The OpenAI debacle (a retrospective)

0:07Welcome to Practical AI. If you work with artificial intelligence, aspire to, or are curious how AI-related technologies are changing the world, this is the show for you. Thank you to our partners for helping us bring you practical AI each and every week. FASI.com, fly.io, and typesense.org.

0:39what's up friends ai continues to be integrated into every facet of our lives and that remains true because you can now index your database with ai you can write more code become that 10xer you always wanted to be and you can even draft a letter for a lease on an apartment or a new property AI is everywhere and it might be time for us to start questioning is AI our friend or our worst enemy and that's the focus of the three part season opener of the award winning podcast called Trace Route Podcast you can listen and follow the new season of Trace Route starting November 2nd on Apple, Spotify or wherever you get your podcasts and this show is all about the humanity and the hardware that shapes our digital world in every episode of Trace Route a team of technologists seeks to untangle the complex question who shapes the internet.

1:26Seasons one and two gave us a crucial understanding of the inner workings of technology while revealing the human element behind tech. And season three tackles not just AI questions, but also how can we use technology to preserve the earth? Who influences the technology that gets made? And what happened to the flying cars we were promised? I think it's safe to say that the future of AI is both exciting and terrifying. So it's interesting to hear the perspectives of experts in the field. Listen and follow this new season of TraceRoute starting November 2nd on Apple, Spotify, or wherever you get your podcasts.

2:22Well, welcome to another fully connected episode of the Practical AI podcast. My name is Daniel Whitenack. I am the founder and CEO at Prediction Guard, and I'm joined as always by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How are you doing, Chris? Doing very well today, Daniel. It has been quite a past week or so. It has been quite a week. In the U.S., we've had usually people take off a couple of holiday days for Thanksgiving, but even leading up to that and during that, there was all of the craziness of what will be remembered as a very unique Thanksgiving season in the AI world, at least.

3:06Yes, it's been a soap opera, to say the least. Yeah, I was trying to think up a good pun from a soap opera title, Days of Our Lives or something, but I couldn't think up one for OpenAI, Days of Our Artificial Lives, or I don't know what it would be. I don't know. But it's definitely, I mean, not even day by day, but at some points, hours by hours, radical changes along the way. Yes. And of course, we're talking about the saga of open AI and all that happened, which is what we're going to talk about today. Yeah, yeah. I mean, I think it's what we have to talk about. And I think rather than just us giving a few hot takes, which I hope that we will have, I think it would be good to kind of step back and kind of look at the history of OpenAI, how it came about and the progression of OpenAI as an entity and their offerings, which kind of frames up some of, I think, the drama that we've seen over the past week.

4:13the week of Thanksgiving, November, 2023. You know, it's interesting, Chris, I was looking back, we had Wojcik Zaremba from OpenAI. One of the founders. Early on. Yeah, yeah. One of the founders. And that was episode 14 of this podcast. We're now on about 250, I think. Somewhere in there. Yeah. It seems like from those early stages till now, we've been kind of going on our own journey in parallel with open AI as they've had this amazing journey as a company or as an organization. And way back then when we were talking to Wocek, we were talking about really reinforcement learning and robots. I don't know if you remember at that time, they were doing these like robotic arms and they would have like stuffed giraffes and they would hit the robotic arms with the stuffed giraffes, which is kind of comical, but it was like a perturbation.

5:10They're trying to create kind of robust reinforcement learning to control robotics. And that was kind of, I need to look back at the episode. I think that was basically what we had talked about back then. That was kind of the focus. I remember being very interested in that episode and he is such a smart person, holy cow, because I had been leading up the first AI team at Honeywell at the time. And we were doing also So robotic arm work within Honeywell's business using convolutional nets to see. It was early days compared to what we're doing these days. But holy cow, he was so smart. That was one of those episodes that's really hung with me over the years.

5:50Yeah. And kind of stepping back and taking a wider view of OpenAI as a whole, I'm even looking at the transcript of that episode now. and at this time would have been, I guess, four more years ago. Wojcik said the goal of open AI is quite ambitious. It's to figure out a way to build artificial general intelligence or artificial intelligence or to be more exact, how to build it in a way that it's safe or that we can control it. And he says, let's say, figure out from a political perspective, how to deploy it in a way that's beneficial to humanity as a whole. So that's kind of one little clip from that episode.

6:30And if you look at OpenAI's founding, if we just take a look at how it was founded and how it progressed to this point of the chaos of last week, it was really focused uniquely on this problem of creating an organization that would steward us towards artificial general intelligence in a way that was beneficial to humanity. And there were various people involved in that, various funding groups. We mentioned Wojcik, but of course, Sam Altman, who has been in the news a lot, Elon Musk, Greg Brockman, and others that were part of that original group when the organization was founded in December of 2015.

7:14I want to explicitly point out, it was set up as a nonprofit at that point in time. That's correct. Which somewhat changed, as we'll talk about. Yeah, which maybe was part of the tension that led to last week. Indeed. Yeah, so it kind of had this initial board. Sam Altman was part of that. Of course, he was in the news a lot this last week. Prior to OpenAI, Sam was the president of Y Combinator from 2014 until 2019. And I know that there was some things in the news kind of trying to imply certain things about why he was let go or fired or left Y Combinator in 2019 and trying to tie those forward.

7:52I don't know that that was totally coherent for me in terms of what it was. But anyway, that's kind of his past in this startup venture backed world. I think that would be the key thing to highlight there is he's coming from this sort of venture-backed startup world, IPO, raise a bunch of rounds of funding and kind of big tech mindset, I guess. Yeah, I agree with you. I think that was the very beginning, in my sense of it at least, of kind of that tension is that it's a very different culture from a nonprofit effort in terms of the kind of the way you run your business and such as that. So, you know, one kind of key other commercial player in this would be Microsoft, who has invested at this point over$10 billion in OpenAI Global LLC.

8:44So that will be an important maybe distinction here in a minute. But Microsoft is a big player in this, which is why you might have seen Microsoft's CEO making statements during the past week, etc. Correct. For whatever reason, you know, kind of going to that point, because that is a structural concern of OpenAI that is, by almost any account, a little bit bizarre, you know, and even this podcast is almost as old as OpenAI. We're not quite there. But we were, you know, when 2019, we were operating when all this stuff was happening. And I remember it. And they've sort of created, they have the parent company, which is nonprofit.

9:23And the short of it is they have this LLC, which is a for-profit entity, which is a subsidiary of the nonprofit. But because of the way they're operating now, all of the funding, the investment, everything has gone into the LLC. And yet you have this group of disconnected investors from the daily operations operating at the nonprofit level above. of. So it's two entities that were never really intended to operate together in a direct legal manner, at least from an intent standpoint. And somehow the attorneys have made this all work so that it is legal. Yeah. When I was looking at like key takeaways in terms of what we're doing here, which is sort of a retrospective on this crazy week of events.

10:10Yeah. One of the things which is not really related to open AI, but if you're in business, it was a key takeaway is like, don't create convoluted corporate structures. No, it's not not going to help anybody. So there you go. That's a non AI tip. If if you're in the process of structuring some complicated corporate thing, try to simplify it. But yeah, you're you're right. The start was nonprofit and NPR even quoted this as kind of like it's almost like an anti big tech company. That's how they quote unquote referred to this non-profit version of OpenAI when it started, that it would prioritize principles over profit.

10:54Again, the idea with the founding of this was it would be a way for the best AI researchers in the world to help steward this really disruptive and potentially harmful technology in some respects into the future in a way that it would benefit humanity. That was kind of the framing. I think that really plays into what ended up happening, which developed over the years is that they were doing two things. They were serving that mission as they founded from 2015 on. But as we have all learned and talked about quite a lot over the years, it's very, very, very expensive to create these large models.

11:35And so you get the sense of They were constantly fundraising and they hit this point where they needed a serious infusion of cash to push forward where they were wanting to go. And I think Microsoft comes along and says, well, we'll give you an initial billion. But that also happened at the same point where this new corporate structure evolved. And I think that that was all tied together as well as I can tell from reading many, many articles on the topic to get that investment going. And so they did that. They got the new thing. They got the initial infusion of a billion. And then the year after that, they got 10 billion more from Microsoft.

12:13But by doing that, you had the driving forward on how fast can we get there and be the leader and do this in that entire startup kind of culture contending with the, our mission says we're going to do this for humanity and we're going to do this safely. And you can see they're crashing together and have been the last week or so. Yeah, yeah. Yeah. So specifically what you're referring to, Chris, is this transition from a non-profit organization to a quote unquote capped for profit, which I had no idea it even existed until the 100 time cap. Yes, until we started talking about this a couple of years ago.

12:57But yes, a cap nonprofit where the profit would be capped at 100 times any investment, which according to the numbers that you just told me about investment would be a significant profit regardless, but a capped nonprofit. And according to OpenAI at the time, and I think what they've said, this really had to do with attracting talent and attracting investment at the levels that they would need to achieve this progress towards artificial general intelligence that would benefit humanity. humanity. Now, there's two kind of key pieces here. The OpenAI Inc., which is the non-profit. So it's very confusing and maybe slightly annoying to have to refer to these differently.

13:45But OpenAI Inc., the non-profit. And we mentioned a second ago, Microsoft investing in OpenAI Global LLC, which is the capped for profit. And what's interesting is that essentially this created a scenario where the board, which had full control of the capped for profit company as a nonprofit, couldn't have a board member that would have some sort of financial stake in the for profit company, which all of that seems kind of convoluted, but let me say it again. So like the board that's controlling the OpenAI entities could not have members on it that would have financial stakes in the capped for-profit OpenAI Global LLC, which means that Microsoft, for example, as a huge investor in OpenAI Global LLC did not hold a board seat on OpenAI Inc., which will play into kind of the timeline that we'll talk about here in a second of what's happened over the past couple of weeks.

14:56And one other thing to throw in on that is, though I am not an attorney, I'm fairly sure that the 100-time cap was one of the mechanisms by which they could make the for-profit fit into the nonprofit because pure for-profits in theory have essentially unlimited ability to generate profit. Nonprofits are not allowed to have profit. You know, you must use those expenses. And the 100 times caps, I think, was something of a bridge. So I say that from a, I did not go to law school to learn this, but I'm pretty sure it's tied in there somehow. Well, Chris, in this saga of what's happened, we have this nonprofit, originally OpenAI Inc., that has spun out this for-profit OpenAI Global LLC, which has received a huge amount of funding from Microsoft.

15:44And as we all know, over the past couple of years has really become the dominant force in the AI industry with releases of all sorts of amazing technology and tools. And of course, most recently, ChatGPT. So it might be worth just kind of giving people some context around this of like, what happened over those years with OpenAI that made it such a driving force, which eventually led to, I mean, you could have a company that has a convoluted corporate structure and it blows up and the CEO gets fired and, you know, no one hears about it in the news and our lives pretty much go on, although it's probably unfortunate for their lives.

16:30But here, this had an impact kind of on the whole AI industry, and I think will have an impact on the AI market moving forward because open AI was such a dominant force in the industry. So maybe we could go back and visit some of those milestones and the history that came along from those early days when we were talking about robots and stuffed giraffes to now where we've got GPTs, I guess. So back in 2016, so this would have been almost when we were starting the podcast, NVIDIA gifted the first DGX1 supercomputer to OpenAI. It's quite a gift of the day. Yeah, so they were kind of early on into this wave of really powerful, cutting-edge, GPU-powered supercomputers that could train larger and larger models, which is one of the things that's, of course, led to these very powerful foundation models that we've seen in recent years.

17:31I don't know if you've got, they haven't gifted you one, you have it in your garage or something, Chris? No, my first DGX one was, which was when that was all there was on the DGX line, the original one, I got at Honeywell and we had it under a desk for a while because we got it and then we had to figure out how to get in the data center and stuff. But yeah, that was, I remember thinking, wow, I got it. I have an AI supercomputer under our desk here. This is amazing. Anyway, a little side story. And in those kind of years after that, you see what maybe I would consider, and I know it's hard to make generalizations like this because there's a lot going on at OpenAI even now.

18:12But in those early years, I think you really saw this kind of exploration phase of setting up the right tooling, setting up the right compute, letting researchers explore various things like the robotic stuff, like RL, computer vision things. and you saw things like the universe software platform open ai gem which i know i've used in workshops for like reinforcement learning type of things so you really saw this kind of wide range of exploration in those early years leading up to february 2019 when gpt2 was announced and this was the first kind of, from OpenAI at least, the first major language model or foundation model, base model that gained a lot of attention for its ability to output really human-like coherent text.

19:10Of course, is the first in the line leading up to the latest GPT models that we see now like GPT-4. But that was in February of 2019. Still something that I think was basically just being looked at by people like us. Yeah, episode 32 of our own podcast was covering that. Oh, there you go. Episode 32, GPT-2, episode 32. It's interesting now, Chris, when I'm helping people learn how to fine-tune a language model or something like that, I'll often use GPT-2. And it's interesting to see because now this would be considered a very small model. And not only that, but it's open, right? So OpenAI was, by its very nature and stated aims, going to be very open with its research and models and IP and all of that.

20:06So GPT-2 is on hugging face, and it's a great model to use even now to kind of figure out how to fine-tune language models. But it's just interesting to compare that to, for example, GPT 3.5 or GPT 4. I'll never see those models in terms of downloading the weights and all of that. It became a different org. Yes. And with their own reasons for doing that, and that's their decision. But yeah, in 2020 then, so a year later, OpenAI announced GPT-3, which was a language model that was more extensively trained on this sort of huge corpus of content from the public internet. And, you know, so it was a lot of scraped data and all of that.

21:01And I think this is really where you started to see some pretty crazy outputs from these language models, what people might have thought not possible. They started to see this with GPT-3. This is also where you kind of see this shift within OpenAI to from releasing GPT-2 as a model to releasing GPT-3 as an API with a more gated release, a slower release. I was even reading in some of kind of the lead up and debriefing from the chaos of last week that you could interpret some of this slower release of GPT-3 as an API and a product to really this tension that you're seeing between a startup kind of mentality wanting to fail fast, release things fast, learn from things in public, and this kind of more shielded, non-profit, good of humanity side that is wanting to do things maybe in a more slow way that ensures safety and releases things not with harmful outputs, that sort of thing.

22:12So you see this kind of starting to really collide and come together, I think, around GPT-3. GPT-2, GPT-3 were language models, caused a big stir in the AI world. Again, still not something that public, you know, as far as chat GPT, wasn't something that a lot of people knew about in the general public. I think even when they released DALI in 2021, this was kind of text to image model where you could, you know, put in your prompt, the astronaut riding the horse on the moon and you could get that. And I think the public saw that as still kind of like a novelty and that sort of thing. And that led up all the way to December 2022, of course, when AI started, when ChatGPT was released, at least in free preview.

23:04Which changed the world. Yes. When we were talking, even going back to Dolly, people were aware that there was this thing and they saw all these crazy AI pictures. but if you would ask someone who wasn't following this industry the way we do most of them out there they would have said oh i've seen some of those photos but i couldn't tell you what the organization was at the time they they didn't know yes that really changed with chat gpt the whole world woke up to this stuff yeah and i mean we've been talking about opening on this podcast pretty much non-stop since then and everyone else has as well of course chat ChatGPT isn't the only thing going on in the AI world, and hopefully we're representing that on this podcast, but it is certainly a driving force.

23:49And we've mentioned it a lot, and that's why the events of last week caused so much stir. It's worth noting, still things happened after ChatGPT, right? We had GPT-4. We had, I think what you saw as a shift in public discourse from Sam Altman and Greg Brockman at OpenAI really related to, hey, we need to really provide recommendations for governance of super intelligence, governance of artificial intelligence, and also at the same time, rapidly releasing new products as well. So you see this, again, this tension, right? Like, hey, we want to talk publicly about governance of super intelligence and regulations around this.

24:35And at the same time, we're going to have our OpenAI Dev Days and release four new offerings, which are going to blow your mind and immediately permeate all industries, right? So you've got GPT Vision and the GPTs, plural, which is kind of easier ways to create customized models and systems and RAG workflows, along with other things like their assistants, playground and API. So you have, again, this like you just see these two things kind of coexisting where I'll use the words of NPR. They quote say, two competing tribes within OpenAI, adherence to the serve humanity and not stakeholders credo, and those who subscribe to the more traditional Silicon Valley MO of using investor money to release consumer products into the world as rapidly as possible, end quote.

25:34So you see this even in this past year, I think, in public discourse and in the release of products. And that leads us all the way to November 16th. So I don't know. Where were you on November 16th, Chris? You know, we're talking like this great historical moment. As we're recording this, this is what Sunday afternoon. This was a week and a half ago that we're talking on a Thursday. And so the whole thing happened in that week before Thanksgiving. And it was done in one week from a Thursday, basically to a Thursday to a Thursday for all practical purposes or Thursday to the Wednesday. and it was working and uh happened to look at the news and the first thing i saw was open ai had fired sam and uh was like which was crazy because it was like right after his keynote i know a day it was a day later it was a day after his keynote yeah yeah it was just mind-boggling yeah so apparently he got a text on thursday night from one of the co-founders asking him to join a google meet on Friday.

26:39And they had already kind of tapped the CTO, Mira Mirati, as the next CEO. And on Friday, literally, Microsoft learns of Altman's firing, quote, a minute before the world does again. So it was on Friday that we heard it actually, not Thursday, but yes. Yeah, yeah, exactly. So it happened while it got into motion on Thursday, and then some of us heard about it on Friday and apparently also Microsoft, which was just, you know, so I knew about the kind of convoluted corporate structure and all of that, but it was just mind boggling to me that Microsoft was not informed about this given their investment in the for-profit entity.

27:28Which I think they have 49 % stake in, if I recall correctly. Correct. Yeah, it's something like that. I don't remember the exact figure, but certainly they've invested over$10 billion with plans to invest 10 plus more billion dollars. And of course, they've integrated OpenAI, ChatGPT, etc. into Azure, into Bing, into Microsoft Office 365. and so yeah it just sort of boggles the mind that this happened in the way that it did so then on the 17th then open ai releases a statement that sam altman was ousted and uh later that evening greg brockman quit um announces he's quitting they got rid of him as the chair uh so he was fired as chair but then he turned around he was still president and he was and he quit as president standing uh to show sam altman that he stood with him okay well that brings us to the point chris sam and greg are out and uh microsoft on that next saturday then kind of going day by day microsoft releases a statement saying that they and i don't know if you saw the video um the Microsoft video, but essentially they immediately offered to hire Sam Altman into Microsoft.

28:56And Greg and anybody that left open AI would enroll. Yes. And I think I could tell in those videos, just the sort of still shock and unbelief in a lot of people. So Microsoft releases these statements. People of course start wondering like, why did this happen? This had to be something really bad that Sam did. Why would this ever happen? And there was some reporting and some statements that basically were kind of confusing and murky that they were saying, no, it wasn't like any violation of security or privacy practices or kind of malfeasance on Sam's part. But you kind of got like these vague hints of, oh, he wasn't completely forthright with the board and his communication didn't allow them to properly make decisions.

29:52And so you get that kind of mix of stuff going on until November 19th when Altman announces that he has been hired into this new research unit of Microsoft and posts on Twitter slash X. The mission continues. And in the meantime, Mira was the first designated CEO. They had another CEO for another day on Sunday. And then a third one, they went through CEOs, one a day for a while there. Yes. I was on my Zoom calls when I was hopping on. I was asking people if they had been asked by OpenAI to be the next CEO yet. Because it seemed like they were working their way down a list. I don't know how far down the list I was.

30:40Kind of sad I wasn't in that top 10, but they were definitely working their way down some type of list. And I forget the name of the one that came after Mira. It's escaping me right now while we're talking. But the OpenAI employees were reported to, on their internal Slack, were using, I'll gently say FU, and showing the middle finger on their Slack. Apparently, the employees had had enough. By the time we got to mid-weekend, going from Saturday into Sunday, the employees started finding their voice, as we'll hear about next. And there were reports basically that up to 95 % of the employees within OpenAI were going to depart OpenAI if some deal wasn't struck to have Sam Altman return as CEO, which obviously would be the end of OpenAI, at least as we know it.

31:31But so I don't know, I could definitely tell you that. And this is one thing I'll highlight later on is all of these people who have relied on OpenAI as the sort of bulwark of of the industry and integrated this across their products. We're really in a state of panic, right? Because it's like, hey, you know, the OpenAI is still saying they're going to provide great support for all this stuff. But also you're saying like potentially 95 % of your employees are leaving. So where does that leave all of these? So it was really kind of a time of reckoning where it's like, hey, you know, we've built a whole strategy around OpenAI's products and models.

32:15So now what? And, you know, the resilience of relying on this kind of single family of models, I think, really, really was showing in that moment. If only there were services that could help people find a variety of different models and not be entirely dependent on a single family. Prediction card. I didn't say that. Yeah, yeah. Sorry. This is a point we'll make later, but I think it does. we even a few weeks ago had some comments about with the executive order, is there firming up of the market around these kind of foundation models because of the regulatory burdens that would start starting to be put on them, right?

32:56And I think that this kind of opens up the field a little bit more, regardless of what has played out with OpenAI. I think people are really wrestling with the fact of, hey, what else is out there? And of course, you've got amazing players in that space from Mistral to Mosaic ML to Meta and what they're doing with Llama 2. And people really finding, you know, of course, we're helping Prediction Guard helping people build with these models, but a lot of people are finding a lot of success with these models. And I got a number of messages during this chaos about, hey, you know, we weren't thinking about trying open models or models outside of the GPT family prior to all of this craziness going on.