Black Forest Labs has established itself as a pioneer in visual intelligence, with its open-weight FLUX models reaching over 50 million downloads on Hugging Face and rivaling models from Google, OpenAI, and DeepSeek in developer adoption. The company has distinguished itself not only through technical capability, but through a strong commitment to open research. Black Forest Labs recently published comprehensive safety evaluations showing its latest FLUX.2 model family demonstrated more than 10 times fewer vulnerabilities for serious risks than other leading open-weight image models, and its post-training mitigations reduced vulnerabilities by up to 98%.
In this conversation, Black Forest Labs’ Adam Chen and Ben Brooks, who lead the company’s legal and policy work, join Matt Perault to discuss what it means to build frontier visual AI openly. They explain the role of open models in advancing transparency, driving down the cost of innovation for developers, and strengthening security and sovereignty by reducing the world’s reliance on a handful of closed APIs. They also outline the unique policy challenges facing open-weight model developers.
For policymakers, their message is clear: supporting open innovation does not require abandoning oversight. It requires targeted rules, analysis of where harms arise, and a better understanding of how proposed regulations land on smaller frontier labs, not just the largest incumbents.
The conversation also offers a window into what it looks like to build a policy function at a startup. With a legal team of four and an even smaller policy operation, Black Forest Labs is navigating the same regulatory landscape as firms with thousands of employees on their legal and policy teams. Adam and Ben offer a candid view into how they enable their small team to have outsized impact, rather than trying to match Big Tech’s playbook.
Topics covered:
00:00: Intro
01:09: What is Black Forest Labs?
02:25: The makeup of a legal team at a frontier AI startup
06:26: The role of visual intelligence in the AI ecosystem
09:01: Core risks and baseline safeguards for visual models
09:46: Unique policy challenges of open-weight models
11:50: Restricting access to general-purpose technology should be a last resort
15:04: What’s at stake: open models as soft power and the China dynamic
19:19: BFL’s approach to being open and responsible
21:38: BFL’s model testing results
24:11: How a four-person legal team approaches disclosure and compliance
27:02: What works and what doesn’t in transparency proposals
30:19: Navigating the state, federal, and international patchwork as a startup
32:46: BFL’s advocacy goals
36:25: The Little Tech voice as a competitive advantage in the policy ecosystem
This transcript has been edited lightly for readability.
Ben (00:04)
I think a lot of developers and a lot of policymakers think you’re either open or you’re responsible, you can’t be both at the same time. And what this data, what these evaluations help to show is that you can be both at the same time.
It’s really important for policymakers to understand that open models are a vehicle for projecting soft power abroad. And the values and the choices and the vulnerabilities encoded and embedded in these models are going to determine behavior of all sorts of real world systems for the next few decades.
Adam Chen (00:34)
We believe that a voice that can articulate a Little Tech point of view is important to ensuring that policymakers have a more robust understanding of the entire market and what regulations they may be contemplating, and what impact that could have on innovation and competition.
Ben (00:47)
Restricting access to general purpose technology should be a measure of last resort, not a measure of first resort.
Matt (01:00)
Welcome to the a16z AI Policy Brief.
Adam Chen (01:03)
Thank you, it’s great to be here.
Matt (01:05)
Adam, can you tell us a little bit about Black Forest Labs? What is the company and what are you guys doing?
Adam Chen (01:09)
Yeah, it’s my pleasure to do so. So BFL is a frontier research lab building the standard for visual intelligence. It was founded in Freiburg, Germany, so the heart of the Black Forest, hence the name Black Forest Labs. And really the founders were in university studying visual intelligence and at a certain period in time decided to kind of continue their work by building their own startup from Germany.
And so our mission is to build foundational technology in visual intelligence. And so what that will mean is essentially a high performance, high quality, lean, open and fast visual intelligence models. We do so, and we want to do this by building openly, by investing heavily in open research. And we ultimately believe that this will result in better gains for the company and for society as a whole.
As terms of myself, I am working the legal team here at Black Forest Labs, heading up the legal team. I joined about a year ago and I’m very excited for all of the work that the team has been building towards.
Matt (02:18)
Adam, you are referencing the legal team at Black Forest. I think sometimes when people hear legal team, they’re thinking of dozens or hundreds or thousands of lawyers. Can you give us a little bit of a sense of like, what does the legal team look like at Black Forest Labs?
Adam Chen (02:25)
Very small. So in terms of people who are acting as lawyers, it’s about four people. And that’s two on the commercial side and myself and another lawyer on the product counseling side. And so it’s very minimal, I think, in terms of the scale as compared to some of our competitors out there.
Matt (02:45)
And you’ve recently grown the function to include policy as well. Ben, you’ve recently joined in a full-time role. So Ben, are you doing at Black Forest?
Ben (02:57)
Yeah, I lead public policy and a lot of what we’re calling model assurance works. In other words, how do we validate the claims that we’re making about the risk in our models?
I’ve known the team for quite a while now in one form or another. I first reached out to the team when they were working on stable diffusion back in the day. We had a congresswoman who was writing to the White House saying we need export controls to stop the release of this technology. And that’s how I got involved.
And then most recently joined to lead our public policy work. Before that was a fellow at the Berkman Klein Center at Harvard. And before that, a whole lot of regulatory advocacy and some really spicy high stakes, usually commission based domains, right? From drones through to crypto, ride sharing and a whole bunch of stuff in between.
So for me, this has always been an exercise in how can we regulate in the public interest, but do so in a way that is maximally compatible with open innovation? And this is more important in AI than ever before. And this team has been at the forefront of showing the world that there is an open and capable future out there. It doesn’t just have to be closed source API. So it’s really exciting work for them for some time. And as Adam says, it’s a lean agile team but we’re punching above our weight in all ways that matter.
Matt (04:20)
So you guys have both done a range of different things in tech policy and law. Ben, maybe starting with you and then Adam, I’d love to hear your thoughts. Like why BFL? What is compelling to you about this opportunity versus Ben’s work in academia or other companies large and small that you might be at?
Ben (04:35)
Yeah, I mean, think it’s a combination of things. From the perspective of our team, having a team that is comprised of missionaries, not mercenaries, right, they deeply care about research. They care about moving the frontier in all the ways that matter to real world developers and deployers And they care about releasing this technology and this research openly. And to me, that was a huge draw card. And it sets them apart from a lot of the other players in this space.
I think there are also some interesting, fairly unique challenges that a team like BFL faces. We’re doing frontier work, which means that we have some of the most capable models out there. We’re doing it openly and open models have unique challenges and unique properties that we can maybe discuss today. And we’re a small team, right? So we’re up against the Googles of the world, but we don’t necessarily have the same ability to absorb risk.
And we need to be very careful about how we go about this work. We need to show the world that we develop this stuff thoughtfully And so, you know, a really compelling mission, a really important mission, a great team coupled with some very unique headwinds, I think makes this a really attractive place to continue to continue that mission with government and policymakers.
Adam Chen (05:49)
On my end, the team is exceptional here. Everybody is really, really talented at what they do, and they’re great people. And that was a really great draw for me. And I think, honestly, one of the most exciting things to do in law and policy right now is working on the frontier. And right now, AI is the frontier of both law, policy and also technology. And so this team is looking to push that frontier to expand the capabilities of what AI models can do. And that makes professionally a very exciting challenge.
Matt (06:26)
So I want to get into the specifics of the law and policy issues, but to get there, I think we need to understand a little bit more about the technology itself. So when you talk about building visual capability at the frontier, what does that actually mean in practice?
Ben (06:37)
Well, with visual intelligence, think about language models, right? Language is a really abstract and highly compressed representation of the world. And so we have models out there, fantastic models that can learn to manipulate text for a whole range of applications, right? Coding, search, everything in between. But their understanding of the world is fundamentally limited in important ways. And our team’s belief has always been that pixels encode a much richer representation of the world. If you think about us as people, vision is primarily how we perceive and understand and interact with the world around us. And so, you know, the bet that we’re making as a team is that visual AI is going to transform the world, not just in these very creative, highly expressive ways that we see today, but in more functional ways as well.
If you think about today’s sort of current crop of visual models, we have relatively narrow capabilities, but we still have some high impact applications, film, gaming, design, concepts in architecture. But where we’re moving towards is more of a unified visual intelligence architecture, where you have versatile models that can bring together perception of the visual world, simulation, visual reasoning across different modalities and do so in one scalable architecture. So in the short term, that means better simulation environments for agents, embodied systems that are using models to learn how to navigate complex environments. That’s really exciting.
I think long-term, this means potentially embedding these models in high-stakes systems themselves, right? The model isn’t just predicting or generating the next frame in a video sequence, but it’s actually outputting an action or a recommendation. And this is fundamentally how we’re going to unlock specialized robotics, web agents with computer use, embodied systems that can manipulate physical tools in really consequential environments. So for us, it’s not push a button, get an image which is where the technology is historically associated today. But long term, there are some really impactful, important functional applications of this stuff. And our team, as I say, is pushing the frontier there.
Matt (08:43)
Adam, when Ben was talking in his opening just about the nature of the work and he was talking about visual capability being compelling from a technology perspective, he was also saying it raises unique challenges in the regulatory context. What are those challenges and how are you guys responding to them?
Adam Chen (09:01)
I think at core base, there is some visual content that is just outright illegal in many jurisdictions, right? And so you’ll see that with CSAM, child sexual exploitation, abuse material, and non-consensual intimate imagery, NCII. And so one of the things we really try to do is to tackle those types of core risks that exist right now with visual models at a really baseline model layer and not to just kind of try to paper over it with some sort of like inference moderation layer itself.
Matt (09:35)
And the other policy issue that you were flagging in your opening was how open you’ve decided to be. So what is the feedback that you hear from the policy community about your decisions around openness?
Ben (09:46)
I think open models, as opposed to closed models, pose some unique challenges. You can modify them for purposes that were not anticipated by the original developer. You can run inference on them independently without moderation layer filters. And if there’s something wrong with the model, if there’s a vulnerability, it can be very difficult to withdraw that model from circulation once it’s out there on the internet on platforms like Hugging Face or GitHub. And so we’ve always been very sensitive to the challenges.
I think one of the difficulties is that when policymakers look at AI, they expect us to mitigate a lot of these risks in the model layer. And we’re everything from political deepfakes, sexual deepfakes, automated decision making, negligence, defective product design, the copyright issues. All of this is expected to be done at the model layer. And the challenge is that many of these risks are highly context dependent.
And it’s not necessarily possible to mitigate that in the model. The models are versatile. You can do many things with them. Most of it good, some of it bad. Those capabilities are often entangled in ways that can be difficult to tease apart. And then, if you embed a model with particular safeguards, you can, with enough data, compute and expertise, start to unwind some of those safeguards. So I think that mismatch is one of the biggest challenges we face with policymakers. The expectation that models are the silver bullet and we’re going to fix all of these challenges in the model. And the reality that actually the model is just one component in a complex integrated system, these risks are context dependent. And while there are many important safeguards we can put in place in the model, that isn’t the end of the story. And from a regulatory perspective, it’s really important that we capture some of that nuance in these future legislative and regulatory reforms.
Matt (11:37)
I assume when you raise that point with lawmakers, they say, okay, we understand that it’s difficult, but are you just expecting us to live with a world where all these harms are created? So what is your vision for what the right approach is from a policy perspective if you’re concerned about the harms, but you recognize some of the limitations of addressing that at the model layer?
Ben (11:29)
I think it’s a few things. I think one is just being very clear that restricting access to general purpose technology should be a measure of last resort, not a measure of first resort. And too often, I think policymakers are looking at those kinds of interventions as the first thing we should do and not the last thing we should do. You think about just the past couple of years, the number of bills we’ve seen around licensing model developers, around export controls for model weights, some really interesting, fairly exotic liability proposals that make life very difficult for open source researchers and developers, those kinds of reforms, we need to be very clear, have a huge impact on open innovation and they should be interventions of last resort. I think the second piece though is like, I would love to see governments run a systematic gap analysis of where our existing regulatory systems fall short. And so often we’re kind of legislating before we do that gap analysis.
And the truth is there are gaps. It was only last year that the US government introduced federal criminal liability for non-consensual intimate imagery. NCII is a huge proportion of the risks that policymakers and real families out there in the real world care about. But it was only last year that we plugged that gap at a federal level. I think the same is true on a range of other different AI systems too.
So we’d love to see more of a systematic gap analysis there, figure out across the supply chain from developers through to users and everyone in between. Where do our existing systems fall short and where can we introduce some targeted reforms to up level the confidence in the AI supply chain? So a lot of different things. can, can double click on any of them, but generally speaking, it’s really important that we make sure that whatever we do for whatever risk, it’s compatible with the culture of experimentation and open innovation that brought us here in the first place and is going to help to make AI useful and accessible in the future.
Adam Chen (13:38)
It’s also really critical to also remind ourselves why opening innovation really matters in this space, right? AI advances have in the past critically been all done out in the open. Neural nets were reinvented out in the open. The transformer paper was published readily out in the open. And being able to put this research and these advances and these models and be able to openly release them will allow broader innovation in society. It will allow us to advance science and allow many people to have broad access to critical technology. I think those are all positive things for society.
Ben (14:21)
If I could just add, I think a lot of folks forget that there’s a reason Linux is everywhere. There’s a reason Android powers our smartphones, right? Like we fundamentally understand the contours of the open source conversation. Like we know it’s important for transparency. You can inspect the technology. It’s important for competition. You can build on it without having to reinvent the wheel. And it’s important for security and privacy. You can fine tune AI models, run inference. You can build systems without all of that data going to an API in California. But I think there is a bit of a disconnect as Adam says between understanding the importance of open source in these previous tech cycles and understanding just how important open weights and open research have been to AI and will be in the future.
Matt (15:04)
Okay, I’d love to talk a little bit more about what’s at stake in thinking about open AI. Can you explain a little bit about what’s at stake here? What are other countries doing and why is it important that you guys are a strong competing force?
Ben (15:20)
Yes, the micro lens on this is, as with Linux, as with Android, developers care about having access to capable open alternatives to closed source technology. AI is going to be critical infrastructure across the economy. It’s transforming in the language model space how we access and interact with information. It’s transforming how we perceive and shape the visual world, visual AI. And we know that there is huge demand for capable open alternatives.
Our team has had 50 million downloads of our Flux open weight models on platforms like Hugging Face. The team collectively has released models of over 400 million downloads. These models rival Google, OpenAI, and we’re the only U.S. or European lab to rival DeepSeek by developer likes, to give you a sense for the size of this community. So there’s huge unmet demand out there.
And I think it reflects a few things. It reflects developers’ need to be able to inspect the technology before they deploy it. They want to be able to build exciting new applications that we can’t imagine today and do so without having to spend millions or hundreds of millions of dollars training their own models from scratch. And it’s important for security and privacy and for sovereignty as a result. If folks want to fine tune, optimize, integrate these models and then deploy them into real world systems, they need to be able to do so without sharing all of that data with an API on the other side of the planet. So we know that this huge unmet demand is out there in the global developer community.
I think the challenge is that if all of these models are coming from one particular player or one particular country, well, they embed certain values. They embed certain design choices and certain vulnerabilities. You see this most clearly with the führer around Deepseek right? DeepSeek was actually, in many respects, a center left model that would talk about most issues much as you’d expect an OpenAI or an Anthropic model.
But in certain areas, it had been heavily censored under Beijing’s cybersecurity regulations. And so it wouldn’t talk about Uyghurs, wouldn’t talk about Hong Kong, Taiwan. So what I say to folks is, imagine a world where the next search platform can’t talk about Taiwan, or in visual space, if the next generative tools can’t parody Xi Jinping. These are some really serious weighty issues.
And we’re seeing this too in our evaluations of models. If you take a look at our NCII and CSAM evaluations, our team has shown that there are sensible safeguards that can be put in place to drastically reduce the risk of widespread misuse. We know that certain models and certain teams based in China are not necessarily putting many of these mitigations and these safeguards in place. And so what does that then mean for the health and the integrity of the content ecosystem?
So it’s really important for policymakers to understand that open models are in some sense a vehicle for projecting soft power abroad. And the values and the choices and the vulnerabilities encoded and embedded in these models are going to determine behavior of all sorts of real world systems for the next few decades.
Adam Chen (18:18)
And open models also overlay on top of something else that’s critical for BFL, which is open research. And we believe really strongly that open research leads to more innovation with an AI and leads to more scientific advances and leads to that innovation and advances in a very responsible, acceptable way.
And so we believe that if you don’t allow the models, you basically increase the cost of innovation by developers, you increase the cost of innovation. And that has real implications on a strategic level, at a national level, at a geopolitical level, all of which needs to be considered before we decide whether open is bad and closed is good.
Matt (18:56)
And then how do you respond to the concerns that policymakers raise about openness? So you’re talking about all the benefits of openness for the good guys. But I think what we often hear when we’re talking about the benefits of open sources, people will express concerns about, well, if it’s open for the good guys, it’s also open for bad actors. And so it can be misused in different ways. So Adam, when you get to that point in the conversation with lawmakers, what do you say?
Adam Chen (19:19)
I think there’s a lot that one can do to kind of make sure that models have an appropriate level of safety when they’re being released, right? A lot of the work that Ben has done at BFL has been setting what the right levels and putting in the right testing mechanisms in place so that robust evaluations are being done on whatever models that are placed in the broader market.
And I think having a clear understanding of what we’re testing for, why we’re testing for it, and how we’re evaluating all of these results is very critical and making sure that we’re addressing these concerns that policymakers have, but also is overall just a good thing to do.
Ben (20:01)
I think a big part of getting policymakers comfortable with open weights is also just explaining the reasonable safeguards that we can put in place today to help lower that risk. We can talk about what we do with our models in the visual space, but the bottom line is there are mitigations that we can embed in the model themselves to help reduce the risk of misuse and we think reduce the risk of malicious modification as well.
I think part and parcel of that is also explaining to policymakers that there is always going to be a residual risk. And we need to think about other ways to mitigate that residual risk across the supply chain. So if you think about deepfake content, for example, sexual deepfakes, NCI, synthetic CSAM this isn’t just something that can only be fixed in the model. We also have to think about how content is distributed downstream, who’s using it, and what tools law enforcement has to respond to these forms of misuse. So I think when we just, we’re very candid and we’re very thoughtful and open with government, whether it’s members of Congress or leadership in the EU Commission, about what we can do and the state of the possible, but also the fact that there is that residual risk and those residual vulnerabilities. I think generally the conversation is in a much better place. And it also means that any reforms are going to be much more precise, more targeted, and less likely to interfere with that spirit of innovation.
Matt (21:19)
One thing that’s really impressed me about your approach to these issues is you haven’t been just saying, we’re great at safety, and published pretty statements about how great you are on safety. You’ve actually published data. You’ve really tried, I think, to say, like, here’s how we perform on various metrics that are important. So can you talk a little bit about the research that you recently released?
Ben (21:38)
Yeah, there are certain risks that don’t require a lot of context. You know it when you see it. And it’s really important to mitigate this in the model itself, not just in APIs, applications, and content distribution platforms. And so two of those risks are synthetic non-consensual imagery sexual deepfakes and synthetic child sexual abuse material. It’s really important to us that we make it easier for developers and users to do the right thing with our models and at the same time make it harder for bad actors to potentially misuse or modify those models. And so one of the things we’ve been working on is how we can better quantify these vulnerabilities in our models and use that data to make an informed pre-release decision.
And so working with one of our partners, Cinder, we built up a very comprehensive evaluation and benchmarking process to compare how our proposed checkpoints, our proposed release candidates perform compared to models that we’ve released previously and compared to other powerful open-weight models that are out there in the market. And the findings were really promising. They showed that with our pre-training and our post-training mitigations in the model, there are more than 10 times fewer vulnerabilities than models released by big tech firms in China. We showed that the post-training mitigations alone can help to reduce those vulnerabilities by more than 90% in some cases.
And we showed that while there is that residual risk, that residual risk can be nearly eliminated through the application of some very industry standard moderation practices of the API and the application layer. So it’s really exciting for a few reasons, but chief among them is that we can show that you can be both open and responsible at the same time, right? I think a lot of developers in this world and a lot of policymakers think of them as mutually exclusive or you’re either open or you’re responsible, you can’t be both at the same time. And what this data, what these evaluations help to show is that while there’s still a lot of work to be done and there’s still a lot of scope for improvement, that you can be both at the same time. And that’s really important if we want to continue to release this technology openly long into the future.
Matt (23:44)
So there’s a lot of debate right now, I think, on how developers can publish information that’s actually useful for people, but not too burdensome for developers. And it seems like you think that you’ve found a nice balance here. Adam, you were describing a legal team of four people, obviously significantly smaller, more limited resources than larger companies. Why is this approach to performance evaluations and disclosures one that works for BFL?
Adam Chen (24:11)
I think a large part of it has to do with focus and making sure we’re focused on what we actually really need to solve. Just to pull back a bit to, BFL is really about trying to build that kind of foundational layer for visual intelligence in the market. And so our customers can be quite wide ranging that may utilize our models in very different contexts.
For example, we’ve had customers before who are trying to [build] bespoke children’s book stories using AI for parents to kind of create for their children, very interesting for the child’s stories through the use of AI images. On the other hand, you can also have video game developers who may want to create AI assets as inspiration for their final visual assets within a game, right? And so you can see there what you may want to permit from a violence or gore perspective will vary widely. And so from our perspective, given our focus on providing that more broader infrastructure layer for all sorts of clients, we’re focused on just removing the stuff that is clearly illegal, and then provide guardrails and moderation standards for our developers who will then be able to make it more bespoke for their customers.
Matt (25:30)
It seems like part of this is a way that you have developed a company culture, company identity around combating the idea that openness is unsafe. It seems like your way to flip that is to say, well, if we provide openness, but we also provide measurable mitigations, we evaluate ourselves. We try to encourage downstream norms. We can get to a better place where we raise the bar on these types of issues. Is that a fair characterization?
Ben (25:58)
Yeah, I’ll just say I think part of this exercise is also helping policymakers, civil society, wider public think about open technology in a different way. Again, it’s not access to open technology that is our primary choke point and our primary intervention. It is how do we mitigate these residual risks right across the ecosystem, knowing that that technology is out there. We’ve done this with open software. We’ve done this with the open internet.
And I think fundamentally the same is going to be true of AI. Now, when I say that, I don’t want to diminish the fact that there are some very acute, very concrete risks associated with this technology. The uplift with AI and with visual AIs is meaningful. It’s significant over the baseline in some areas. But again, I think we need to think in a more joined up way about how these mitigations come together right across the tech stack. And to Adam’s point around moderation, if we can make available the tools and the resources that downstream developers need to implement this technology safely, then that’s an important way of mitigating some of these risks as well.
Matt (27:02)
For lots of people who do a policy for living, you spend a lot of your day reviewing various different policy proposals, lots of bills at the state level, lots of bills at the federal level. You guys are an international company, strong European presence. So you’re reading bills internationally as well. Well, there’s a lot of discussion about what the right disclosure model is now for companies and particularly for small companies.
Just based on your behavior, you’re not saying no disclosure is the best course. What’s the delta between what you guys are doing now and what then you see when you’re reading through disclosure mandate proposals from various different governments?
Ben (27:44)
I’ll be clear, I think a lot of transparency proposals that I’ve seen are actually trending in the right direction. I think the devil’s in the details, right? There’s a lot of challenges and there are a lot of challenges that can disproportionately affect startups compared to Big Tech. As Adam says, we have a small compliance team. You go to Google and there are these compliance teams that are much larger.
But I think the proposals that have been most challenging have been on the mitigation side. So proposals that try to draw a line in the sand about acceptable risk. They don’t just say disclose what evaluations you’ve performed, but they try to actually stipulate what acceptable risk looks like. And the challenge with that approach is that you create a world where it’s relatively easy for a closed source developer to comply. You make it much more difficult for an open source or an open weight developer to comply. You’re not really comparing apples to apples, you’re comparing apples to oranges. And so most of the proposals that we’ve seen become challenging when they veer out of pure disclosure, pure transparency territory, and they start moving towards either overt or implicit restrictions on the release of the model itself.
A good example of this is actually California, right? I think if you look at where SB 1047 was back in the day, and you look at where SB 53 and some of these proposals are going in the future, the trajectory is the right one, which is don’t stipulate secondary liability from model developers in relation to some exotic and fuzzy ill-defined harms. But there may be a well-designed targeted intervention focused on disclosure of your risk assessments, disclosure of the evaluations that you’ve performed. And without endorsing one bill or the other I think that is the right direction.
Matt (29:28)
That’s helpful in terms of understanding on substance where you come down. I’m curious as well, just how you deal with the jurisdictional complexity. Ben, you mentioned California and I know you have written in the past about 1047 and various different transparency models there. Adam, again, I’m seized by this four person team image in my head. That’s a small team for navigating like a compliance framework in Europe, a compliance framework that’s being set for the United States and Washington and then navigating all various different state compliance frameworks, including California, which has passed a lot of AI legislation in the last couple of years. How does that, even if you’re seeing a positive direction of travel, how do you guys navigate that state, federal, international patchwork?
Adam Chen (30:19)
I think maybe just addressing this from a more operational level. Number one is to hire really great people. So I think we’re very fortunate to have people like Ben and others kind of join us and agree to kind of tackle this crazy hodgepodge and emerging AI regulations all at the international, federal, European state level that is quite complex and interlayered.
More specifically though, I think trying to figure out the trends in which you take one set of actions, let’s say, disclose one certain way on the transparency side, and allow you to meet regulations across the board that oftentimes will have disproportionate results as well, which will be great and something the team looks to do. I think, though, there’s probably a larger question of if there’s a lot of regulations that conflict with each other.
What do you actually do? And when it comes to that, we’ll take on those issues when they come up. But I think those are hard, hard questions, which the team is well-equipped to tackle, given the experience we have.
Ben (31:23)
I think as a small team, it’s really important to triage in a really disciplined way. We are not a big tech advocacy operation. We need to be very careful about which battles we pick and where we can have an impact. From our perspective, you know, we look at the world through the lens of how can we best promote and protect open innovation and digital intelligence.
Fundamentally, it means researchers must be able to share the technology widely. It means the developers must be able to build on the technology responsibly. And so when we boil the ocean on 1800 state and federal bills just this year alone, and we look at 700 pages of the AI act during the trial process, for example, and 100 pages of codes of practice, we look at it through that lens. We say, where is this going to set back that mission? We’ll talk with anyone anywhere to be clear in any jurisdiction.
But we’re not interested in photo ops and roundtables and naval gazing just for the sake of it. Everything we do from an advocacy perspective on law reform has to serve that overriding mission. And when you’re that disciplined, you’ll find a lot of work kind of falls away. There’s a lot of busy work out there in the policy world. And we’re, as I say, ruthlessly focused on what really matters to researchers and developers.
Adam Chen (32:34)
Where can we have outsize impact? That’s really fundamentally one of the central questions we always ask ourselves, whether it’s a policy team, legal team, and the research team, and engineering team, and so on.
Matt (32:46)
A year ago, Adam, you didn’t have a policy team and then you decided that you would start one and that you bring Ben in to lead it. Can you explain a little bit about that decision and why did you decide that this was the time that you needed to have a policy function?
Adam Chen (32:59)
I think one of the central things too is when you find exceptional talent like Ben is, then you bring them on board no matter what they do. I think also more practically, AI regulation is a central part of any foundation model lab. It will be something that any foundation model lab will have to tackle throughout its journey from a startup to a larger company or if at a Big Tech company and the AI lab was in them.
And I think we recognize that early on that this would be such a unique thing for BFL on its journey that kind of tackling the problem ahead of time would give us some outsize returns. I think we’re seeing good results there too, engaging with policymakers and regulators early on makes it so that we’re not just some strange random lab operating out of the middle of Germany that no one knows about and is looked at more suspiciously and more fearfully by regulators. Instead, there’s somebody that regulators and others can interact with, we can engage on these issues and we can try to work through thoughtful solutions.
Matt (34:05)
Ben, as the founding member, the lead of this policy team, how are you thinking about the function and potential growth of your team?
Ben (34:13)
I just don’t think that the Big Tech playbook translates very well into startups and potentially even into model development. We have to think differently about how we do advocacy. I and our growing team, I think we need to be fundamentally committed to the idea that meaningful public oversight is compatible with open innovation. Our job is to explain the upside of this technology to policymakers, help to mitigate the downside in concrete ways, and help policymakers to reconcile the two through any more legislative or regulatory reforms.
From my perspective, with the number of jurisdictions we have, the outsized impact that we have as a business, I think it’s really important that we have, as I say, relentlessly focus on meaningful lasting impact in the regulatory system. And we don’t just create noise for its own sake. I am interested in targeted reforms and helping policymakers to get there. I’m not interested in generating PDFs and having photo ops and staging roundtables just for the sake of it. And I think if you’re a small team, it’s really important to have that discipline. We don’t want to have the kind of hyper-specialization that I think you see in some of the larger organizations. We need to stay focused and we need to be very clear about what matters to us and our team and our community and what maybe doesn’t matter.
Adam Chen (35:25)
And the focus as well, or the point of view that we bring as a small startup is somewhat unique, I think, from the Big Tech policy team and what the few points of view that that can bring. And I think we can see some outsize influence there or some disproportionate impact from making those voices heard.
We believe that a voice that can articulate a Little Tech point of view is important to ensuring that policymakers have a more robust understanding of the entire market and what sort of regulations they may be contemplating, and what impact that could have on innovation and competition.
Matt (35:58)
Lots of people speak about what that point of view is, we have a Little Tech policy agenda. We’ve talked about it extensively. How would each of you characterize what the Little Tech perspective is? And another way to say that is kind of what’s the competitive advantage that you bring as a little tech player to the policy ecosystem? You’re going up against these policy teams that have hundreds of employees, sometimes thousands of employees. What is the little tech voice that you’re bringing to the table?
Ben (36:25)
I think the value of a small team punching above its weight is that we have the agility to take some really bold positions on critical issues or major reforms in a way that might be more difficult for a larger team. We can sit down more as a thought partner with policymakers and work through the menu of options. I think one of the challenges with larger organizations, especially those that have come out of the kind of traditional internet platform space, everything is so colored by that Section 230 view of the world.
We need to maintain immunity on the one hand, but don’t worry, we’ve got these sort of privatized regulators internally, these trust and safety teams who will take care of it for us. And I think that’s important for all sorts of reasons, but it’s fundamentally a different model of advocacy and a different set of risks and options than you have in something like model development. So I think being able to take a bold view, sit down with policymakers, have a considered dialogue about the menu of options, I think that’s where small teams can have as much and sometimes more impact than large organizations.
Matt (37:26)
Of course there are downsides too. Are there policy debates you’ve been in so far where you’ve thought over significantly disadvantaged relative to larger players?
Ben (37:36)
I think larger players are sometimes perhaps blind to how these proposals impact small teams doing frontier work, right? Teams in our position. You have proposals out there that are very straightforward for a vertically integrated provider and very difficult for a provider who just forms one part of the tech stack. Those proposals are a relatively minor compliance overhead for a big tech firm and an absolute massive compliance ordeal for a small business in our position. And as I’ve said, you’ve got proposals out there that are relatively straightforward for a closed source provider and very, very difficult for an open source or open-weight developer. And so I think sometimes, larger organizations have so many competing interests in equities internally that they sometimes don’t understand the second and third order consequences of these proposals. And as a result, the effect of these proposals on small teams like ours gets lost in the noise.
So one thing I think has been promising is that a lot of policymakers, including many who we might disagree with on principle, are realizing that there is a more diverse community of developers out there than just the big fang or mango companies. And they’re willing to sit down and listen to these smaller teams and see how they can adjust these proposals in more targeted, more precise ways.
Adam Chen (38:54)
And overall, we’re very heartened by the fact that policymakers are aware that they often listen to a lot of policy discussions or have a lot of policy discussions with Big Tech policy teams and have less of them with smaller tech and smaller companies just because of a pure resourcing issue. And I think the fact that they’re willing to talk to us and willing to take our point of view into consideration means that these discussions do have an audience and it does have an impact. Providing that point of view, especially as to how some sort of policy could have an onerous impact on much smaller teams, I think is quite critical too. We often saw and heard when GDPR was put into place, as well as when it was implemented, that GDPR could actually kind of slow down Big Tech expansion within Europe when really after it was implemented, the exact opposite was the case because of how many lawyers the Big Tech companies could throw at this, being able to articulate this point of view to policymakers, I think that’s really powerful.
Matt (39:52)
Adam and Ben, this is a great conversation. Thanks so much for coming on the AI Policy Brief.
Ben (39:56)
Thanks, Matt. Appreciate it.
Adam Chen (39:58)
Matt, it is such a pleasure. Thank you so much.
This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.
