Steam Halts Web3 Games; FDA Approves Prostate Cancer AI

Welcome to 16 Minutes, our show where we talk about tech trends in the news. I'm Zoran. We have two segments today, including news of a first-time FDA approval of AI software helping detect prostate cancer in biopsy images. But first, we cover the announcement recently that Valve Software, which operates the massive gaming platform Steam, added a rule prohibiting games that use blockchain technologies or that allow users to exchange cryptocurrencies or NFTs. This rule appeared on its What You Shouldn't Publish on Steam onboarding list for developers. This sparked headlines contrasting Valve's move to, quote, ban blockchain games NFTs, while gaming competitor Epic claimed, per its CEO, that it would, quote, welcome them. But we go beyond the players to the trends at play here, putting the news in context, as is the premise of the show, because it not only immediately impacts gaming developers and gamers using the platform, but has implications for gaming business models and the arc of innovation in gaming as part of the Web3 movement. We've talked a lot about the topics of crypto, gaming, and more, which you can find at a16z.com/gaming. And you can also check out a16z.com/nfts to learn more there specifically, as we don't cover that in this episode. Our expert guests are A16Z partner Jonathan Lai on the consumer team, who helps lead gaming and has joined us before to talk about Steam and other trends. He previously led the North American games investments team at Tencent. Also joining the segment is A16Z investing partner on the crypto team, Eddy Lazarin. We begin with Jon on the significance of Steam and Eddy on the Web3 movement before going into the specifics and the implications.

So Steam is a digital distribution platform. It is a desktop client many folks that they download onto their computer and then use it to play your games. But at the same time, you can also click into the store, which then shows you a catalog of games from all over the world that you can purchase. It has over 120 million monthly active users, has really popular games, Counter-Strike, Team Fortress, Dota. And as a game developer, it's great because instead of going through a publisher to distribute your game, you can just go on to Steam, create a developer account, and upload an executable file for your game and then with just a couple of button pushes distribute your game to the entire world. Think of it as the App Store for PC games.

Okay, so Steam is a big player in gaming, but why is it in the news now? What does it actually mean?

This is a huge signal for the market because Steam has generally had a positive reputation in the game developer community as one of the most progressive slash developer-friendly companies. They were one of the first to open source a lot of the software that they built around Steam. They also permitted peer-to-peer trading in a number of their most popular games. And so it's especially significant that one of the most progressive game platforms is declaring that they're banning NFTs and blockchain games outright. It's a cause for concern for many Web3 game developers who are still figuring out where they can distribute their games.

Let's talk about some of the bigger trends at play. What's really happening with Web3 and gaming based on Steam's decision here?

The way I think about Steam as a gamer is that Steam allowed us to disintermediate one important agent in the distribution of games. And that was the game store. Games used to be distributed in kind of a consumer packaged goods style model where you go to a store and you actually pick up the game disc or game cartridge. And Steam was one of the leaders in removing that part of the distribution cycle. So Steam allowed the publisher to distribute the game directly to the gamer's device. And that felt like cutting out so much waste in the middle. Now Steam stands as an intermediary between the game publisher and the players because they control the in-game economy. Now they're even controlling the kinds of technology that you can use. It forces game developers and publishers to rethink what distribution channels they're going to use. And I think it encourages them to own those distribution channels directly, like we've seen in other kinds of media.

So we hear about NFTs in a gaming context. We hear about blockchain-based games and Web3 games. So what's going on here? What is Web3 gaming and how is it actually different than what we've been accustomed to?

Well, we're in the early days of Web3 games, but it's financialization and composability. Financialization is the one that we've seen more. In the most basic sense, that's like paying for games or game-related items. Using a blockchain or using cryptocurrencies as the medium, but much more interestingly, Web3 allows players to own the items in those games, like really deeply own them in a way that even the game developer can't necessarily seize the asset. It also opens up the opportunity for gamers to trade those items in totally purpose-built, very legitimate, and very difficult to circumvent or subvert secondary markets. And using a blockchain in any of these games allows the capital or assets or money within the game economy to exit Steam. A concern that I have is that Steam may be throwing the baby out with the bathwater because they're limiting the interesting kinds of non-financialized applications that Web3 enables, specifically interesting kinds of composability, the ways that blockchains and Web3 allow developers to remix and combine different elements from games. Like for example, items or land or skins, all the different kinds of entities in these games, allowing other people to build on them or extend them kind of like modder culture, things like that. Web3 offers some really, really interesting methods, really high quality 10x improvements to this modding, extending, remixing, recombining process. And by limiting that kind of game, any game built using blockchain technology, Steam is also limiting that kind of remixing and recombining and extension of modder culture that will emerge in Web3.

So let's separate the hype from the real, John. What's the actual impact here with Steam's decision for Web3 game developers and their users?

Yeah, there definitely have been projects on Steam that employ NFTs or have a portion of their gameplay on-chain, they're going to have to remove their games and then figure out ways to reestablish that community elsewhere, which I think is always challenging. I think it's not an insurmountable challenge in the sense that most Web3 games have actually done a fantastic job of building communities directly. And so if you look at most of the largest projects, whether that's Zed Run or Axie Infinity, they have massive Discord servers where you have hundreds of thousands of people that are talking and chatting about the game. And so in a sense, they sort of own their own distribution directly to the players.

Given all this, you would think that Steam would want to find a way to coexist with this new style of game development and gameplay, and yet they seem to be ceding all that ground. So what are some of the other reasons they might be doing this?

So I think there are two potential views as to why Steam is taking this particular stance. I think there's a charitable one and an uncharitable one. The charitable view is that Steam is just trying to avoid a lot of the financial reporting and the legal requirements that might come with the trading of crypto assets. I think the uncharitable view is that fundamentally they want to maintain control over their marketplace. And that includes everything from content, the types of games that are on it, pricing, how much those games are sold for, down to the take rate, which is the tax they extract in return for providing the platform that these games are distributed on. But it's hard to control those things if large swaths of games are Web3 and their economic activity The trading of NFTs, for example, primarily happens off-platform.

Bottom line, what's the main takeaway for you here?

The bottom line for me is that Web2 companies, including game companies, have to rethink how Web3 affects their fundamental business models. And that includes taxing distribution, which is the main revenue source for these kinds of application platforms.

Web3 is fundamentally disruptive to app stores like Steam because Decentralized distribution removes power from gatekeepers. Steam blocking NFTs and blockchain components from its platform, I think, is a defensive move for them. And so the bottom line for me is that if you're a developer building a blockchain-based game or incorporating NFTs into your game, you just unfortunately lost one of the largest digital distribution platforms for PC games. Over time, I would expect to see new sort of Web3 native distribution platforms arise to fill the gap caused by Steam, Apple, Google all stepping back here.

Eddie, John, thanks so much for being with us today.

Thank you.

It's a pleasure.

In our next segment, we talk about the FDA's announcement last month that it authorized marketing of the, quote, first artificial intelligence-based software designed to identify an area of interest on the prostate biopsy image with the highest likelihood of harboring cancer so that it can be reviewed further by the pathologist if the area of concern has not been identified on initial review. The FDA reviewed the technology from Paige Prostate through its de novo regulatory pathway designed for low to moderate risk new types of devices. In the company's clinical study, the software improved pathologists' ability to detect cancer on individual slide images from 90% to 97%. So to discuss this news, we have 3 expert guests. Ellie Van Allen is a computational biologist, medical oncologist, and associate professor of medicine at Harvard Medical School, as well as chief of the Division of Population Sciences at the Dana-Farber Cancer Institute. A16Z Biogeneral partner Vanita Agarwala, who also sees patients at Stanford as an adjunct clinical professor in the Division of Primary Care and Population Health and is a former director of product management for Flatiron Health, and A16Z Bio partner Jay Rougani, who was also previously at Flatiron Health, where they helped build data and analytics software products used to study cancer care. We begin by discussing the impact of this FDA approval for other AI technologies. The first voice you'll hear is Elise.

I think perhaps the most exciting part about this particular approval is that it is actually now creating a mechanism for these kinds of approvals going forward. And that if the FDA is thinking about this kind of technology in this kind of a way, it opens the door for lots of other decision support technologies that use advances in machine learning to inform and improve healthcare delivery. And that's actually huge. You can maybe help with this decision support tool find a little bit more prostate cancer in an image.

So let's separate the hype from the real. Is there something else we should think about when it comes to this approval?

You know, one could very easily argue the opposite and say, well, we already overdiagnosed a lot of prostate cancer to begin with, and we might be finding more pixels that are consistent with prostate cancer. However, is this actually helping us diagnose more indolent disease or disease that's not clinically relevant, or is this actually helping us diagnose disease that really is mission critical that we intervene on immediately? And it's complicated to actually tease that apart. I think there's a lot of information that we still have to learn about this specific use case.

The only thing I'd say is that the top-level numbers that you tend to read in headlines, like percentage improvements in sensitivity and specificity and detection rates can be a little misleading because if you actually look at some of the published data, and this is not even specific to PAGE, the vast, vast majority of slides analyzed are actually concordant, you know, between a prior diagnosis or a pathologist's initial diagnosis and the algorithm. And so then it does become a challenge to kind of ask the question, sort of, how much efficiency and time gain did you really squeeze out? And a lot of the action is really in a minority of cases where there are discrepancies and then The discrepancies, some of them matter, some, to Ellie's point, clinically may not matter. And so we're in somewhat uncharted territory. If there's an ML-derived feature that maybe correlates with past human characterization of cancer, but 3 pathologists still don't see it, we do have to kind of ask clinically what that means, and we don't really know.

And how do we work toward answers to those questions?

So this was a retrospective study, and which is still admirable and it's good. I think the way that gets addressed ultimately is a prospective clinical trial that's randomized and representative and likely with large patient numbers. The question in a prospective trial would be what's the right endpoint? So maybe what they really want to assess is, is that does it help with pathologist efficiency? And is that the right outcome metric or is the right outcome metric Did application of this decision support tool improve outcomes for men with prostate cancer? And there's a wide range of different kinds of outcomes one can aspire to obtain somewhere in between those two ends of the spectrum, and those are the kinds of things we tend to look for when we're thinking about implementation of new devices, new technology into medical care to really sort of prove that it's actually not just better in a retrospective way, looking at old data in a specific task, but rather prospectively going forward, we can show that it's even better and it really matters in a way that's meaningful to clinical care.

Help us understand where this fits on the long arc of innovation. What have we seen in recent years with AI and medical devices, and is this approval an inflection point of sorts?

You know, there was an analysis in Nature by Stan Benjamin et al. In 2020. They looked at the number of AI and ML-enhanced algorithms, medical devices, and diagnostics that were approved by the FDA to date, and they found, I think, less than 100 approved. But more interestingly, there were only a handful approved through this new de novo pathway clearance, the same pathway that the PAGE prostate algorithm was approved through. So this is the de novo pre-market review pathway, which is a regulatory pathway for low to moderate-risk devices of a new type. And so what's interesting here is that there's a pathway that is emerging for these types of AI-enabled diagnostics. From a performance standpoint, the question is, how do we think about the upside of an AI-enabled diagnostic algorithm that is providing additional data to the human decision maker, the clinician, relative to the downside of a result that is potentially incorrect? Effectively, what is the risk of a false positive and a false negative? And I'd be especially curious on how should we think about the performance of computationally enhanced diagnostic methods that are going to be adjunct to the human pathologist decision but not the primary driver of decision.

So I certainly think that the sort of adjunct style is a softer way to do it, and it's certainly the most reasonable way to sort of gradually dip one's toes into the water of this new space and ease people into this new world. And I actually think that's reasonable. My concerns are actually more upstream at the moment of that, of actually still the more fundamental technology itself and sort of what the issues that might arise when it starts to be applied en masse that we can't actually reliably know in advance. I try to think more practically about sort of Where is this technology perhaps most useful? It's most useful in the context of overburdened and overwhelmed healthcare systems and low-resource environments that maybe don't have the technical expertise to get 2 or 3 or 4 pathologists' opinions. That's the healthcare gap you're trying to fix. I then go back to, well, how are these algorithms developed? What data were used to represent them? What patient populations were used to train these models, are those representative of the patients you're trying to actually fill the gap for? And in the other end of the spectrum, is the technology you've developed accessible in an equitable way for all of those sort of resource environments that you're actually trying to help?

I think, I think none of us are actually calling into question the approval or the underlying data, but it is notable that it's not actually public. Ali's raising good questions that we just don't know exactly which of these evaluation questions were definitively answered versus answered as best as they could be with limited data.

So this has been approved for marketing. What does that actually mean? Does that mean that it can start making its way into oncologists' offices, or what are the next steps here?

So actually, I think this is one of the most interesting questions, is the meta question of do you even need FDA approval? There are lots of places where clinical decision support is not formally FDA approved in real life. There is a very gray spectrum when, as long as a physician and provider, you know, who's certified is in the loop, the whole field of clinical decision support, it's not clear what parts of it need FDA approval versus not. So PAGE actually has kind of set a precedent now in the market regardless of whether it was really officially required to pursue FDA approval to augment pathologist decision-making. But they have now set a precedent. And so this is a stamp that could very well make it easier for providers, health systems, pathology labs to adopt the technology. I do think if you just put yourself in the shoes of a patient and you ask, hey, if my pathologist could have a tool that provided some oversight or some degree of protection from the possibility that I'm the 80th slide they reviewed on a busy Wednesday? Would you want that knowing that the pathologist is still making the final call? I have to say my intuition is that most patients would say yes. You know, I may not want a robot instead of my pathologist, but do I want computational kind of oversight where discordance could still be adjudicated, I think the answer is yes for a lot of patients. So I do think that's kind of another piece to keep in mind here. We're sitting against a backdrop of patients who have variable access to high-quality care with respect to, you know, very, very specialized academic centers versus other places.

I agree completely, right? You wouldn't want a doctor to have a hand tied behind their back if there's an actual technology to help. This is just in some sense another tool in the toolbox. Where I get concerned is, you know, some of the best models we've trained doing this exact kind of work are to predict what hospital the image came from, or things like predicting the body mass index of the patient from the image, and other sort of like features that are not things you think you're trying to predict but are confounding it and are basically a function of the upstream data. And I keep coming back to that point because the long-term technology play is super exciting. The short to medium term implementation is something that we got to think about really carefully.

Okay, let's get to the bottom line. What's the main takeaway here?

There are very reasonable concerns about data absenteeism and equity and what technologies like this may do to the healthcare system when applied in larger patient populations across more diverse clinical settings. And as with any technology, it opens up a lot of new things or exacerbates a lot of pre-existing things that need to be addressed in the system that open up more doors for innovation and opportunity.

So I think this in many ways provides some framework to entrepreneurs out there who are looking to understand the level of and the type of evidence that the FDA is looking for, for these kinds of computationally oriented diagnostic tools and get them in the hands of clinicians that they're trying to support.

I think at the end of the day, human decision-making and human compassion as a physician is an extraordinary thing we have. And I think this follows the right framework of enabling, empowering, and increasing the productivity and capacity of that unique physician workforce. I respect Paige for going through the FDA process and setting a precedent that's actually good for the clinical ecosystem. And I really look forward to learning more about the data and getting more insight into how the real-world application of algorithms like this play out.

Ellie, Jay, Vanita, thank you so much for being with us today.

Thank you.

Thanks for inviting me.