The Dispute Over Scientific Credit

2:48:26going? So, we are back now with our lighting in not nighttime mode. For those listening 3 hours in, uh, welcome. We are going to now take a look at a little bit of a timeline and some of the connected issues that have arisen around this OpenAI release. And so, just to give us a grounding here, um, you know, a little bit about how did we get here and everything that happened around it. So the announcement that set off these two conversations, what does the mathematics establish and who gets the credit for the work behind it came out on September 8th. So that's our first

2:49:07date in the timeline. Now, interestingly on September 9th, it was reported that OpenAI had revised the PDF from the previous day. For those who saw on the first day, one of the complaints was this is not citing any of the people we actually just talked about. And at the time I saw the names and didn't know what it meant. >> Now we've built the construction of why the work of Cordoba and Zurua >> was was fundamental to this solution. And so they they sort of changed that version and now it includes those references. >> Yeah. Yeah. They went back and they included the citations. So the marketing

2:49:49team put it out too early. They felt time pressure to put it out too early. Um and this obviously matters because um citations are fundamental. >> Huge. Yeah. >> Uh to you got to give credit. >> You got to give credit where where it's due. And so acknowledging the earlier published work that led them to do this is sort of the answer to one issue. Uh, but there's another issue that's related here, which is whether or not OpenAI had access to unpublished work from folks who were using OpenAI's

2:50:31models to build on Cordoa and Zurua's work independent of their own announcement. And we're going to get to what that allegation is, but on September 10th, OpenAI said that an investigation has been put out that rules out Tristan Buckmaster, who we talked about earlier, his allegations that they used the work of himself and Levant El Pog to accelerate their conclusion to get to this solution. So, we've we've talked about Leavant El Pog before. A new entrant into this conversation is

2:51:11Tristan Buckmaster. And so who are we who are these people? Who are we talking about? Right. So Tristan Buckmaster is a mathematician at NYU. As we mentioned, uh Alpog is a member of the technical staff at Anthropic >> who makes the model Claude that people may be familiar with not at OpenAI. And this is interesting because this story goes back according to Buckmaster all the way to August 15th and 22nd as it relates to the work that they were doing together. And to kind of frame this up, this is a statement which you talked about earlier in page one where

2:51:53Buckmaster is talking about this collaboration that he was having with Alpog where they credit uh Diego Cordova and Luis Martineza for the work they were trying to do using both Claude and anthropics models to build on their understanding to try to get to this finite time blowup result. Um, and he noted that they used substantial AI assistance, but in his first page of his statement, um, and this statement, uh, comes out, uh, in this time frame, he's he notes that he believes that, uh,

2:52:34Zorua deserves the Fields Medal. >> Yeah. for this solution. >> These are the guys who created that Russian nested doll of vortices that ultimately OpenAI used to create the solution. >> This this is exactly right. So in his statement he says that himself and El Pog they date their busines and Oiler blowup results >> Mhm. >> to August 15th and the lean verification for that to August 22nd. and they wanted to work on the proofs a little bit more because what the models put out was not understandable. Yeah. And it was not sufficient. >> Um, while this is happening, there are

2:53:16rumors starting to fly around that Anthropic is leading leading up to potentially an IPO init initial public offering where they'll go into the public markets and raise a ton of money. And as a part of the rumors, it was alleged that Anthropic was working on some open math problems, potentially some Millennium problems. And um there were rumors that Anthropic had two millennium problem solutions. Yeah. >> And I think some of these rumors were exacerbated by the fact that Levant Alpog tweeted um Austous Mirabilis. >> That's this is this is exactly right. which is like a reference to the annus

2:53:56miraabilis of 1905 of Einstein the miracle year of Einstein where he had four papers that put him on the map this is him saying that August is a miracle month for anthropic and what's interesting is this allegation that open AAI was uh triggered to start taking one of their new internal models and start pointing it at these millennium problems Sam Alman actually tweeted that one of the reasons they started doing this was because of the rumors they were hearing. So this is it's it came from the source itself. >> Um and what's interesting to note is >> OpenAI was using an internal model to

2:54:37see what they could do >> and they were using this with this idea of using agent teams like a swarm of multiple agents to do so. Okay. So that started on September 1st, >> right? >> 7 days before our September 8th, you know, announcement, right? On September 3rd, >> uh, Buckmaster contacts Open AAI because he'd been hearing rumors through his network, which he describes in his statement that maybe OpenAI was looking at this and wanted to kind of >> uh clarify what was going on there to potentially deconlict. In the second page of his four-page statement, um

2:55:19it was understood that OpenAI wanted to date its result as September 5th, followed by 17 hours of lean verification. Um and then Buckmaster ended up having a call with members of OpenAI on September 6th, which included a Sebastian Bubck. Al Pog was not on those calls. And the conversation, the substance of the conversation was, you know, Buckmaster was trying to say, "Hey, we've been working on these things. I heard rumors that you guys are working on something." And there was a back and forth about how to kind of deconlict the release of these things. >> Um, ultimately, Boobeck was >> then trying to provide some options for what could happen. uh you know it was you can release your

2:56:02your papers first on Oiler blowup and then we'll follow up with our Navier Stokes um where uh you are the author uh talking about Buckmaster. Yeah. >> Um or if you want to be uh a named author on our Navier Stokes paper, we can work around that. But hey, like you know, we can't really involve Alpog. Yeah. uh the reason being it would be you know uh Bubbeck's point was it would be a little bit weird for an open AI research paper to include an anthropic researcher and in the statement he talks about um how there was it got a little contentious um and he asked you know when did you guys start and there was a little bit of

2:56:42like oh well we're not going to kind of tell you when we started and then he was curious as to whether his work in codeex opens tool was used as training data for them to get to a solution. >> Yeah. >> You know, not really a lot of commentary. >> Yeah. >> Around that. And there was a lot of back and forth. And basically Buckmaster um part of his accusation was that you know there were comments made about this idea of like how you know why would you want to ruin your career by going up against us which we'll come back to Bubck's response to that. There's just there's a lot of weird tension because >> Buckmaster basically senses that you

2:57:24guys may have stolen our work >> and are trying to frontr run us. >> Yeah. >> And that's not going to happen on my watch. >> This is September 5th and 6th. >> Total side note. On September 7th, our buddy Terry Tao, as you like to say, uh talked about three related papers about smooth forcing results for incompressible porous media for Businessesque and three-dimensional oiler. And this was from a totally separate research team. And but what was interesting to the point you were bringing up earlier is people were coalescing >> Yeah. around this and it just so happened that these papers had come out

2:58:05noting that they were two of them I think were fully formed. One still needed a lean verification but everyone was hearing these rumors within the math community and so everyone was kind of trying to effectively frontr run a lab. Yeah. >> Taking all of the credit. um Boo Beck on the ETH, which is when OpenAI put out the result after we had uh Buckmaster basically frontr run this and put out his four-page statement that we talked about outlining this. He talked about the result, how he worked collaborated with Alpog, explained how they built off of Zuru and Cordoba's work and made the allegations. Bubbeck comes out with a

2:58:46response where he basically says, "All these allegations are false." >> Yeah. >> Didn't happen. >> Didn't happen. >> I regret the language I used about why would you risk your career and I said so on the call, but everything that Bugmaster is saying is nonsense. >> Okay? >> Right. Total denial. They're crazy. Okay. Total denial. So part of the tension here, there's still an unanswered question about whether they used the sessions from Alpog and Buckmaster Open AAI. >> Yeah. >> To have a starting point. They heard the rumors.

2:59:29>> It it was known that these guys were working on this because they were working on it for months. >> Yeah. And and they even said that the the recent model that they're using was trained like late started training late August. >> Late August. >> So it's after Buckmaster and Alpog have been using this codeex thing for a very long time because again they got their results uh Buckmaster and Alpogi had the results August 15th and 22nd for Oiler Blowup. >> Yeah. And they had the results which means they had started using this thing for a long time. Right. and they mentioned that they were using the previous version I believe it was soul uh 5.5 or 5.6 of open prior even to Astra and at the end only at the end

3:00:10>> uh did they do a little bit of Astra to plug in but this is an important point they were using the previous state-of-the-art version of open AAI to get to to models to get to that point open had an internal version >> that ultimately came Astra but they had even a better version than the public version of Astra that they were using internally to do this >> yeah and one one thing I want to say is that like um the that new version that started training in August that could be using all of the codec data that users have put in. Bingo, right? It might be anonymized. >> Yes, that's the thing. And so they can't

3:00:50say it was specifically your sessions, but if in their settings they did not turn off the use my data for training setting, >> it this is where the gray area arises around a lot of this. And I dude, there's even a grayer area. It's like what if they turned off the settings, but all that means is that now whatever chat they have, whatever chat log they have is just not tagged to their p specific profile. Like they'll still take the text and put it into a giant bucket of anonymized text. The question of theft here is, in my view, still unanswered. Mhm. >> Um,

3:01:31and these are two independent people. Buckmaster making his comments. Bubbeck defending that OpenAI didn't do anything unourred. OpenAI posting on their account that they didn't do anything unourred. Uh, it's a he said, he said. >> Yeah. >> And it's still unresolved. And what's interesting is this now has totally brought up so many colliding arguments that starts with academics versus the labs, but then now spreads out to a whole variety of conversations about who benefits from this stuff, who gets credit for these

3:02:13discoveries. um why are now people talking about human instin human extinction and there are a couple of cohorts I kind of want to talk about that are a part of this conversation. >> Mhm. >> So you have the labs and the people financing those labs. They have a financial incentive to show that these models are doing incredible work. >> Yeah. >> Because it lines their own pockets. There is some tension between the frontier labs in this case in the US it's anthropic and open AI and the financeers or the investors because the investors also have bets on AI at large

3:02:57and generally that are accelerated by the frontier labs but are not solely dependent on the success necessarily of the frontier labs. You have the research community and the research community uses these tools in ways we've talked about on the show in narrow contexts not in a general or super intelligence context that is really valuable. But there is now this open question about is the work that we're doing by using these tools and both anthropic and open AI have free access for researchers that get more powerful things or free credits. is that basically a carrot to get them in so that they can train on the frontier of

3:03:39the smartest people to then frontr run them on any number of different things. I'm not saying that that's true, but people are asking that question now and it's making it difficult for researchers to understand how should we think about these tools. You have the governments asking a different question. They care about jobs. >> Now we're looking at multiple industries, not just coding. this thing's getting very good in other areas. Our job is to keep people employed because unrest arises when people are not employed. >> Yeah. >> Is this going to be a problem that we have to deal with? And then the and China >> national security. >> Yeah. >> We have to compete in the global stage.

3:04:20And so this is the thing that's always brought up. It's like, oh well, we have to compete with China. And I think the central point I want to get to as we look at all of these things is that multiple things can be true at the same time. Arguments from each of these different parties as it relates to the AI conversation can coexist and are not necessarily mutually exclusive. It can be true that the capabilities are increasing >> and that matters and there's risks associated with that. It can be true that the labs and investors have a financial incentive. It can also be true that researchers in these organizations who are bringing up fears about the pace

3:05:02of progress and our inability to contain it in a way that causes damage is sincere because the researchers at these places and the executives at these places do not have the same incentives. They might diverge. >> It is true that a regulatory infrastructure probably needs to be in place. It is also true that these companies, the frontier model companies might want to have a regulatory capture approach because that prevents other entrance from coming in. It is true that understanding what this does in our relationship to China matters. It is not necessarily true that we cannot get global collaboration just because we're in geopolitical competition with China. Because ultimately, do you think that the CCP,

3:05:44the China Chinese Communist Party, wants runaway super intelligence that's accessible to everyone in their environment where control over every aspect of people's lives? >> Wow. Yeah. >> Is the way in which they maintain power. No, >> that's a good point. >> Why are they going to build and release a thing that destabilizes their strangle hold over their social, political, and economic system? That to me is nonobvious that it makes sense for them. Yeah, >> they also approach it from a very different perspective. They look at AI in narrow implementations in manufacturing, in scientific research, but their power dynamics are very different. So these are all aspects of this conversation

3:06:24and a lot we we just need to take them step by step. So one thing I want to talk about in this context right is for most of us we experience the idea of AI through going to chat GBT or going to claude and we have a single instance of what is or was a chatbot. If you have not used any of the frontier model companies in the last even 3 to 6 months you have no idea what we're talking about. It's totally different than December of last year and January of this year. Um, so

3:07:06what we now have with these systems are the transition from a question and answer conversation that does not maintain context over multiple conversations to this concept that folks have talked about about uh agents. And