Yesterday Microsoft held a special event in Redmond to announce the new Bing, which incorporates a ChatGPT-like model into search. Microsoft was coy about what OpenAI model undergirded the new Bing, other than saying it was more advanced than the GPT-3.5 model behind ChatGPT; perhaps this is the launch of GPT-4.
The biggest difference from ChatGPT is that while the interaction model is similar — basically a chat-bot, which launches on the side of search results, but is easily accessed in a pure chat interface — it is dramatically more aware of the current state of the web than ChatGPT is. You can, for example, ask it about current events:
There a few other differences from ChatGPT. Notice the links to references for the information, for example. In addition, while you are waiting for the answer Bing tells you what it is searching for. I think the purpose of those prompts is to provide assurance that you are getting real results, but it did feel a bit less magical than ChatGPT.
Here is another example of Bing’s ability to ingest recent articles that is pertinent to Stratechery:
That’s a decent enough summary — and again, a big departure from ChatGPT in its ability to incorporate new information — although I posted an example on Twitter where Bing basically made up an entire paragraph that I didn’t write. By the way, I do wonder if Bing received an update in the last hour or two; recent answers have been a lot more concise than what I was getting when it first launched. I asked Bing just that:
As this answer notes, there are two parts to the new Bing: the first is the OpenAI model, whatever version it is; that provides foundational knowledge and the ability to communicate in natural language, and, like ChatGPT, it is apparently only current through 2021. The second part is traditional Bing search, and Yusuf Mehdi, Corporate Vice President of Search, argued that Bing’s key breakthrough was in marrying these two approaches:
For the billions of queries that are going unanswered, we have seen new attempts to try and address the problem. As you all know there are vertical search attempts: Amazon has done a better job for shopping, YouTube’s great for video, Reddit is a great place to get advice. The benefits of Search are well-known: it’s fast, it’s timely, and there’s a great business model.
And then, more recently, there has been another vector; more disruptive ideas like leveraging AI to answer questions directly and to generate content. These are amazing as well, and they show what’s possible.
But what if you could get the two to come together? Not only would you get two things in one, but we think you could actually solve the problems with each and you could get to something that is really 1+1=3. We have done that with the new Bing.
Microsoft clearly thinks this is a big deal: not only did this announcement warrant a special event, but Satya Nadella labeled this announcement as heralding a new era akin to the PC/server and mobile/cloud eras, and was explicit that Microsoft has Google in their sights:
We think there are two things that are emerging: one are these conversational intelligent agents. I think these are things we’re going to have everywhere we go. All computer interaction is going to be mediated with an agent helping you. In fact, we’re going to have this notion of a co-pilot that’s going to be there across every application canvas, inside of an operating system shell, in a browser. And so we want to show you some of this innovation starting with how it’s going to reshape the largest software category on planet earth — which I’ve been working on for a long time — which we are very, very excited about: Search.
It’s a new day in search. It’s a new paradigm for search. Rapid innovation is going to come. In fact, a race starts today in terms of what you can expect, and we’re going to move. We’re going to move fast, and for us, every day we want to bring out new things and most importantly, we want to have a lot of fun innovating again in search, because it’s high time.
Nadella told Joanna Stern:
AI is going to completely change what people expect from search. We are grounded in the fact that Google dominates this space. We feel like a new race is starting with a completely new platform technology. I’m excited for the users to have choice finally, and a real competitive race out there…
The last time I checked Search was the most profitable category there is on Planet Earth, so all I need is a few more users, and someone else I am competing with has to keep all of their users and all of their gross margin. I’m looking forward to that…There is enough surplus that goes to one place, and I think it would be nice if it were evenly distributed. distributed
I’m going to reserve any further analysis of Microsoft’s announcement for now; I am looking forward to Google’s search event later today, which was proceeded by a blog post from CEO Sundar Pichai on Monday about Google’s plan to release Bard, a new conversational AI service. I don’t think it was a coincidence that the blog post and event bracketed Microsoft’s announcement; what would be nice is to have an actual product that people can use.
After Microsoft’s event I had the opportunity to talk with Microsoft CTO Kevin Scott and OpenAI founder and CEO Sam Altman. Given that I had only a short amount of access to two executives from two different companies because of a single product announcement, I decided to focus on better understanding the Microsoft-OpenAI relationship. It is, as far as I can remember, nearly unprecedented in tech: Microsoft is fronting all of the money and seems to be in line for nearly all of the upside, but only by virtue of betting their company’s future on a then-unproven startup with a very unconventional structure and ambitions. How did such an arrangement come about, what has it been like working together, and how has the relationship between the two companies evolved over the last several years? These are the questions I asked of Scott and Altman.
To listen to this interview as a podcast, click the link at the top of this email to add Stratechery to your podcast player.
On to the interview:
An Interview with Kevin Scott and Sam Altman About the Microsoft-OpenAI Partnership
This interview is lightly edited for clarity.
The Genesis of the Microsoft-OpenAI Relationship | Goals and Risks | Integrating Independent Companies | Costs and Business Models
The Genesis of the Microsoft-OpenAI Relationship
Kevin Scott and Sam Altman, welcome to Stratechery. We have a limited amount of time, so I will have to skip the usual biographical deep dive I like to do in these interviews, but I am interested in some historical context as to why I’m talking to the two of you. Sam, you’re obvious, so I’m going to start with Kevin.
After hinting that I had an interview scheduled for today, I got some speculation along the lines of, “Oh is Ben talking to Satya and Sam tomorrow?”. Instead, I’m talking to Kevin Scott, which I think is even better: from what I gather you’ve really been at the center of the relationship between Microsoft and OpenAI. What was the genesis of the partnership there?
Kevin Scott: Yes, to what you just said, we have been in the current instantiation of our partnership with OpenAI for about three and a half years now. I think very early in OpenAI’s history, they were on Azure briefly and then moved to Google for a bit and then after I joined Microsoft, I was looking after a bunch of the AI work that we were doing. Sam and I, funny enough, have known each other since he tried to recruit me to be the Head of Engineering at his startup, Loopt, a million years ago.
So this is really going full circle.
KS: Yeah. So Satya asked me to take point on having the conversation with OpenAI and seeing what we ought to be doing with them. I think that was in 2018, actually, we just ground through a bunch of things, decided that it would actually be a very good idea for Microsoft and OpenAI to partner, and we’ve been working super closely together ever since.
Sam Altman: Yeah. I agree with all that, but I think that’s underselling Kevin in this whole thing. I think the way it worked is that Satya and I had a few-minute conversation at a conference together in the summer of 2018 and said, “Hey, maybe we should figure something out more.” But then from then on it was, really, we’ve worked with Kevin from the very beginning and very in-depth, and Kevin is the one that got what we were doing. Kevin also is most of the reason of why we’ve wanted to partner with Microsoft from the very beginning. Now we like many, many people there and feel very aligned, but I think it’s mostly been Kevin and I driving things and it’s been great.
KS: Yeah, a ton of fun.
So tell me about this when you were thinking about this partnership. I mean, you have this connection with Kevin, but then just broadly speaking from a business perspective, was it just that Microsoft came to the table with more funding, or were there competitive concerns like in the long run you saw yourself competing with Google? What things went into your decision making, Sam, as you were deciding to commit in this direction?
SA: We trusted Kevin and Satya too, of course, but really I’m a big believer that you do a deal like this with people, not a company. And we had a lot of concerns. There are many reasons I don’t like many of the big tech companies, and there weren’t that many options for us that could have done the scale not only of the capital we need, but compute and hardware and just general muscle as well. We were delighted that of that shortlist of people one of them was Microsoft, and, again, specifically Kevin and Satya. I don’t think we would’ve done the deal had we not had the relationships and felt the very high degree of alignment.
How has the partnership changed over time? I mean, you just signed a new agreement. I’m sure you can’t disclose specific details, but I’m curious — you make an agreement up upfront, I’m sure there are a lot of disputes about exactly where the lines are drawn. Has there been a big shift in that as you’ve gone through this experience of working together or are things still mostly the way you thought they would be at the beginning?
KS: I think at the beginning the clear thing was that we had alignment, that we both really powerfully believed in this vision of powerful — what we call foundation models now — but these big models that we could use as platforms to develop lots of things on top of. We had a shared vision of what you needed to do at scale. We, Microsoft, believed that OpenAI was one of, if not the best — and I would argue actually it is the best AI team pound-for-pound on the planet — and that we could work together. They would help us do the highest ambition things with their infrastructure, which would be beneficial not just for OpenAI, but for Microsoft and all of the other customers who were using us to do AI things, and then we could help OpenAI commercialize some of the things that they were doing.
I think one of the things people miss is that OpenAI has been able to accomplish all of this amazing stuff, and it’s a very small group of people. The last thing I think you would want a world-class team like OpenAI to do is have to be distracted by building an enterprise salesforce, although obviously they can build whatever salesforce they want because we are independent companies.
I actually want to touch on both of those. But Sam, what’s your perspective on how things have changed?
SA: Look, I won’t pretend we never get annoyed at each other. It happens once in a while where one of us gets spun up about the other. I think that happens in any good relationship, but if you zoom out and look at the magnitude of what we have jointly accomplished and how unlikely it seems that you could do a partnership like this and have it work, I think it’s pretty remarkable — definitely the most successful partnership at this scale that I know of in tech.
I think we have extremely aligned vision and values about where we think this is going to go and what we want for the world, but extremely complementary skill sets and assets and things that we want. The things that OpenAI wants and the things that Microsoft wants are compatible and not very overlapping. That’s led to where I think we’re crushing the game together jointly on technological progress, soon and increasingly so on business progress, and in terms of pushing for what we want about responsible deployment of these very powerful systems and massive shared benefit to the world.
I think we’re really doing that and it’s a crazy thing: the level of dependence that companies have taken on each other is a little bit scary. It probably makes us both a little bit uncomfortable in different ways, and no contract in the world can protect you for that. So you just end up sizing each other up and saying, “All right, you know what? It’s been good these last three-and-a-half years. We trust each other. Let’s just go for it.”
Goals and Risks
That’s an excellent segue to the questions I wanted to get to. Sam, starting with you. The way I’ve thought about OpenAI’s willingness to partner with Microsoft, you mentioned this bit about compatible but not overlapping goals, and you’ve talked about this overarching goal of achieving artificial general intelligence. My perception is you’re happy to give away incremental gains, whether that be better search results or integration with word processors, or what have you, in exchange for the resources to make that happen. One, is that a fair characterization, and two, how exactly do you define AGI? How do you know if you’re actually achieving this goal?
SA: I don’t quite agree with the characterization of “giving away” because I think we’re getting far more value back here. We’re happy for Microsoft to be doing at-scale commercialization, but there are things that we want back like more computers, human data from using these products, things like that, and our partnership allows for that. Also Microsoft is very good about being extremely understanding of when we need to go off and commercialize something on our own for some reason.
But in the same way that we try to be very understanding about Microsoft’s needs, and if Microsoft needs a model that’s a little bit different than what we had in mind, we always try to make that happen. I think that’s called being in a good relationship. I don’t think it’s accurate to say we’re giving away all short term value, we’re just both optimizing for what we’re good at and what we care about. And as you said, we, OpenAI, definitely care a lot about AGI and that’s going well, but the reason I don’t quite like the characterization is because I know Microsoft cares about that a lot too. We just think we can jointly do it together and get it deployed together and get the public policy and engagement of all of society done correctly together.
So I don’t think it’s like, “Okay, Microsoft gets all short-term value and OpenAI gets AGI”, but it’s like we’re just going to do the whole ramp from here to there together. And again, we’ll piss each other off once in a while, but on the whole will be very good partners.
Kevin, on this point, particularly with today’s announcement, at this point the benefit to Microsoft of this partnership seems super clear, but what were the risks?
Something that stood out to me about today’s presentation is that you talked a lot about, “We established this AI principles thing in like 2017” and you said, “Oh, we have all this compliance and legal teams X, Y, Z.” But what about your actual AI efforts? That’s something that Microsoft used to talk about, and you mentioned when you came into the company, partnering with OpenAI was a function of reviewing Microsoft’s AI efforts. Is that an admission that they weren’t where they needed to be? You can certainly see now, it’s so clear that you’re committed to OpenAI, how would you even build something internally at this point? Are you completely dependent on a separate company for arguably your most important tech going forward?
KS: Yeah, it is a very good question. I think the way that we look at things is we still do a ton of AI research and a ton of AI development. We have a bunch of long-term things that we are working on together with OpenAI and long-term things that we’re working on that are not necessarily AGI things that are orthogonal to some of the stuff that OpenAI is working on.
But by-and-large, what we really have been able to do is I think efficiently allocate resources in a field where it really doesn’t yield much benefit to peanut butter scarce resources like these incredibly talented people, and very expensive compute resources, and large volumes of data that you need to train models across a huge number of potentially competing things. So we have stopped some things that we were doing internally and we have accelerated some things that we were doing internally and I think it’s been incredibly complementary what we’re doing at all.
In some cases, I would say that our AI research that is just Microsoft that’s happening inside of Microsoft research and inside of Teams is maybe more advanced now than ever from us being able to take a bunch of OpenAI technology as starting points to build on top of as well as just being inspired and collaborating by and with OpenAI.
Is Microsoft’s goal, is it perhaps AGI but you’re the computer of AGI and OpenAI is the model of it? Obviously you do that together and work together, but what’s the balance between making our productivity apps way better, which is Microsoft’s traditional arena, versus we’re going to be the world’s supercomputer when all these productivity apps become pointless because it’s just AI everywhere?
KS: I think the way that I would characterize it, and this is where Microsoft and OpenAI are very aligned, is we believe that what we’re building is a platform of powerful compute and powerful AI models that lots and lots of other people are going to build on top of. Some of those people are product teams at Microsoft, some of those are product teams at OpenAI, some of them are entrepreneurs that are starting companies right now that are already building successful businesses on top of all of this. So for us it’s about whether we want to be part of completing that whole platform picture alongside OpenAI so that this technology can emerge and scale, and be as useful as possible to as many people as possible over time.
I would argue that that’s really Microsoft’s heritage. We’ve always seen ourselves as a technology platform company. You go all the way back to the invention of the PC. It was, “What do you need to do in order to make personal computing more accessible to everybody on the planet”? And it started with operating systems and programming languages and development tools and then it became a little bit about productivity tools and eventually just became a whole bunch of other things. I think we’re in the very early stages right now of seeing this new platform emerge, and it’s just really exciting to be working on that with OpenAI.
Integrating Independent Companies
Sam, I want to double down on the bit you said earlier about how you’ve been surprised at how well these two companies have worked together, because that’s something that I’ve been thinking about as well. Generally speaking, when developing a new product, you want real deep integration that only happens within a single company, between the technology and the product — modularization and partnerships typically come later.
In this case though, there’s a division between technology, from my perspective anyway, that is built by OpenAI, and product which is built by Microsoft — see the new Bing, for example. How does this work? Is open-AI technology thrown over the wall? If not, how does that partnership work? Is this early access to stuff OpenAI would do anyway? You mentioned earlier that you’ll build some different types of models for Microsoft. I’m curious how that back-and-forth works and if that’s changed over time.
SA: On my side, when we first started, I thought it could have worked the way you described. I had this fantasy that maybe we were just going to train a model and throw it over the wall and Microsoft’s going to build this product and we were never going to think about it again.
But in practice, that turns out to be not how it works at all. That wouldn’t work for either of us. We need the close partnership in both directions. We need Microsoft’s input early for the model to behave, Microsoft needs our help on how to get the model deployed for a particular use. Sometimes we have early product experimentation at smaller scale that can help. We need the data back from Microsoft and we need input into the product so that we can iterate in a meaningful way. There are a lot of pieces to this and so the way it has worked is not what I would’ve thought, which is the clean, bright line delineation, but actually just very close collaboration at each step. Kevin, I don’t know if that seems accurate to you too.
KS: Yeah. I think that is actually how it’s worked over and over again. If you think about GitHub CoPilot, for instance, it started with this realization with GPT-3 that you had a natural language model that could write code, that it could translate a English language intention of something you wanted to accomplish with a piece of code and produce actual code. Then it was just a lot of work to figure out how to turn that into a product and it was work across a whole bunch of different teams. So multiple parts of OpenAI, multiple parts of GitHub, and multiple parts of Microsoft — and that one was especially complicated because it was three organizations that are working together, not just two.
I think, again, to say that everything was smooth 100% of the time would be a mischaracterization and actually just sort of a ridiculous fantasy for how things work, but we were just determined to take this really interesting capability of a model and get it into the hands of people who would find it incredibly useful and I think that’s in the end fulfilling for everyone.
Did that lead to meaningful changes in how you work together? Because you obviously did CoPilot first and now you’re doing Bing. I’m particularly interested from your perspective, Sam, if you had to rethink how you allocate resources: “Oh, we don’t have to build an enterprise salesforce, but we do need to have a meaningful component of our resources focused on integration with Microsoft”?
SA: I’m sort of a big believer that you let these things evolve fairly gradually and extremely organically. I think it’s a total fantasy to think we could have designed all of this upfront correctly the first time in a vacuum. But the thing that we have done as just a rapid iterative loop, I’m quite happy with.
KS: Sam, you should say whether you agree with this, but I think CoPilot was the first major AI-most products, in that the product couldn’t exist without the really advanced AI. It’s the first thing that we did together and we learned a lot from the process and then we just went through this experience and the experience is not over. We will continue to work together on what we just launched with Bing. I think even though that was a bigger undertaking, it was a smoother collaboration than the first because of all of the things that we learned from doing it the first time. It’s like a thing where you practice and you get better.
SA: I would say that with these things, you never get the org design right. You never get the partnership right. You never get the agreement right. It’s either you trust each other and you like each other and you work together in good faith and eventually you work it all out or you don’t and you never work it all out and none of the rest of it can save you.
That’s why I was pleased to get the two of you together, because I think that that’s making a really good point.
Costs and Business Models
Just a couple of practical questions. Sam, you mentioned in a tweet that ChatGPT is extremely expensive on the order of pennies per query, which is an astronomical cost in tech.
SA: Per conversation, not per query.
Oh, okay, that’s a good clarification. Are there similar cost concerns about the Bing? One of the interesting queries, and this is maybe a question more for you Kevin, but I did a really interesting query this morning where I asked Bing to generate a bit of code. The first answer used an old version of an API I told it to access. So I told it to use the new version, and what’s interesting is it did it, but it took a really long time, several minutes in fact. It felt like it was going out there, finding the new API, parsing it in real time, and figuring out the answer. That’s both very cool and it also sounds extremely expensive. How is this going to be managed as you expand access?
KS: One of the things that we’ve gotten a lot of confidence in over the past handful of years is our ability to performance-optimize all of this stuff, both on the training and the inference side of things. Already the cost envelope for Bing looks very favorable to us. We think we actually have a lot of flexibility in bringing it to market as an ad-supported product that we wouldn’t have had if we hadn’t spent so much money on performance optimization and the cost is just going to go down over time.
SA: And I think this is a great example of where the partnership works together, which is like, we all want this to be much cheaper at inference time for various reasons. Can we share resources, share ideas, can we talk together? Can we do this part? You do that part. Can we get really good at driving this cost down? We didn’t expect ChatGPT to be such a success, so we had not gotten as disciplined about optimization at that point as we needed to. But since then we really have, and it’s been a great example of the partnership really working.
Sam, will some of these new Bing features come to ChatGPT, and will any of them be open source?
SA: We will continue to open source stuff. We won’t open source Microsoft’s — their stuff is their stuff — but we will continue the open source stuff and yeah, we’re going to collaborate a lot on features. I don’t think everything will make sense for ChatGPT, but we want to continue to improve that product.
KS: Yeah, for sure.
Kevin, I actually want to delve in real quick, you said advertising will make sense, which is really interesting. The new Bing does have a combined interface with ads, but what was striking to me about watching the demos is that everyone, almost every demo, goes to the chat interface right away because it just feels more compelling, it’s not stuff smooshed together. However, that really kills the user-selects-the-winner-of-an-auction dynamic that drives search monetization. Satya in another interview said something along the lines of, “Well, maybe there will be other business models.” Is this going to be a subscription? Is it going to turn out that these models are not necessarily making search more profitable, they’re actually value-destructive, and that’s actually fine for you because you have 4% share, and maybe not so fine for someone that may be more dominant in this space?
SA: Can I say something because Kevin will be too polite?
SA: I think it’s fabulous for both of us. I think there’s so much upside for both of us here. We’re going to discover what these new models can do, but if I were sitting on a lethargic search monopoly and had to think about a world where there was going to be a real challenge to the way that monetization of this works and new ad units, and maybe even a temporary downward pressure, I would not feel great about that.
Well, you could always write a blog post and put together a quick press conference in response.
SA: Or both of those things! You can do both, it turns out.
From Microsoft’s perspective, is this going to be a funnel into new products or do you see it as an end goal in and of itself, winning search?
KS: So I think you hit on a very important point which is even if the ad economics of this system doesn’t have the same economics that “normal search” has, if we gain share, it’s just great for Microsoft. I think we have a lot of ability here, partially because we’ve done so much performance optimization work and we’re really confident around costs, that we can figure out what the business model is. The thing that I know having been a pre-IPO employee at Google is the search business that you have now is very different from the search business that we had twenty years ago, and so I really think we’re going to figure out what the ad units are, we will figure out what the business model is, and we have plenty of ability to do all of that profitably at Microsoft.
SA: There’s so much value here, it’s inconceivable to me that we can’t figure out how to ring the cash register on it.
Right, and you guys have the luxury of figuring it out because Microsoft, you have plenty of other great business models in play, and OpenAI, you have funding from Microsoft.
SA: What do you think of the product, Ben?
It’s good. Is it GPT-4 or is that not being disclosed?
SA: I think the model numbers thing is a dumb framework anyway. The thing people thought, there’s been many versions of GPT-3 so it’s a better model, we need to figure out our naming at some point.
I think that one thing that was really interesting to me, just as I’m using it, is before it generates the answer, it gives some prompts about, “Oh, this is where we’re searching”, and I couldn’t decide if I liked that or disliked it. On one sense, it sort of broke the magic because it’s like I’m not talking to an AI. On the other hand, I can get the sense that it conveys some confidence about what’s going on.
I was very impressed at the recency, how it captures stuff. For example, I asked it, “Does Ben Thompson think there’s a recession?” and it actually parsed my Article on Monday and said, “No, he just thinks tech’s actually being divorced from the broader economy,” and listed a number of reasons. On the other hand, you have some very weird Ben Thompson results so I got some other weird stuff, but I was doing narcissistic searches.
What I think what I’m really interested in is how this interface plays out. This bit about you are launching it as a sidebar to search but every demo immediately went to chat was very striking to me, and ChatGPT was obviously such a breakthrough just because it resonated, the interaction model did. This makes me worried for a lethargic search company as you put it, because it’s just hard to see the compatibility with the business model.
KS: (laughing) I don’t think that’s what we said.
No, that’s what Sam said! Does that old format even make sense with this? Or does it have to be new interaction models for it to work?
SA: You know what? The answer is that I don’t know. No one else does know, but competition in search is going to be a big win for actual users. I think we’re going to see a ton of innovation this year and I’m thrilled about that.
Can I make a closing comment? I didn’t realize this was going to be so much about the Microsoft-OpenAI partnership, but I’m happy that it was and it made me think of something. I think 75-80% of the success of the partnership goes to Kevin personally, and it’s a rare thing for any big company executive to actually go say, “I’m going to do this thing,” which was probably not a very popular or risk-free decision for Kevin to make at the time of the initial investment, and then stick so steadfastly behind it and put up with our antics and whatever else. So very thankful to that and I think 75-80% of the credit due there and it’s a rare thing.
KS: I think Sam’s being too generous.
Well, the things I’ve heard, it affirms Sam’s statement. A lot of people are going to be talking about the product itself, but to me, this bit about two different organizations, with Microsoft basically taking such a dependency on OpenAI and OpenAI in some respects being very generous with what Microsoft gets out of that to me is very compelling, and I agree with you, Sam, it’s something I don’t think we’ve ever seen in tech before, and that’s why I wanted to dive deep on that, because I think it’s very striking.
KS: I have been very humbled and grateful for all of the work that we’ve been able to do together, and it really is a tough act of coordination because it’s required a ton of people at OpenAI and a ton of people from Microsoft to change, sometimes significantly, the things that they were doing or how they were thinking about the world in order to accomplish these shared goals that we have, and I think a lot of people deserve a ton of credit for the flexibility and getting on board a thing that, to your point, looks a little bit unusual in the history of tech, but so far, is working really, really well.
This Daily Update Interview is also available as a podcast. To receive it in your podcast player, visit Stratechery.
The Daily Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.
Thanks for being a supporter, and have a great day!