I create social media posts. And so it's knowledge of everything that I've done for that particular goal makes it a lot easier to create images that are relevant to that. And I just find it a very straightforward flow, different than working image by image and trying to fix it while I'm working with MidJourney or Ideogram.

17:57So as I mentioned, if you have the free version, first of all, I suggest paying for the full version. It's the best 20 months a month you've ever going to invest in any piece of software. But if you don't want to do that, you can get access to two images per day on ChatGPT, free version. And the last piece of news from OpenAI, actually doesn't come from OpenAI, but directly related to them, comes from Elon Musk. So if you remember, Elon Musk sued OpenAI a while back, and he dropped the lawsuit a couple of months ago without explaining exactly why. But the previous lawsuit was about the fact that they're betraying their fiduciary responsibility as a nonprofit organization.

18:37Now, the lawsuit that he claims is significantly stronger is saying that OpenAI has breached its founding mission to develop AI for the benefit of humanity. and that Sam Altman and Greg Brockman that I mentioned just took a sabbatical, manipulated Musk to fund while co-founding OpenAI using false statements. So basically they're claiming they used fraudulent ways in order to make him fund their initial steps. So the actual lawsuit is about violation of federal racketeering laws and conspiracy to defraud Musk himself. it was pretty obvious in the previous lawsuit that he's going to lose, especially after OpenAI shared that he himself suggested to buy and basically roll OpenAI into Tesla.

19:25So his suggestion that moving OpenAI from a nonprofit organization to a for-profit organization is something he personally suggested that he was trying to pursue. So he dropped that lawsuit and now there's a different version of it. There's a lot of bad blood between Elon Musk and Sam Altman, specifically about the role of OpenAI in the universe that, as I mentioned, Elon Musk was a co-founder of and wrote some of the original and probably bigger checks in the beginning of the company. Now, where's that going to go? I don't know. As I mentioned, there's a lot of turmoil going on in OpenAI right now.

20:00That just adds some fuel to the fire. I don't think he's going to win this lawsuit because I think it's going to be very hard for him to prove that they were planning this thing all along versus this is how it evolved. But that being said, as I mentioned, that's going to make the whole thing more interesting. And I will report to you as things progress. Now, I mentioned earlier that OpenAI slashed the cost of ChatGPT to keep them more competitive in this highly competitive market. They don't live in a vacuum and Google immediately the next day slashed the prices of Gemini 1.5 flash models. So Gemini 1.5 is their smaller model and they just cut its cost by 80 % starting on August 12th.

20:43So it's not yet available, but the new pricing is going to be 0.75 cents per million input tokens and 0.3 dollars per every output tokens. That's going to be 50 % cheaper than OpenAI's GPT-40 mini, which is their smaller bottle. That being said, from a capability perspective, GPT-40 Mini is performing significantly better than Gemini 1.5 Flash. And when I say significantly better, on most benchmarks, it's doing much better. And GPT-40 Mini is also performing much better on the chatbot arena we mentioned earlier. It is currently ranked number three just after GPT-40, its bigger brother, basically.

21:30So the GPT-40 and GPT-40 Mini, They're very close on the ranking, which tells you that GPT-40 mini being a smaller, much cheaper model is probably the best way to go right now if you're looking for something fast and small. And number one is still, by the way, Gemini 1.5 Pro, so the latest big model from Google. But Google's small model, Gemini 1.5 Flash, is only ranked number 17 on the list, so very far behind. So from a capability perspective, Gemini 1.5 Flash is still behind GPT-40 Mini, but it's going to be much cheaper than it to use. So I'm sure developers will find the right use cases for each of them.

22:11But as you can see, this competition is intensifying, especially when there are more and more very capable open source models like Llama 3.1 that was released two weeks ago, as well as other open source models, which we're going to talk about later on in this news episode. Now, to put things in perspective for a minute with this whole down spiraling price competition, there was a very interesting article by David Kahn from Sequoia, and he wrote a post called the$600 billion question. This is actually a follow-up from an article he wrote in September of 2023 called the AI's$200 billion question.

22:50And basically what he's showing is in order to get positive returns on the huge investments that these companies are making on chips, he's not even talking about the rest of the infrastructure. These companies jointly need to make, as of right now,$600 billion in revenue. They're not even in the ballpark. So if you remember last week, we talked about the revenue from OpenAI. OpenAI will probably make this year anything between$3.5 and$4 billion. Google has not shared yet what kind of new revenue it's generating from AI. But even if it's in the same ballpark and the same with Microsoft, we're talking about double digits in the teens from all these companies combined, even if you throw in Anthropic and some of the other companies, and they need to be making$600 billion instead of $12 or$16 or even$20 if you go wild with your assumptions.

23:47And so that's not a sustainable business model, but all these companies, the big ones, have so much cash they're sitting on that they can bet that this will actually going to work. And I think their biggest fear is what happens if they do not bet and this actually generate these kind of returns. So it's a very weird game that they're playing with very high stakes where they're throwing insane amounts of money, bigger than any company has ever thrown on a single initiative ever before, assuming that it's going to pay off and fearing the situation where they don't make that bet and somebody else actually makes it and is able to generate that kind of revenue.

24:29Now, still about Google, an interesting research came out this past week from Sonana Insights, sharing that Google search traffic has not declined despite the rise of AI search engines. So Google's search traffic has actually grown by 1.4 % from May 2023 to May 2024. Now that's despite the fact that perplexity that I'm a very big fan of, that is probably the most capable AI search engine right now, has grown 42 % from earlier this year to now, but that's still a very small number of searches compared to Google's complete dominance in search. So people have been searching Google 290 times more than people have been searching perplexity in May of 2024.

25:15So that's a very recent data that kind of shows you the trend. In addition, Google users perform about 200 searches per user per month, while perplexity average user performs about 15 searches per month. I can tell you that Advanced users like me, I use Perplexity now more than I use Google. I think the transition to AI-based search that gives you an answer versus just a list of websites to visit and research yourself is something that is going to happen. It's going to happen across the board. Google themselves are testing it as well. I think what's going to make the biggest difference in this particular field is the recent release of Search GPT, which is the AI-based quote-unquote search engine that is starting to roll out from OpenAI.

26:01The reason I'm saying it's going to make a much bigger impact is not necessarily because I think it's better than perplexity. I haven't had a chance to test it myself. I've seen some very positive initial reviews from people who got access to it. But the reason is perplexity does not release the amount of users they have, but let's be very loose and positive for them, let's say they have 15 million users that are using perplexity, OpenAI has hundreds of millions of users. So if OpenAI rolls this out, that's going to be a very different kind of impact from just a user-based perspective to what perplexity is doing right now.

26:35And this may put a dent into what Google is doing, putting more pressure on them to start changing what they're doing. What does it mean to you as a business? What it means to you as a business that this thing is happening. Whether it's going to happen this quarter or this next year, it doesn't really matter. If your business depends or some of your revenue depends on organic traffic, and especially if most of your business depends on organic traffic, you need to start diversifying your sources, start a podcast, have a YouTube channel, set up local meetings, whatever it is, start a newsletter, anything that will give you traffic and eyeballs on what you're doing and the offering that you have that does not come directly from Google search.