Bye Bye Pi? Inflection CEO and Team Head to Microsoft

0:00Today on the AI Breakdown, the Inflection team, including their CEO and two of their three co-founders is headed to Microsoft. Before that on the brief, rumors about GPT-5 and Saudi Arabia starts a$40 billion AI investment fund. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.network for more information about our Discord, our YouTube channel, and our newsletter. Welcome back to the AI Breakdown Brief, all the AI headline news you need to know. Yesterday, as all this crazy inflection stuff was happening, which of course will be the subject of our main episode today, some other folks were noticing an article from Business Insider that seemed to have a little bit more information about the potential launch of GPT-5.

0:43The piece was called OpenAI is expected to release a materially better GPT-5 for its chatbot mid-year. Business Insider's source was two people familiar with the matter, and it seems like potentially it was some enterprise customers who have recently received demos of this latest model. They quoted one CEO who recently saw a version of GPT-5, quote, it's really good, like materially better. That CEO said that OpenAI had demonstrated the new model with use cases and data unique to his company. He said the company also alluded to other as-yet-unreleased capabilities of the model, including the ability to call AI agents being developed by OpenAI to perform tasks autonomously.

1:17Another one of the folks who talked to Business Insider said that OpenAI is still actually training GPT-5, even though they're showing it off a little bit as well. Given that, there will be a process of safety testing internally as well as red teaming, and so because of that, any timeline could be subject to change. Ethan Mollick points out the stakes of this, not just for OpenAI, but for the field itself. He writes, we will learn a lot about the future of AI from how much GPT-5 beats GPT-4. Agents are an obvious use if it is much better. Now, some are skeptical with the report. The AI Safety Memes account tweets, am I missing something or does GPT-5 this summer not make sense?

1:50One, why would they skip releasing a model called GPT-4.5? Two, if they haven't finished training GPT-5 yet, that would give them very little time to red team. They red teamed GPT-4 for six months. In my opinion, this is 4.5. We don't have any confirmed information from OpenAI, but the rumors from the rumor mill have been that they might be skipping 4.5 and going straight to 5, although ultimately we'll just have to wait and see. Next up today, we could be on the verge of seeing a new biggest investor in the world when it comes to artificial intelligence. According to the New York Times, Saudi Arabia is planning a$40 billion fund to invest in AI.

2:23The New York Times has three sources who have been briefed on the plans, and apparently the idea is a partnership between Saudi Arabia's public investment fund and Silicon Valley's Andreessen Horowitz. The Times notes that Saudi Arabia would be pursuing these goals through its sovereign wealth fund, which has more than $900 billion in assets. They also note that, quote, the$40 billion target would dwarf the typical amounts raised by U.S. venture capital firms and would be eclipsed only by SoftBank, the Japanese conglomerate that has long been the world's largest investor in startups. This is actually a very relevant point.

2:52One of the factors that has shaped the way that AI has played out is that traditional venture funds in Silicon Valley just haven't had enough capital to compete at least in the foundation model space. There aren't funds, for example, that could put in$13 billion to open AI like Microsoft has done. This has created an opening for those big tech companies to have even more control over this new generation of technology, whereas previously that sort of investment would have come from a third party in venture capitalists. This is also not the first time we've seen AI companies turning to the Middle East for access to potentially huge amounts of capital.

3:23You'll remember those reports that Sam Altman was out seeking$7 trillion for a global network of chip manufacturers, and one of his stops in those conversations was the United Arab Emirates. The Times points out that as much money as Saudi Arabia has spent on tech, it hasn't necessarily been all that successful. They write, Saudi Arabia, which poured$3.5 billion into Uber in 2016, has largely struggled with technology investing. It handed$45 billion to SoftBank for the Japanese firm's$100 billion vision fund, which was channeled into dozens of enterprises including the now-bankrupt real estate firm WeWork and other failed startups, such as the robotic pizza-making company Zoom.

3:56Moving back to American markets for a moment, one of the big stories in tech generally, not just AI, is that Reddit is going public. According to the information, a big part of their push to investors is their ability to sell data to data-hungry AI companies. As evidence, of course, they point to their deal with Google, where they will give Google access to Reddit data for helping train Google models in exchange for around$66 million next year. According to the IPO filing, that will likely amount to between 6 % and 7 % of Reddit's expected revenue. On the flip side, some in AI have been skeptical of the value of any one platform's data.

4:30When Sam Altman was asked about OpenAI's reliance on data from companies like the New York Times, which was suing them, he said, Any one particular training source doesn't move the needle for us that much. Ali Golshan, the CEO of synthetic data provider Gretel, said that things are also changing because developers aren't really training LLMs from scratch anymore, instead building off of existing open-source models like those available from Meta and Mistral. Still, Wall Street is very AI-hungry right now, so we'll have to see if the pitch actually works in practice. Finally today, there has been a lot of discussion recently on Twitter slash X around AI self-awareness and inner monologues, and according to a report in Live Science, a new study showed that giving AI systems an inner monologue made them better at reasoning.

5:10A group of researchers applied an approach that they called QuietStar to the open-source Mistral 7b, and really what it was was a slightly different way to respond to prompts. The article reads, The method trains AI systems to think before they respond to prompts, just as many people consider what we should say before we speak. This is different from the way scientists have trained mainstay AI chatbots like ChatGPT, which don't think quote-unquote about what they write or anticipate different possibilities for the next steps in a conversation. Dubbed QuietStar, the new method instructs an AI system to generate many interrationales in parallel before responding to a conversational prompt.

5:41When the AI answers prompts, it generates a mixture of these predictions with and without a rationale, printing the best answer, which can be verified by a human participant depending on the nature of the question. Finally, it learns by discarding rationales that proved incorrect. When it came to a reasoning test, the regular version of Mistral 7B scored 36.3%, while the QuietStar trained version scored 47.2%. Before training, the Mistral 7B scored 5.9 % on a school math test, whereas after the QuietStar training, it scored a much better if still failing 10.9%. Seems to me quite likely to provoke much more studies like this, but for now, that is going to do it for today's AI Breakdown Brief.

6:18Next up, the main AI breakdown. Welcome back to the AI breakdown. The AI community was quite shocked yesterday when news broke that Mustafa Suleiman, the CEO and co-founder of Inflection AI, was headed to Microsoft to run their new consumer AI division. Inflection is of course the maker of Pi, the personal AI, that has been trying to do something a little bit different than ChatGPT or Claude or Gemini or anything like that, and be a sort of AI companion to people. Still, it was only a couple weeks ago on March 7th when Inflection announced its 2.5 version. They claimed at the time that they had 6 million monthly active users, that the average conversation with Pi was lasting about 33 minutes, that they had 60 % week-over-week retention, and that Inflection 2.5 was in fact approaching GPT-4 levels.

7:04And so again, this announcement came as a bit of a shock. The way that the news first broke, it seemed like it was just Mustafa Suleiman, the CEO of Inflection, who was leaving. Suleiman was, of course, previously the co-founder of DeepMind, which sold to Google, and his new position at Microsoft was going to be something with the big title CEO of Microsoft AI. He tweeted yesterday, I'm excited to announce that today I'm joining Microsoft as CEO of Microsoft AI. I'll be leading all consumer AI products and research, including co-pilot Bing and Edge. My friend and longtime collaborator Karen Simonian, who, by the way, was one of the other two Inflection co-founders, will be chief scientist, and several of our amazing teammates have chosen to join us.

7:39Inflection AI will continue its mission under a new CEO and look to reach even more people than ever by making its API widely available to developers and businesses the world over. It's been an amazing journey with so much more to come. Thank you to everyone for your support. Things really are just getting started. The first reaction from the AI sphere was incredulity. Binyu Reddy tweets, Why would you ever give up after raising such a huge investment? Entrepreneurship is resilience above all else. She shared a screenshot of the TechCrunch story from June of last year, announcing that inflection had raised$1.3 billion at a$4 billion valuation.

8:09Neil Koslow writes, This inflection thing is honestly wild. They raised$1.3 billion nine months ago from Microsoft and were founded less than two years ago. And the CEO just bailed to go work at Microsoft? Is there any historical precedent for this? Story has to be more complex than meets the eye. Journalist Eric Newcomer writes, Yeah, I feel Microsoft's terms with inflection and inflection shareholders are a big black hole. People are writing the narrative here without knowing some key facts. There was also some skepticism of the decision from a personnel standpoint. Israel Gonzalez Brooks writes, Mustafa Suleiman, the non-technical of DeepMind's founders, getting continually funded, promoted, and appointed to positions where he has influence over AI is demoralizing.

8:46He's a corporatist, statist, anti-open-source Doomer. People in the industry are shocked. Technium went even farther, writing, Brain-dead move by Microsoft. Doomer grifter master book salesman hired by Microsoft to lead their new AI initiative. The guy started Inflection just a few months ago, raised$2 billion to fund his book tour, then dips. LMAO. Oh well, I guess this takes Microsoft out of the competition for good models. The book thing that he was referring to is the fact that earlier this year, Mustafa Suleiman released a book called The Coming Wave, Technology, Power, and the 21st Century's Greatest Dilemma.

9:15Holding aside any of the details of the book, it did seem weird to folks that the CEO of a company that was less than a couple years old was spending time writing and then promoting a book. So what were the possible reasons that this could have come about? Some speculated that it had to do with the terms of their deal with OpenAI. D. Krantz, for example, tweets, So Microsoft signed a deal with OpenAI until AGI. Now why is Microsoft suddenly investing in other AI companies? Are they preparing for something? The logic here being that if OpenAI is getting closer to AGI, or if Microsoft is afraid that the board is going to say that they've achieved AGI in such a way that it nullifies Microsoft's access to OpenAI models, perhaps they're diversifying ahead of that happening.

9:54The much more common analysis, though, however, was that Inflection just hadn't been able to find a sustainable business model and didn't like their prospects going forward. Adapai summed it up like this, Too much capital and talent needed for next generation. No real way to exit. Reid Hoffman looking to engineer an acquihire, Microsoft unwilling to bite at$4 billion. So they engineered an earn-out deal, allowing founders and research team to leave, investors get back some capital over time. Surprising, Reid managed to persuade everyone to get this done. Reid Hoffman, who is of course the founder of LinkedIn, a co-founder of Inflection, and a board member at Microsoft himself tweeted,

10:53So how was this deal actually structured? The information called it an unusual payout. They write, Microsoft's decision to hire most of the staff of Inflection AI, including the co-founders, leaves the would-be OpenAI rival a shell of its former self. But Inflection has arranged an unusual deal for investors that should take the sting out of that blow. At the same time it revealed the staff departures, Inflection disclosed a licensing deal with Microsoft to make its models available for sale on Microsoft's Azure and said that other cloud platforms would follow. It has also told investors that they will fully recoup their investment and more as a result of the licensing agreement, according to a person briefed on the arrangement.

11:27It's unclear what the size of the payout will be and whether it will come in a lump sum or over time. In addition to the payout, investors will keep their equity, ensuring they can profit if inflection is bought or goes public. An agreement to pay private startup shareholders before the startup sale or public offering is rare and adds to the unusual nature of the arrangement between inflection and one of its main backers. The information also speculates that part of the deal structure might have been a goal to avoid regulatory scrutiny of an outright acquisition. As the information writes, since Microsoft is not buying any assets, only hiring workers, it does not automatically trigger a review from antitrust regulators.

11:59Now, the information also had a little bit more details about how the deal started to come together. Apparently, a few months ago, the CEO of Microsoft, Satya Nadella, had approached Mustafa Suleiman and tried to get him to come join Microsoft. That, of course, would have put it right around the time of OpenAI's board troubles. And it seems to me, in many ways, like the deal that they architected with Mustafa and the inflection team is kind of like what they had planned to do with the OpenAI team, should Sam have remained on the outs from that company. Is it possible that Microsoft actually had started to get excited about the new structure of the team that they would have built around Sam Altman and the rest of the OpenAI crew, and just decided to keep pushing for that?

12:34Coincidentally, in Elon Musk's lawsuit against OpenAI, the complaint at one point claims, quote, Microsoft CEO boasted that it would not matter if OpenAI disappeared tomorrow. He explained that, We have all the IP rights and all the capability. We have the people, we have the compute, we have the data, we have everything. We are below them, above them, around them. There's also a broader line of discussion that this has kicked off about whether it's even possible for startups to compete in this incredibly resource-intensive environment. Ethan Mollick, for example, shared a chart that's been doing the rounds on the companies that had received the most H100 and H800 shipments from NVIDIA this year, with Microsoft and Meta way out ahead of everyone else.

13:10Those two companies had each received around 150 ,000 of those H100 and H800s, as compared to$50 ,000 for Tencent, Oracle, Amazon, and Google,$40 ,000 for CoreWeave,$30 ,000 for Baidu, etc., etc. Ethan writes, It is interesting to see a key truth about the current state of AI play out over and over again. Only a few organizations have the compute needed to train GPT-4 class and beyond models. Key small AI labs may have a lot of talent and ideas, but only four to five tech giants have the servers. Gavin Baker writes, Inflection will likely be the first of many VC-backed foundation model companies to fail.

13:40Foundation models without proprietary real-time data and massive distribution for reinforcement learning from human feedback are the fastest depreciating assets in history. Irony is that while models are commodities today, the ultimate future is likely one where there are only a few truly massive models with proprietary real-time data and vast distribution. Only a few will make it, and they will be super valuable. Smaller open-source models will be used for most vertical tasks to save on inference costs. As ever, there are no barriers to entry on the internet, only barriers to scale. And once it's scaled, the returns are very high.

14:09Foundation models are becoming a game of emperors, and the empires on the other side of this winnowing are really large. Respect to the inflation team for trying. They built a great model, just wasn't valuable enough without proprietary real-time data and massive internet distribution. And yes, technically it hasn't entirely failed yet, but I think fair to say has effectively failed. Signal president Meredith Whitaker put it even simpler, writing, There's no such thing as an AI startup, just guys with models competing to be absorbed by one of the three actual AI companies. So now speculation moves on to what this new Microsoft team is going to look like.

14:41The information points out that it, quote, has to mean something that a longtime Googler will run Google's only formidable competitor in Bing. Continuing, they write, Suleiman will now need to keep Bing competitive not just with Google, but with OpenAI as well. OpenAI CEO Sam Altman said on a podcast earlier this week that he's interested in integrating ChatGPT with a search engine. They also point out that one of the most challenging jobs will now be of uniting disparate AI teams, which, as they put it, have been grappling with overlapping mandates. However it shakes out, this is a big deal that has a lot to say about the state of AI, and so we will continue to watch it closely.

15:12However, that's going to do it for today's AI Breakdown. Until next time, peace.