OpenAI’s Windsurf Crash, Grok’s Wild Week, Replace Tim Cook? — With Aaron Levie

Good to be here. What a week in tech.

Yeah. I mean, in this industry at this point you never get even just like one piece of news. It's always multiple pieces of news embedded in one major thing. So this is sort of a multi-part announcement I guess. So Windsurf has been one of the faster growing AI coding platforms. It's an IDE that is built off of VS Code that lets you have agents that automate your coding and quite successful particularly in the enterprise. They were one of the first to really kind of nail the enterprise-oriented sales motion, lots of protections for data and your codebase that they had a good fit on. Varun's a fantastic founder entrepreneur and the expectation I think was that they were going to be acquired by OpenAI, help OpenAI really kind of boot up their coding efforts and clearly that's now off. Obviously the rumor in that process was there was some structural issues with maybe the Microsoft terms and different parts of that deal. No one has ever kind of explained exactly the problems there but now with that deal off and Varun and team going over to Google, it's a recalibration of the market. It's actually interesting. So the thing that everybody should have been thinking this entire time actually was where is Google in AI coding? Because right now you have Anthropic from a model standpoint tends to be seen as the leading model for coding. And then they also now launched Claude Code. OpenAI launched Codex which is a very strong offering in an agentic coding experience. And so the odd man out there is Google where Google is a very deep engineering-centric organization and so one would have imagined that they would want to be front and center with AI coding. The Gemini 2.5 model is seen as very good at coding. But again it's in this a little bit of no man's land because they neither have an IDE nor do they have like the best coding model from Anthropic. So they had to do something in this space. And this is a pretty exciting move to launch into.

Yes, although it's not obvious to me what leverage they have in that dynamic. So I think everything at this point is basically just rumors and conjecture. It's very clear that we do best on this show. Yeah, exactly. Sure. Or basically our entire industry at this point. But I don't perceive that OpenAI is constrained by anything strategically at this point vis-à-vis the Microsoft relationship. So I doubt, I think the rumor was more there was things like IP issues and other dynamics with the acquisition but again it's all rumors so impossible to know. These deals fall through for a variety of reasons. But I would not be surprised if OpenAI continued their motivation for needing to be in this space more aggressively. And so I doubt this is the last that we hear from OpenAI on either IDEs or coding in general. And certainly they're very committed on the Codex side. And people have had great experiences with Codex, their AI agent.

I think it's fine. They are somehow juggling building some of the world's largest data centers, massive energy needed for that, massive GPUs, they're acquiring Jony Ive's company, they are releasing models at an incredible cadence, the rumor next is an open source model. So I think they have probably 50 different things going on, this being only one of those activities.

Oh yeah, for all AI tokens in general. Yeah, I think it'd be fun to look at a graph of this. I mean, it tends to be one of the highest volume, if you look at what's the relationship between a human and the amount of AI that they can consume, coding absolutely would be the peak use case right now. There's no other human task that one person could cause so many tokens to be produced. Like deep research is great but you do it maybe once or twice a day and it's relatively confined. Summarizing information, very efficient, not very token-heavy. So coding is definitely the one that is like this incredible, one person could cause thousands of dollars per day of GPU expense if they really want it. So I think this is going to be the killer app for the foreseeable future in terms of just sheer volume of tokens. And so this is why it's such a big prize. And Google again, it's funny actually just timed well with your Sergey and Demis interview. I mean, with Sergey back at Google, there's these little small nuances that you probably don't want to overly extrapolate from, or anecdotes that you don't want to overly extrapolate from, but I think Sergey being back at Google is a very interesting thing to consider around how this is a company that has a great operator in Sundar, an incredible AI innovator in Demis, Jeff Dean deep on research and science, and then now this kind of hardcore founder in Sergey. This is not a company that is going to lose the coding battle. There's no way that Sergey is sitting around being like, oh, I'm going to use Anthropic to code the next version of a feature I'm building. He has to make sure that they're using Google's technology. That is obviously a point of pride for any founder is to make sure that you're building the technology that you're using for the domains you're going after. And so you just have to imagine how committed they are to solving this problem and Varun now and team are going to be one more way to accelerate that.

Well, it's not crazy. I mean think about all of these founders right? Greg Brockman at OpenAI, Sergey, Zuck, I mean they don't want to walk around their office and find out that the thing that everybody's really excited about is somebody else's model. So that would be like if you worked at Facebook and everybody was on X all day long and not using Facebook. So these things are very major points of pride for these founders which makes the race so exciting to be watching.

So, I'll give you an example of how crazy this gets. I was talking to a founder this week where I mean every day I see something that I'm just like I have to completely reassess my estimation of the future. This founder is so right now a solo founder. He has many different, I don't know if it's five or ten or whatever the right number is, many different agents in the background going off doing individual parts of his codebase as well as the marketing website that he has to build for this product that he's working on. And so he is effectively as a solo founder a manager of multiple agents doing all of this work. And then his job and basically the new form of engineering work out there is to come up with incredibly precise prompts that are super tuned for his use case and then kick off all these agents in the background that are going off and doing work and then he goes and reviews their code and he integrates that code into the broader codebase and then effectively is there reviewing and auditing all of their work. The reason why that's so impactful or meaningful is that one person could literally be causing tens of thousands of dollars a month in AI consumption because of just the single actions that he is doing. So while that's not going to be the behavior of everybody on the planet, that is a massive force multiplier of human to compute ratio that we've just never seen in computer history.

Yeah, I mean I think it is working empirically. Obviously you can, and we saw this with Meta a little bit, you can sort of train your models to perform better at some of the evals or the benchmarks which then somewhat can delude you into thinking that the model is better than it is where it's really just better at these kinds of tests. However, right now I think most of the evidence is that this is a very high performing model kind of across the board. It continues to align with the theory that more compute, more data generally is going to produce better models. And then they're doing some novel things that I think are emerging across the industry, but maybe this will be the first kind of real commercial model at scale that does this, but they have a model called Grok 4 Heavy that has multiple agents go off and execute basically the same task. And then they go and review their answer for which answer these agents think is the best result. And so this is a great example of how you can have a lot of compute in the training process but then also have lots of compute in the inference process where you just have the model working harder and harder and harder to produce better answers which is clearly producing great results. They show the scores of what Grok 4 Heavy can produce. And I think that will become a standard across the board. So, I think it's absolutely a continued improvement in model quality, model performance, and we're super excited that the scaling laws are continuing to play out and this is just more evidence of that.

Well, I would kind of say that both can be true. Diminishing returns is first of all it's a relative concept. So diminishing relative to what rate? But I think the way to think about it is if you think about a curve that eventually sort of asymptotes, all that matters is where are you in that curve. So if a curve is sort of like this and it's asymptoting right here, well if we're right here that's bad. But if we're right here it will be quote unquote diminishing returns but you haven't asymptoted or plateaued yet. And so all that matters is where you are on that curve and trajectory. And you can see based on some of the evals, it's not as if there's going to be a 10x improvement in intelligence anytime soon simply because some of these evals we're already at 80 or 90% of where the eval is. And so there isn't even room for the model to be 10x better. So that might mean though that you have to apply five or 10x more compute to get to that next final, that last mile of intelligence which again would be both diminishing returns but also something we would still continue to drive as an industry because you're still going to appreciate that quality difference. And so that I think is totally fine. In general talking to enterprises, we're already for the most part, with many exceptions, we're already in a position where the technology well exceeds anybody's ability to adopt all of these benefits so far. So simultaneously we want the progress to continue at this exact rate and there's most use cases on the planet it still could be benefited just by even what today's models can do. So we want more innovation, we want more compute, we want more intelligence. But even if you stopped right now, you'd still have massive amounts of economic gain delivered from what we've already created.

I'm not going to make any predictions on that front. Because the, but do you think this is evidence for or against? I think the smartest people on the planet have two totally different views. And so I'm not going to get in the middle of that one. I mean clearly you have people like Ilya where it's rumored that he's working on a different architecture or maybe a different path. And then obviously you have other people that are, let's just throw more compute and data at the problem. I think you can start to sense actually as an industry that the AGI term has actually kind of gone into the back seat and obviously more of the conversation is around super intelligence. And I think there's more and more comfort around this idea that actually the race really is just how do we build intelligence that far exceeds a human and what will the economic and societal benefits be of just even accomplishing that which are massive. And I have always sort of found the AGI thing to be particularly squishy as a concept. In the B2B world I deal way more with just utilitarian concepts and so super intelligence and this idea of we have AI that will far exceed a human, that alone is enough of a breakthrough to be shooting for and I think what you're seeing with scaling is we will be able to certainly accomplish our collective definition of super intelligence with the current path we're on with scaling laws.

Well, if you, is that right?

Who would that be? Dario probably.

He said we know what to do, we know what to do, and he's investing in Stargate so seems like he's a scale maximalist. But what's interesting is actually I think that you'd be able to get them all to say the same thing which is this category that says we need a new idea are probably AGI maximalists and then there's another category which is like actually it's already proving out the economic and societal advantages of even our current approach to AI. So just let's keep running that for as long as possible and we'll just keep eating out more and more benefit. Like you could already dramatically improve every healthcare experience on the planet just by using whatever the latest state-of-the-art model is in every area of healthcare. Like everybody will absolutely get better doctor diagnosis. They'll get better healthcare. The doctors will be happier when they transcribe all of their conversations with patients with AI. And that's just today's state-of-the-art. We don't need any new breakthrough just to have that ripple through everything that we do. If every engineer on the planet had background agents that were checking for bugs or writing new code for them or updating their libraries, all that longtail work that's really inefficient and not enjoyable, already the economic advantage of just today's architecture would be massive. So I think you can basically be happy about both outcomes. The super intelligence track with more scale is a great track to be on and we're just going to get more and more benefits and the sort of we need a new idea AGI maximalist, that's fine too and that's just upside if and when we discover whatever that thing is.

That's a question for me. I mean unfortunately I don't know if there's been a full postmortem as to whether that was a training issue, all of it's in the weights to be Mecha Hitler, or if

That was a system prompt issue, in which case you can do quite a bit with a system prompt to effectively change the direction or path of what you want the AI to respond to. So to the extent that it was as simple as they used to have a system prompt that said, you know, please be politically correct and be thoughtful and make sure not to say anything offensive. If they used to have that and then they basically said, actually no, say anything you want, then in that latter mode, users could certainly kind of coax it into doing Mecca Hitler stuff.

And so I think it's sort of unknown how they trained that model, how much of this was system prompt. For being able to remove that as a risk factor, I think it's sort of well understood what you need to do post-training and what you need to be doing from a safety standpoint. And then it's really just a decision of the model provider and the application layer of how to implement those things. But I thought it was obviously a ridiculously bad situation. Deeply obviously offensive and dangerous, but also not really that much of a meta story about AI simply because you can get these models to do anything you want. And the whole thing is as an industry, you're kind of working toward trying to keep these things confined within a particular pattern of behavior and communication style.

It's an experiment of how to achieve it. So listen, he's always claimed he's fairly centrist. So that would make it pretty aligned. Yeah, I mean, they clearly keep stepping on the rake and the rake keeps hitting themselves in the face. But I have faith that they will find a way to work through some of these kind of ridiculous situations.

Yeah.

Well, I think it's a fantastic question and it's absolutely worth thinking about. Do you remember like 10 years ago Microsoft had an AI chatbot I think called Tay or something?

Yeah.