OpenAI's $7 Billion Buyback: Key Developments

0:00OpenAI has just bought back 7 billion dollars worth of employee shares at about an 852 billion valuation. Anthropic's unreleased model has advanced the Riemann hypothesis. They spent 31 million tokens to do this. OpenAI's Astra model has solved a 10 long open math problem for$2 ,000 worth of tokens. So it wasn't cheap, but it was able to get it done. Alibaba released Quen 3.8 Max, which is a 3.4 trillion parameter model. It's basically competitive with Claude. Researchers right now are extracting hidden reasoning from Claude. So Chai Chibuti and Gemini, they're using a trick with the API and they're comparing the reasoning of different models with the open source or open weight models that are Chinese versions to see if they ripped off those models.

0:47Almost everyone I know is paying$20 a month for some AI tool, whether that's Chai Chibuti, Claude, Gemini, Canva, there's so many out there. And the one thing that I will say to all of them is I would love for them to check out AIbox.ai, which is my own startup that lets you access 80 different AI models all on one platform. Our cheapest tier is$8.99 a month. It is incredibly cheap. It's the price of a coffee and you get access to over 80 different AI models. There's image models. There's audio models. There's video generation models like Google VO3. And of course, you have all of the text and reasoning models like Claude, ChatGPT and Gemini.

1:22If you want to get access to all of that in one place and also have an MCP, which basically is a link that lets Claude, if that's your main model that you use, generate images or let's cloud generate audio or let's cloud generate video. So you can pull any model into any model. If you want to check that out, it is AI box.ai. I'll leave a link in the description. OpenAI has just completed a$7 billion buyback of employee shares. They did this, which I think a lot of people are kind of shocked by at$852 billion valuation. Now, if you remember, this is the exact same valuation that they had back in March, we are now many months later.

1:57And if you look at a company like Anthropic, which month over month is adding like$100 billion to their valuation every single month, basically, this is not I mean, you could look at this a couple different ways. But like, technically, that's not great. If their valuation hasn't grown since March, I think this is basically they're trying to keep the staff from leaving and going to Anthropic or going to Gemini. But also, I think this is basically showing they're not in a big rush to go public anytime time soon. This$7 billion tender offer is one of the largest employee buybacks a private company has ever done.

2:31I think that's no shocker, but this is very consistent with what OpenAI has done. They have a pattern of doing really regular secondary sales. They've done this for the last three years where they'll let the employees sell at a regular basis. Sam Altman told the staff last month that OpenAI missed a bunch of internal financial targets over the last year, but he expects that the next year is going to be stronger. I mean, there's the elephant in the room and the no surprise is that Anthropic just absolutely mopped the floor with their growth. And OpenAI, I don't think has actually shrunk. They probably just went stagnant or they kind of plateaued for a minute.

3:05And the reason why is just because Anthropic grew so fast, they just took a lot of the oxygen out of the room. Is it going to stay that way forever? I personally don't think so. I'm heavily testing both Claude and ChatGPT. And right now, my main tool that I'm using is ChatGPT work. I switched from Cloud Cowork about two weeks ago, and I've been really impressed. Cloud Cowork or ChatGBT work can do a lot of things that Cloud Cowork didn't. It feels like OpenAI felt pretty threatened and they put a lot of time and energy into getting this right. So I love a good healthy rivalry. I think it was going to go back and forth.

3:38I'm excited if we have other competitors, if it feels like Google gets a little bit more in this space, you have perplexity in there. But overall, I think OpenAI has a superior product right now. And so I think they might start getting back a little bit of that enterprise that they've lost. Anthropic was reportedly profitable earlier this year. I think that's definitely going to threaten OpenAI as far as if Anthropic can reach the stock market first. If they can do their IPO first, they're going to get a lot of buyers that would possibly have invested in OpenAI if not. So I think there's definitely a timing issue or there's a motivation to get to an IPO first for OpenAI and Anthropic.

4:18but right now if they're holding their valuation flat and they're doing their their tender offer for all their employees now rather than waiting for the ipo open ai is basically buying time to prove that their business model is going to work at scale especially for enterprise and their api products that have real recurring revenue also i mean if you want to be like sort of pessimistic maybe you could say look perhaps sam altman believes that the company is is worth more and it's growing but he just wants to buy back all the shares at a really good deal so he can go buy you know,$7 billion worth of shares from the employees.

4:50And when they IPO, they can bump up the IPO price 50%. And now those$7 billion of shares will be worth$14 billion. The company that's worth over$850 billion. I don't know if that's really the main incentive here. But there's definitely a case to be made that perhaps those shares are at a discount. I think more likely, he's just trying to keep his employees happy and keep them around. But you know, it's got to make you think that an IPO is not just weeks away if they're able to or if they're going to be buying back these shares from their employees. You know, if an IPO is around the corner, they probably just let them all sell on the open market.

5:24But if it's going to be months away, they probably will just buy them and wrap that up. Now, there's an unreleased anthropic model that is advancing the Riemann hypothesis. So basically, this is a 150 year old math problem. There is a$1 million bounty on this math problem, by the way, which is kind of cool that this exists. But by extending the range of numbers for which it's been verified, this new model from Anthropic ran itself for a day and a half. It organized 60 subagents to test 650 different ideas and formalize the results in a way that mathematicians could check because not only do you have to solve these problems, but you have to explain how you got it.

6:03So this is a big part of it. The model orchestrated its own workflow without a human mathematician directing it. It was assigning tasks like idea generation, validation, paper drafting to different subagents. It used about 31 million tokens to produce this. Two anthropic mathematicians confirmed the results, which was then formalized in Lean, which is an open source proof assistant that lets other researchers verify it step by step. OpenAI's internal astro model proved 10 major results this year and a separate anthropic effort disproved the Jacobian conjecture, which is a 1939 problem. So I think we're seeing a big shift in how AI is tackling big problems in math right now.

6:47I think AI is producing verifiable mathematical progress on a bunch of unsolved problems. But there's a lot of people in the math field that are debating what it means for credit and accountability when a machine is coordinating all this work instead of a person. But also you can imagine, well, what if a person just used something like OpenAI? Like right now, this is being done by the actual researchers or anthropic. But what would happen if they put out this unreleased model that's good at math and a regular person or a regular mathematician uses it to solve one of these problems? Who gets the credit for it?

7:19And maybe they say like, oh no, I did it all by myself and I never even used AI. Would they get credit for it for solving these really hard kinds of problems? This is what everyone in the math field is currently debating. Now, I mentioned that problem or those 10 problems that Astra solved. Basically, the way this worked is this is also an unreleased model, by the way. So, OpenAI and Anthropic, it's so funny. It feels like they do all their PR stunts at the same time. So, it's like OpenAI goes and hacks Hugging Face. Then Anthropic's like, we hacked people. And then Meta's like, we hacked people too.

7:52And then Quen's Kimmy K3's like, we hacked people. Or not Quen, Moonshot's Kimmy K3's like, we hacked people as well with our unreleased model. and now we have like, I don't know, all of these different models are like, we solved, you know, these unsolvable hundred year old math problems. And then they're like, oh, we did too with our unpublished model. Anyways, I find it funny, but I mean, this one's pretty cool. So in this case though, the Astra model, so not released, it solved 10 longstanding math problems that mathematicians had worked on for decades. It published a 250 page proofs that verified with formal proof checking software, the solution, like to actually get the solution to this cost about$2 ,000 in tokens, um, to generate it.

8:33And on the one hand, I'm sure people are like, well, that's a lot of money to solve a, you know, solve these math problems. I was like, if these are like famous math problems that have never been solved, they've been around for like hundreds of years. I mean, 2000 bucks, that seems pretty cheap, especially for a PR stunt like this. I mean, this is worth a lot more than $2 ,000 to open AI. This is probably worth, you know, uh, like 50 or a hundred million dollars worth of good PR for them. So makes sense why they would do this. But yeah, these are the math is getting cracked by these AI models.

9:03And this is stuff that's been, you know, pretty tricky for a very long time. Alibaba has just released Quinn 3.8 max on Monday. This is a huge AI model. It has 2.4 trillion parameters. It is on a bunch of benchmarks. It is right behind Anthropics Claude Fable 5 on the public leaderboard. So it's I mean, it's right up there is number two, The company said that they're going to release the model's code and weights next week, making it free for developers. So any developer can go and download it. They can go modify it. I mean, the thing that's cool here, it's not like they're like, oh, we released a number two model.

9:36That's great. Meta just released like a number four model yesterday or whatever. The thing that's amazing about it is that it's a number two model and they're going to release the code's open weights and like give this to anyone to go and download and use. on arena.ai's text leaderboard. It is behind Fable 5 and 3 Claude Opus models. It's ahead of a bunch of other ones. Alibaba is coming to the open weight distribution. They had a little bit of a proprietary pivot earlier this year, which was kind of showing how a lot of Chinese labs like Moonshot and ByteDance were ever, there's like kind of this weird, and it feels like it happens in sync, but there's this weird moment for a lot of these companies that are making these open weight models.

10:19It feels like if they fall a little bit too far behind, then they're like, they go proprietary, they fall a little bit behind, then they're like, okay, now that we're getting caught up, we got our mojo back, we're going to release an open weight model. Even Meta did that. Moonshots AI's Kimi K3 released last week had a 2.8 trillion parameter model. So I mean, this is 2.4 trillion. It's very similar, but it's definitely, you know, Kimi K3 is a little bit more than Quinn 3.8 max. But the parameters don't 100 % matter, especially because this is currently ranking above it or on a bunch of different benchmarks.

10:52Researchers have found a way to extract hidden reasoning from Claude, GPT, and Gemini. This is a really interesting one for me. What they're doing is they go and trick smaller AI model variants into decrypting what larger models keep secret. So inside of any time that you ask an AI model like Chachapi or Claude a question, it's not just running it to one model and spitting back the results. It takes your question. It gives it to like 16 or a whole bunch of different sub models that all review it and they all come up with their own response. And then it sends that to another model that looks all the responses and it consolidates it.

11:30Anyways, there's like these, you know, these orchestrations of models in the background, looking at it and reasoning through it and trying to come up with the best response possible for you. But apparently that is the vulnerability. So the same flaw has also leaked API keys and passwords before the companies patched it last month. So essentially, the main model was kind of better protected, but these little sub models were they could actually go and exploit them. This particular attack exploits shared decryption keys between large and small model versions. So you're feeding encrypted reasoning to a less restricted sibling model, and then you're making it decrypt and expose the hidden chain of thought from the bigger and better model that is, you know, more encrypted or, you know, higher security, but the little models, like you're using their own models against them, kind of.

12:19Moonshot's open weight model Kimi K3 produced reasoning outputs that were very similar, apparently, to Claude Opus 4.8 and GPT-4.0. Their reasoning traced across 90 different test questions. So basically the idea here is that Kimi K3 probably used model distillation. I mean, this isn't a shocker. Why wouldn't they? They They obviously used model distillation on Opus 4.8 and GPT-4.0. OpenAI, Anthropik, and Google all patched that particular vulnerability after they were notified. Researcher Alexander Panfulov says that some reasoning content can still be reconstructed and closing the hole entirely would require rebuilding how APIs handle offloaded computation.

13:00So in some of these cases, it's really, really hard to patch this kind of exploit. and they'd have to really rebuild a lot of their infrastructure and rebuild how APIs handle stuff altogether, which obviously that's a massive problem. I think this exposes a really big structural weakness in how frontier models offload reasoning. The companies share infrastructure for cost reasons, but that shared infrastructure becomes a backdoor when alignment training is different. So the distillation angle matters most for policy. I think when we're talking about distillation, if US models are being systematically copied by all these openweight competitors, which some of them are American, right?

13:38We have Meta doing that with Meta's Muse Spark. And then of course we have K3. But basically this is a technique that is this particular method is a very plausible avenue. A couple other interesting things that happened this week. Brad Lightcap, OpenAI's longtime COO is leaving after eight years to start something new. And Anthropic is going to start watermarking claw generated text. They have to do this to comply with the EU AI Act. So there's a lot going on. Thank you so much for tuning into the podcast. If you enjoyed this episode and you haven't left a review on the show yet, it helps the show out so much.

14:15I'm sure you hear me say this all the time, but honestly, if you're one of the people that hasn't left a review yet, it really helps me show up in the algorithm. I read them all. So I just appreciate hearing from you guys. If there's any topics you want me to cover or any things that you enjoy, drop it in a review, leave a comment. I will read it and I really appreciate it. Also make sure to check out AI box.ai if you want to get access to over 80 different image, audio, video and text models all in one platform for$8.99 a month. We also have 20 % off annual subscriptions. And if you want to get more tokens, there's a bunch of different tiers.

14:45You can do some really, truly incredible things with AI box.ai, especially our MCP that lets you connect AI box to Claude. And Claude can generate images, video and audio all inside of Claude, which is what I use it for a ton. All right. Catch you guys all in the next episode.