Latest / a16z AI Policy Brief / Open Models, Measurable Safeguards

Transcript

Open Models, Measurable Safeguards · a16z AI Policy Brief

0:00
I think a lot of developers and a lot of policymakers think
0:03
that you're either open
0:05
SPEAKER_02: or you're responsible.
0:05
SPEAKER_02: You can't be both at the same time.
0:07
And what this data,
0:08
what these evaluations helped show is that you can be both at
0:11
the same time.
0:12
SPEAKER_02: It's really important for policymakers to understand that
0:15
SPEAKER_02: models are a vehicle for projecting soft power abroad.
0:19
And the values and choices and vulnerabilities encoded and
0:22
embedded in these models
0:23
are going to determine behavior of all sorts of real world
0:27
systems for the next few
0:29
SPEAKER_02: decades.
0:30
I believe that a voice that can articulate from a little tech
0:32
point of view is
0:33
important to ensuring that policymakers have a more robust
0:36
understanding of the
0:38
SPEAKER_03: entire market and what regulations they may be contemplating,
0:41
SPEAKER_03: what impact that could have on innovation and competition.
0:43
Restricting access to general purpose technology should be a
0:47
measure of last
0:48
SPEAKER_02: resort,
0:49
SPEAKER_02: not a measure of first resort.
0:50
SPEAKER_00: Welcome back to the A16Z AI Policy Brief.
0:54
SPEAKER_00: Today,
0:54
SPEAKER_00: I'm joined by Adam Chen,
0:55
SPEAKER_00: Head of Legal,
0:56
SPEAKER_00: and Ben Brooks,
0:57
SPEAKER_00: Head of Public Policy at Black Forest Labs,
1:00
a company that has established itself as a pioneer in visual
1:03
intelligence through
1:04
SPEAKER_00: both its technical work and its commitment to open research.
1:07
SPEAKER_00: In this conversation,
1:08
we talk about what it means to build frontier visual AI and to
1:11
build it openly,
1:12
SPEAKER_00: the role of open models and advancing transparency,
1:15
SPEAKER_00: lowering costs for developers,
1:16
SPEAKER_00: and reducing dependence on closed APIs.
1:19
We discuss Black Forest Labs efforts to show that openness and
1:22
responsible
1:22
SPEAKER_00: development aren't at odds.
1:24
The team recently published safety evaluations showing their
1:27
latest models have
1:28
more than 10 times fewer vulnerabilities than other leading
1:30
open weight image
1:31
SPEAKER_00: models.
1:32
And we explore the broader policy challenges that come with
1:34
building in the open.
1:35
SPEAKER_00: Finally,
1:36
Adam and Ben share a candid look at what it takes to build a
1:38
policy function as a
1:39
SPEAKER_00: startup,
1:40
including the steps they have taken to enable their small team
1:43
to punch well above
1:43
SPEAKER_00: its weight.
1:44
SPEAKER_00: I really enjoyed this conversation and hope you will too.
1:52
SPEAKER_01: Welcome to the A16Z AI Policy Brief.
1:55
SPEAKER_01: Thank you.
1:55
SPEAKER_01: It's great to be here.
1:57
SPEAKER_01: Adam, can you tell us a little bit about Black Forest Labs?
1:59
SPEAKER_01: Like what is the company and what are you guys doing?
2:02
SPEAKER_03: Yeah, it's my pleasure to do so.
2:03
So BFL is a frontier research lab building the standard for
2:07
visual intelligence.
2:08
SPEAKER_03: It was founded in Freiburg, Germany.
2:10
SPEAKER_03: So the heart of the Black Forest, hence the name Black Forest Labs.
2:13
And really the founders were in university studying visual
2:18
intelligence and,
2:19
you know,
2:20
SPEAKER_03: at a certain period of time,
2:20
decided to kind of continue their work by building their own
2:23
startup from Germany.
2:25
And so our mission is to build foundational technology in
2:27
visual intelligence.
2:29
SPEAKER_03: And so what that will mean is essentially a high-performance,
2:31
SPEAKER_03: high-quality,
2:32
SPEAKER_03: lean,
2:33
SPEAKER_03: open,
2:34
SPEAKER_03: and fast visual intelligence models.
2:36
We do so,
2:37
and we want to do this by building openly,
2:40
by investing heavily in open research.
2:42
And we ultimately believe that this will result in better gains
2:45
for the company and
2:46
SPEAKER_03: for society as a whole in the long run.
2:49
SPEAKER_03: As in terms of myself,
2:50
SPEAKER_03: I am working the legal team here at Black Forest Labs,
2:52
SPEAKER_03: setting up the legal team.
2:54
SPEAKER_03: Joined about a year ago,
2:56
and I'm very excited for all of the work that the team has been
3:00
building towards.
3:02
SPEAKER_01: Adam, you are referencing the legal team at Black Forest.
3:06
SPEAKER_01: I think sometimes when people hear legal team,
3:08
SPEAKER_01: they're thinking of dozens or hundreds or thousands of lawyers.
3:12
Can you give us a little bit of a sense of what does the legal
3:14
team look like at
3:15
SPEAKER_01: Black Forest Labs?
3:17
SPEAKER_01: Very small.
3:19
SPEAKER_03: So in terms of people who are acting as lawyers, it's about four people.
3:24
And that's two on the commercial side and myself and another
3:28
lawyer on the product
3:30
SPEAKER_03: counseling side.
3:31
SPEAKER_03: And so it's very minute,
3:33
SPEAKER_03: I think,
3:34
SPEAKER_03: in terms of the scale as compared to some of our competitors out there.
3:38
SPEAKER_01: And you've recently grown the function to include policy as well.
3:42
SPEAKER_01: Ben, you've recently joined in a full-time role.
3:45
SPEAKER_01: So Ben, what are you doing at Black Forest?
3:49
Yeah,
3:50
I mean,
3:50
I lead public policy and a lot of what we're calling model
3:53
assurance works.
3:54
In other words,
3:55
how do we validate the claims that we're making about the risk
3:58
in our models?
3:59
SPEAKER_02: I've known my team for quite a while now in one form or another.
4:02
I first reached out to the team when they were working on
4:05
stable diffusion back in the day.
4:07
We had a congresswoman who was writing to the White House
4:09
saying we need export
4:10
SPEAKER_02: controls to stop the release of this technology.
4:13
SPEAKER_02: And that's how I got involved.
4:16
SPEAKER_02: And then most recently joined to lead our public policy work.
4:20
SPEAKER_02: Before that was a fellow at the Burton Pine Center at Harvard.
4:23
SPEAKER_02: And before that,
4:24
SPEAKER_02: a whole lot of regulatory advocacy and some really spicy,
4:28
SPEAKER_02: high stakes,
4:29
SPEAKER_02: usually permission-based domains,
4:31
SPEAKER_02: right?
4:32
From drones through to crypto,
4:34
ride sharing and a whole bunch of stuff in between.
4:37
SPEAKER_02: So for me, this has always been an exercise in how can we
4:41
regulate in the public interest but do so in a way that is
4:44
maximally compatible
4:45
with open innovation and this is more important in AI than ever
4:50
before and this
4:50
SPEAKER_02: team has been at the forefront of showing the world that there is an open
4:56
SPEAKER_02: and capable future out there.
4:57
SPEAKER_02: It doesn't just have to be closed source API.
5:01
SPEAKER_02: So it's really exciting.
5:02
SPEAKER_02: Worked with Adam for some time.
5:04
SPEAKER_02: And as Adam says,
5:06
SPEAKER_02: it's a lean agile team,
5:08
SPEAKER_02: but we're punching above our weight in all the ways that matter.
5:12
So you guys have both done a range of different things in tech
5:15
policy and law.
5:17
Ben,
5:17
maybe starting with you and then Adam,
5:18
I'd love to hear your thoughts.
5:19
SPEAKER_01: Like why BFL?
5:21
What is compelling to you about this opportunity versus Ben's
5:25
work in academia or
5:26
SPEAKER_01: other companies,
5:27
SPEAKER_01: large and small,
5:27
SPEAKER_01: that you might be at?
5:28
SPEAKER_02: Yeah, I mean, I think it's a combination of things.
5:31
SPEAKER_02: From the perspective of our team, I mean, I think having a team that are
5:35
SPEAKER_02: comprised of missionaries, not mercenaries, right?
5:37
SPEAKER_02: They deeply care about research.
5:39
They care about moving the frontier in all the ways that matter
5:42
to real world
5:43
SPEAKER_02: developers and deployers.
5:45
SPEAKER_02: And they care about releasing this technology and this research openly.
5:48
SPEAKER_02: And to me, that was a huge drawcard.
5:50
SPEAKER_02: And it sets them apart from a lot of the other players in this space.
5:54
SPEAKER_02: I think there are also some
5:57
SPEAKER_02: interesting, fairly unique challenges that a team like BFL faces.
6:01
We're doing frontier work,
6:03
which means that we have some of the most capable models out
6:05
there.
6:06
We're doing it openly and open models have unique challenges
6:10
and unique properties
6:11
SPEAKER_02: that we can maybe discuss today.
6:12
SPEAKER_02: And we're a small team, right?
6:14
SPEAKER_02: So we're up against the Googles of the world,
6:16
SPEAKER_02: but we don't necessarily have the same ability to absorb risk.
6:20
SPEAKER_02: And I think we need to be very careful about how we go about this work.
6:23
SPEAKER_02: We need to show the world that we develop this stuff fourthly.
6:26
SPEAKER_02: And so,
6:27
SPEAKER_02: you know,
6:27
SPEAKER_02: a really compelling mission,
6:28
SPEAKER_02: a really important mission,
6:30
SPEAKER_02: a great team coupled with some very unique Edwins,
6:34
I think makes this a really attractive place to continue that
6:38
mission with
6:38
SPEAKER_02: government and the policymakers.
6:41
SPEAKER_03: Yeah.
6:42
SPEAKER_03: And on my end, you know, the team is exceptional here.
6:46
Everybody is really,
6:48
really talented at what they do and are great people.
6:50
SPEAKER_03: And that was a really great draw for me.
6:53
SPEAKER_03: And I think...
6:55
SPEAKER_03: Quite honestly,
6:56
SPEAKER_03: you know,
6:56
one of the most exciting things to do in law and policy right
7:00
now is working on the
7:01
SPEAKER_03: frontier.
7:02
SPEAKER_03: And right now, AI is the frontier of both law policy and also technology.
7:07
And so this team is looking to push that frontier to expand the
7:11
capabilities of
7:12
SPEAKER_03: what AI models can do.
7:14
SPEAKER_03: And that makes professionally a very exciting challenge.
7:18
SPEAKER_01: So I want to get into the specifics of the law and policy issues,
7:21
SPEAKER_01: but to get there,
7:22
I think we need to understand a little bit more about the
7:23
technology itself.
7:25
SPEAKER_01: So when you talk about building visual capability at the frontier,
7:28
SPEAKER_01: what does that actually mean in practice?
7:30
Yeah,
7:30
I mean,
7:30
well,
7:31
visual intelligence,
7:31
think about language models,
7:33
right?
7:33
Language is a really abstract and highly compressed
7:36
representation of the world.
7:37
SPEAKER_02: And so we have models out there,
7:39
SPEAKER_02: fantastic models that can learn to manipulate text for a whole range of
7:43
SPEAKER_02: applications,
7:44
SPEAKER_02: coding,
7:44
SPEAKER_02: search,
7:45
SPEAKER_02: everything in between.
7:46
But their understanding of the world is fundamentally limited
7:48
in important ways.
7:49
SPEAKER_02: And our team's belief has always been that pixels encode a much richer
7:53
SPEAKER_02: representation of the world.
7:54
SPEAKER_02: If you think about us as people,
7:57
vision is primarily how we perceive and understand and interact
8:01
with the world
8:01
SPEAKER_02: around us.
8:03
SPEAKER_02: And so,
8:03
SPEAKER_02: you know,
8:04
the bet that we're making as a team is that visual AI is going
8:07
to transform the
8:08
SPEAKER_02: world,
8:08
SPEAKER_02: not just in these sort of very creative,
8:10
SPEAKER_02: highly expressive ways that we see today,
8:12
SPEAKER_02: but in more functional ways as well.
8:14
SPEAKER_02: If you think about today's sort of current crop of visual models,
8:18
SPEAKER_02: we have relatively narrow capabilities,
8:20
SPEAKER_02: but we still have some high impact applications,
8:23
SPEAKER_02: film,
8:23
SPEAKER_02: gaming,
8:25
SPEAKER_02: you know,
8:26
SPEAKER_02: designs,
8:27
SPEAKER_02: concepts in architecture.
8:28
SPEAKER_02: But where we're moving towards is more of a
8:31
a unified visual intelligence architecture where you have
8:34
versatile models that
8:35
SPEAKER_02: bring together perception of the visual world,
8:38
SPEAKER_02: simulation,
8:39
SPEAKER_02: visual reasoning across different modalities and do so in one scalable
8:43
SPEAKER_02: architecture.
8:44
SPEAKER_02: So in the short term,
8:46
SPEAKER_02: that means better simulation environments for agents,
8:49
SPEAKER_02: embodied systems that are using models to learn how to navigate complex
8:53
SPEAKER_02: environments.
8:54
SPEAKER_02: That's really exciting.
8:55
SPEAKER_02: I think long term,
8:56
this means potentially embedding these models in high stakes
9:00
systems themselves,
9:01
SPEAKER_02: right?
9:01
SPEAKER_02: The model isn't just
9:03
predicting or generating the next frame in a video sequence but
9:06
it's actually
9:06
outputting an action or a recommendation and this is
9:10
fundamentally how we're going
9:11
to unlock specialized robotics web agents with computer use
9:16
embodied systems that
9:17
SPEAKER_02: can manipulate physical tools in really consequential environments
9:21
SPEAKER_02: So for us,
9:22
SPEAKER_02: it's not push a button,
9:23
get an image is kind of where the technology is historically
9:26
associated today.
9:27
But long term,
9:28
there are some really impactful,
9:30
important functional applications of this stuff.
9:33
SPEAKER_02: And our team, as I say, is pushing the criteria.
9:36
SPEAKER_01: Adam,
9:36
when Ben was talking in his opening just about the nature of
9:39
the work and he was
9:39
talking about visual capability being compelling from a
9:43
technology perspective,
9:44
SPEAKER_01: he was also saying it raises unique challenges in the regulatory context.
9:48
SPEAKER_01: What are those challenges?
9:49
SPEAKER_01: And
9:50
SPEAKER_01: How are you guys responding to them?
9:54
SPEAKER_03: I mean,
9:54
SPEAKER_03: I think at core base,
9:55
there are some visual content that is just outright illegal in
10:00
many jurisdictions,
10:00
SPEAKER_03: right?
10:01
SPEAKER_03: And so you'll see that with CSAM,
10:03
SPEAKER_03: child sexual exploitation abuse material,
10:06
SPEAKER_03: and non-consensual intimate imagery,
10:09
SPEAKER_03: NCII.
10:10
And so one of the things we really try to do is to tackle those
10:15
type of core risks
10:16
SPEAKER_03: that exist right now with visual models.
10:20
at a really baseline model layer and not to just kind of try to
10:23
paper over it with
10:24
SPEAKER_03: some sort of like inference moderation layer itself.
10:28
And the other policy issue that you were sort of flagging in
10:30
your opening was how
10:32
SPEAKER_01: open you decided to be.
10:34
So what is the feedback that you hear from the policy community
10:37
about your
10:38
SPEAKER_01: decisions around openness?
10:39
SPEAKER_02: So I think the open models, as I said, pose some unique challenges.
10:43
you can yeah you can modify them for purposes that were not
10:45
anticipated by the
10:47
original developer you can run inference on them independently
10:50
without moderation
10:51
SPEAKER_02: layer filters and if there's something wrong with the model if there's a
10:54
vulnerability it can be very difficult to withdraw that model
10:57
from circulation once
10:58
it's out there on the internet on platforms like hugging face
11:00
or github and so
11:01
we've always been very sensitive to the challenges posed by
11:04
open models i think one
11:06
SPEAKER_02: of the difficulties is that when policy makers look at ai
11:09
SPEAKER_02: They expect us to mitigate a lot of these risks in the model layer.
11:13
SPEAKER_02: And we're talking everything from political deepfakes,
11:16
SPEAKER_02: sexual deepfakes,
11:17
SPEAKER_02: automated decision-making,
11:19
SPEAKER_02: negligence,
11:19
SPEAKER_02: defective product design,
11:21
SPEAKER_02: copyright issues.
11:22
SPEAKER_02: All of this is expected to be done at the model layer.
11:25
And the challenge is that many of these risks are highly
11:27
context-dependent,
11:29
SPEAKER_02: and it's not necessarily possible to mitigate that in the model.
11:32
SPEAKER_02: The models are versatile.
11:33
SPEAKER_02: You can do many things with them.
11:35
SPEAKER_02: Most of it good, some of it bad.
11:37
those capabilities are often entangled in ways that can be
11:39
difficult to tease apart.
11:41
SPEAKER_02: And then on robust,
11:42
SPEAKER_02: if you embed a model with particular safeguards,
11:45
SPEAKER_02: you can,
11:46
SPEAKER_02: with enough data,
11:47
SPEAKER_02: compute and expertise,
11:48
SPEAKER_02: start to unwind some of those safeguards.
11:50
So I think that mismatch is one of the biggest challenges we
11:53
face with policymakers.
11:54
The expectation that models are the silver bullet and we're
11:57
going to fix all of
11:57
these challenges in the model and the reality that actually the
12:00
model is just one
12:01
SPEAKER_02: component in a complex integrated system.
12:04
SPEAKER_02: These risks are context dependent.
12:07
And while there are many important safeguards we can put in
12:09
place in the model,
12:10
SPEAKER_02: that isn't the end of the story.
12:12
SPEAKER_02: And from a regulatory perspective,
12:14
SPEAKER_02: it's really important that we capture some of that nuance in these future
12:18
SPEAKER_02: legislative and regulatory reforms.
12:21
SPEAKER_01: I assume when you raise that point with lawmakers, they say,
12:25
SPEAKER_01: okay,
12:26
SPEAKER_01: we understand that it's difficult,
12:27
but are you just expecting us to live with a world where all
12:30
these harms are
12:31
SPEAKER_01: created?
12:32
So what is your vision for what the right approach is from a
12:36
policy perspective if
12:37
SPEAKER_01: you're concerned about the harms,
12:38
but you recognize some of the limitations of addressing that at
12:40
the model layer?
12:41
SPEAKER_02: I think it's a few things.
12:42
I think one is just being very clear that restricting access to
12:46
general purpose
12:47
SPEAKER_02: technology should be a measure of last resort,
12:49
SPEAKER_02: not a measure of first resort.
12:51
SPEAKER_02: And too often, I think policymakers are looking at
12:54
those kinds of interventions as the first thing we should do
12:58
and not the last thing
12:58
SPEAKER_02: we should do.
13:00
SPEAKER_02: You think about,
13:00
SPEAKER_02: you know,
13:01
SPEAKER_02: just the past couple of years,
13:02
SPEAKER_02: the number of builds we've seen around licensing model developers,
13:05
SPEAKER_02: around export controls for model weights,
13:09
SPEAKER_02: Some really interesting,
13:10
fairly exotic liability proposals that make life very difficult
13:13
for open source
13:14
SPEAKER_02: researchers and developers.
13:16
SPEAKER_02: I think those kinds of reforms,
13:17
SPEAKER_02: we need to be very clear,
13:19
have a huge impact on open innovation and they should be
13:21
interventions of last
13:22
SPEAKER_02: resort.
13:23
SPEAKER_02: I think the second piece,
13:24
SPEAKER_02: though,
13:24
is like I would love to see governments run a systematic gap
13:27
analysis of where our
13:29
SPEAKER_02: existing regulatory systems fall short.
13:31
SPEAKER_02: And so often we're kind of legislating before we do that gap analysis.
13:35
SPEAKER_02: And the truth is there are gaps.
13:37
It was only last year that the US government introduced federal
13:40
criminal liability
13:41
SPEAKER_02: for non-consensual intimate imagery.
13:42
And CII is a huge proportion of the risks that policymakers and
13:47
real families out
13:48
SPEAKER_02: there in the real world care about.
13:50
SPEAKER_02: But it was only last year that we plugged that gap at a federal level.
13:53
SPEAKER_02: I think the same is true on a range of other different AI systems too.
13:56
SPEAKER_02: So we'd love to see more of a systematic gap analysis there,
13:59
figure out across the supply chain from developers through to
14:02
users and everyone in
14:03
SPEAKER_02: between,
14:03
where do our existing systems fall short and where can we
14:06
introduce targeted
14:07
SPEAKER_02: reforms to up-level the confidence in the AI supply chain?
14:13
SPEAKER_02: So a lot of different things.
14:14
SPEAKER_02: We can double click on any of them,
14:16
SPEAKER_02: but generally speaking,
14:17
it's really important that we make sure that whatever we do for
14:21
whatever risk,
14:22
It's compatible with the culture of experimentation and open
14:25
innovation that
14:26
brought us here in the first place and is going to help to make
14:29
AI useful and
14:30
SPEAKER_03: accessible in the future.
14:31
It's also really critical to also remind ourselves why open
14:34
innovation really
14:35
SPEAKER_03: matters in this space,
14:36
SPEAKER_03: right?
14:37
SPEAKER_03: AI advances has in the past critically been all done out in the open.
14:43
SPEAKER_03: Neuronets were reinvented out in the open and transformer paper.
14:47
SPEAKER_03: was published readily in the open.
14:50
And I think kind of being able to put this research and these
14:54
advances and these
14:56
models and be able to openly release them will allow broader
15:00
innovation in society.
15:02
It will allow us to advance science and allow many people to
15:06
have broad access to
15:07
SPEAKER_03: critical technology.
15:08
I think those are all positive human,
15:11
those are all positive things for society.
15:14
SPEAKER_02: If I could just add, like, I think a lot of folks forget that
15:17
SPEAKER_02: There's a reason Linux is everywhere.
15:19
SPEAKER_02: There's a reason Android powers the majority of the world's smartphones.
15:22
SPEAKER_02: We fundamentally understand contours of the open source conversation.
15:26
SPEAKER_02: We know it's important for transparency.
15:28
SPEAKER_02: You can inspect the technology.
15:29
SPEAKER_02: It's important for competition.
15:31
SPEAKER_02: You can build on it without having to reinvent the wheel.
15:34
SPEAKER_02: And it's important for security and privacy.
15:35
SPEAKER_02: You can fine-tune AI models, run inference.
15:38
You can build systems without all of that data going to an API
15:41
in California.
15:42
SPEAKER_02: But I think there is a bit of a disconnect,
15:44
SPEAKER_02: as Adam says,
15:45
between understanding the importance of open source in these
15:48
kind of previous tech
15:49
cycles and understanding just how important open weights and
15:53
open research have
15:54
SPEAKER_02: been to AI and will be in the future.
15:56
SPEAKER_01: Great.
15:57
I'd love to talk a little bit more about what's at stake in
16:01
thinking about open AI.
16:03
SPEAKER_01: Can you explain a little bit about what's at stake here?
16:06
What are other countries doing and why is it important that you
16:09
guys are a strong
16:10
SPEAKER_01: competing force?
16:13
SPEAKER_02: Yes,
16:13
SPEAKER_02: I think the micro lens on this is,
16:14
SPEAKER_02: as with Linux,
16:15
SPEAKER_02: as with Android,
16:17
developers care about having access to capable open
16:20
alternatives to closed source
16:22
SPEAKER_02: technology.
16:22
SPEAKER_02: AI is going to be critical infrastructure across the economy.
16:24
It's transforming in the language model space how we access and
16:28
interact with information.
16:29
SPEAKER_02: It's transforming how we perceive and shape
16:32
visual world visual ai and we know that there is huge demand
16:35
for capable open
16:36
alternatives you know our team has had 50 million downloads of
16:39
our flux open weight
16:40
models on platforms like hugging face the team collectively has
16:43
released models
16:44
with over 400 million downloads these models rival google open
16:48
ai and we're the
16:49
only us or european lab to rival deep seek by developer likes
16:53
right to give you a
16:54
sense for the the size of this community so there's huge unmet
16:57
demand out there and
16:58
SPEAKER_02: i think it reflects a few things i think it reflects
17:01
Developers need to be able to inspect the technology before
17:03
they deploy it.
17:04
They want to be able to build exciting new applications that we
17:08
can't imagine today
17:09
and do so without having to spend millions or hundreds of
17:11
millions of dollars
17:12
SPEAKER_02: training their own models from scratch.
17:14
And it's important for security and privacy and for sovereignty
17:17
as a result.
17:17
SPEAKER_02: Folks want to fine tune,
17:20
SPEAKER_02: optimize,
17:20
SPEAKER_02: integrate these models and then deploy them into real world systems.
17:24
They need to be able to do so without sharing all of that data
17:27
with an API on the
17:28
SPEAKER_02: other side of the planet.
17:30
So we know that this huge unmet demand is out there in the
17:32
global developer community.
17:34
I think the challenge is that if all of these models are coming
17:37
from one particular
17:39
SPEAKER_02: player or one particular country,
17:41
SPEAKER_02: well,
17:41
SPEAKER_02: they embed certain values,
17:42
SPEAKER_02: they embed certain design choices and certain vulnerabilities.
17:45
SPEAKER_02: You see this most clearly with the furor around DeepSeek, right?
17:49
SPEAKER_02: DeepSeek was actually, in many respects, a center-left model
17:54
That would talk about most issues,
17:55
much as you'd expect an open AI or an anthropic model.
17:58
But in certain areas,
17:59
it had been heavily censored under Beijing's cybersecurity
18:03
regulations.
18:04
And so it wouldn't talk about the Uyghurs,
18:06
it wouldn't talk about Hong Kong,
18:07
Taiwan.
18:08
So what I say to folks is imagine a world where the next search
18:11
platform can't talk
18:13
about Taiwan or in the visual space if the next generative
18:18
tools can't parody Xi
18:20
SPEAKER_02: Jinping.
18:21
SPEAKER_02: This is some really serious weighty issues.
18:25
SPEAKER_02: And we're seeing this too in our evaluations of models.
18:28
SPEAKER_02: If you take a look at our NCII and CSAM evaluations,
18:31
our team has shown that there are sensible safeguards that can
18:34
be put in place to
18:35
SPEAKER_02: drastically reduce the risk of widespread misuse.
18:39
We know that certain models and certain teams based in China
18:43
are not necessarily
18:43
SPEAKER_02: putting many of these mitigations and these safeguards in place.
18:46
And so what does that then mean for the health and the
18:48
integrity of the content ecosystem?
18:50
So it's really important,
18:52
I think,
18:52
for policymakers to understand that open models are a
18:56
SPEAKER_02: some sense of a vehicle for projecting soft power abroad.
19:00
And the values and the choices and the vulnerabilities encoded
19:03
and embedded in
19:03
these models are going to determine behavior of all sorts of
19:07
real world systems for
19:09
SPEAKER_03: the next few decades.
19:10
SPEAKER_03: And open model,
19:12
I think open models also overlay on top of something else
19:15
that's critical for BFL,
19:16
SPEAKER_03: which is open research.
19:18
And we believe really strongly that open research leads to more
19:21
innovation within
19:22
AI and leads to more scientific advances and leads to that
19:25
innovation and advances
19:26
SPEAKER_03: in a very responsible,
19:27
SPEAKER_03: acceptable way.
19:29
SPEAKER_03: And so we believe that if you don't allow the models, you basically
19:33
increase the cost of innovation by developers,
19:35
you increase the cost of innovation.
19:36
SPEAKER_03: And that has real implications on a strategic level,
19:40
SPEAKER_03: at a national level,
19:41
SPEAKER_03: at a geopolitical level,
19:43
all of which needs to be considered before we decide whether
19:46
open is bad and,
19:48
SPEAKER_03: you know,
19:48
SPEAKER_03: close is good.
19:49
And then how do you respond to the concerns that policymakers
19:52
raise about openness?
19:53
SPEAKER_01: So I think
19:54
SPEAKER_01: You're talking about all the benefits of openness for the good guys.
19:57
But I think what we often hear when we're talking about the
19:59
benefits of open
20:00
SPEAKER_01: sources,
20:00
SPEAKER_01: people will express concerns about,
20:02
SPEAKER_01: well,
20:03
SPEAKER_01: if it's open for the good guys,
20:04
SPEAKER_01: it's also open for bad actors.
20:06
SPEAKER_01: And so it can be misused in different ways.
20:08
SPEAKER_01: So, Adam, when you get to that point,
20:09
SPEAKER_01: in the conversation with lawmakers, what do you say?
20:11
I think there's a lot that one can do to kind of make sure that
20:16
models have their
20:18
SPEAKER_03: appropriate level of safety when they're being released,
20:20
SPEAKER_03: right?
20:21
A lot of the work that Ben has done at BFL has been about
20:24
setting what the right
20:26
levels and putting in the right testing mechanisms in place so
20:30
that robust
20:31
evaluations are being done on whatever models that are placed
20:34
in the broader
20:35
SPEAKER_03: market.
20:36
SPEAKER_03: And I think having a clear understanding of what we're testing for,
20:40
SPEAKER_03: why we're testing for it,
20:41
and how we're evaluating all of these results is very critical
20:45
in making sure that
20:47
SPEAKER_03: we're addressing these concerns that policymakers have,
20:51
SPEAKER_03: but also is overall just a good thing to do.
20:54
SPEAKER_02: I think a big part of
20:56
SPEAKER_02: getting policymakers comfortable with open weights.
20:59
It's also just explaining the reasonable safeguards that we can
21:02
put in place today
21:02
SPEAKER_02: to help lower that risk.
21:04
SPEAKER_02: We can talk about what we do with our models in the visual space,
21:07
but the bottom line is there are mitigations that we can embed
21:10
in the model
21:11
SPEAKER_02: themselves to help reduce the risk of misuse,
21:14
SPEAKER_02: and we think reduce the risk of malicious modification as well.
21:18
SPEAKER_02: I think
21:19
Part and parcel of that is also explaining to policymakers that
21:22
there is always
21:23
SPEAKER_02: going to be a residual risk.
21:25
And we need to think about other ways to mitigate that residual
21:28
risk across the supply chain.
21:29
SPEAKER_02: So if you think about deepfake content, for example, sexual deepfakes,
21:34
SPEAKER_02: NCI, synthetic CTAN.
21:35
SPEAKER_02: This isn't just something that can only be fixed in the model.
21:37
SPEAKER_02: We also have to think about how content is distributed downstream,
21:41
SPEAKER_02: who's using it,
21:42
SPEAKER_02: and what tools law enforcement has to respond to these forms of misuse.
21:45
So I think when we're just very candid and we're very
21:47
thoughtful and open with
21:49
SPEAKER_02: government,
21:50
SPEAKER_02: whether it's members of Congress or leadership in the EU Commission,
21:53
SPEAKER_02: about what we can do and the state of the possible,
21:56
SPEAKER_02: but also the fact that there is that residual risk and those residual
21:59
SPEAKER_02: vulnerabilities,
22:00
SPEAKER_02: I think generally the conversation's in a much better place.
22:03
SPEAKER_02: It also means that any reforms are going to be much more precise,
22:05
SPEAKER_02: more targeted,
22:06
SPEAKER_02: and less likely to interfere with that spirit of innovation.
22:12
One thing that's really impressed me about your approach to
22:14
these issues is you
22:14
SPEAKER_01: haven't been just saying,
22:16
SPEAKER_01: we're great at safety and
22:18
SPEAKER_01: published pretty statements about how great you are on safety.
22:21
SPEAKER_01: You've actually published data.
22:22
SPEAKER_01: You've really tried,
22:23
SPEAKER_01: I think,
22:23
SPEAKER_01: to say,
22:24
SPEAKER_01: like,
22:24
SPEAKER_01: here's how we perform on various metrics that are important.
22:27
So can you talk a little bit about the research that you
22:29
recently released?
22:30
SPEAKER_01: Yeah.
22:31
So,
22:31
you know,
22:31
there are certain risks that don't require a lot of context.
22:35
SPEAKER_02: You know it when you see it.
22:37
SPEAKER_02: And it's really important to mitigate this in the model itself,
22:40
SPEAKER_02: not just in APIs,
22:42
SPEAKER_02: applications and content distribution platforms.
22:45
And so two of those risks are synthetic non-consensual
22:47
intimidatory sexual
22:48
SPEAKER_02: deepfakes and synthetic child sexual abuse material.
22:52
It's really important to us that we make it easier for
22:54
developers and users to do
22:56
the right thing with our models and at the same time make it
22:58
harder for bad actors
22:59
SPEAKER_02: to potentially misuse or modify those models.
23:02
And so one of the things we've been working on is how can we
23:05
better quantify these
23:07
vulnerabilities in our models and use that data to make an
23:09
informed pre-release
23:11
SPEAKER_02: decision.
23:11
SPEAKER_02: And so working with one of our partners,
23:13
SPEAKER_02: Cinder,
23:14
we built up a very comprehensive evaluation and benchmarking
23:18
process to compare how
23:20
SPEAKER_02: our proposed checkpoints,
23:22
our proposed release candidates perform compared to models that
23:25
we've released
23:26
previously and compared to other powerful open-weight models
23:29
that are out there in
23:29
SPEAKER_02: the market.
23:31
SPEAKER_02: And the findings were really dancing.
23:33
They showed that with our pre-training and our post-training
23:36
mitigations in the
23:37
SPEAKER_02: model,
23:38
there are more than 10 times fewer vulnerabilities than models
23:42
released by big tech
23:43
SPEAKER_02: firms in China.
23:44
We showed that the post-training mitigations alone can help to
23:48
reduce those
23:48
SPEAKER_02: vulnerabilities by more than 90% in some cases.
23:53
SPEAKER_02: And we showed that while there is that residual risk,
23:55
that residual risk can be nearly eliminated through the
23:59
application of some very
24:00
industry standard moderation practices of the API and the
24:03
application layer.
24:04
SPEAKER_02: So it's really exciting for a few reasons,
24:05
but chief among them is we can show that you can be both open
24:10
and responsible at
24:11
SPEAKER_02: the same time,
24:11
SPEAKER_02: right?
24:12
I think a lot of developers in this world and a lot of
24:15
policymakers think of them
24:16
SPEAKER_02: as mutually exclusive,
24:18
SPEAKER_02: one or the other.
24:19
SPEAKER_02: You're either open or you're responsible.
24:21
SPEAKER_02: You can't be both at the same time.
24:23
SPEAKER_02: And what this data,
24:24
what these evaluations helped show is that while there's still
24:27
a lot of work to be
24:28
SPEAKER_02: done and there's still a lot of scope for improvement,
24:30
SPEAKER_02: that you can be both at the same time.
24:32
And that's really important if we want to continue to release
24:34
this technology
24:35
SPEAKER_02: openly long into the future.
24:37
SPEAKER_01: So there's a lot of debate right now,
24:39
SPEAKER_01: I think,
24:39
on how developers can publish information that's actually
24:43
useful for people,
24:44
SPEAKER_01: but not too burdensome for developers.
24:46
SPEAKER_01: And it seems like you think that you've found a nice balance here.
24:50
SPEAKER_01: Adam,
24:51
SPEAKER_01: you were describing a legal team of four people,
24:53
SPEAKER_01: obviously significantly smaller,
24:55
SPEAKER_01: more limited resources than larger companies.
24:57
SPEAKER_01: Why is this approach
24:59
SPEAKER_01: to performance evaluations and disclosures, one that works for BFL?
25:04
I think a large part of it has to do with focus and making sure
25:08
we're focused on
25:09
SPEAKER_03: what we actually really need to solve.
25:11
SPEAKER_03: Just to pull back a bit to, you know, BFL is really...
25:14
about trying to build that kind of foundational layer for
25:17
visual intelligence in the market.
25:19
And so our customers can be quite wide ranging that may utilize
25:24
our models in very,
25:25
SPEAKER_03: very different contexts.
25:26
SPEAKER_03: For example,
25:26
SPEAKER_03: you know,
25:27
we could have a customer and we've had customers before who are
25:30
trying to create
25:31
bespoke children's book stories using AI for parents to kind of
25:35
create for their
25:36
SPEAKER_03: children.
25:37
SPEAKER_03: Very, very kind of interesting for the child's stories.
25:41
SPEAKER_03: through the use of AI images.
25:42
SPEAKER_03: On the other hand,
25:43
you can have also video game developers who may want to create
25:47
AI assets as
25:48
SPEAKER_03: inspiration for their final visual assets within a game,
25:51
SPEAKER_03: right?
25:52
And so you can see there just from a kind of pure like what you
25:56
may want to permit
25:57
SPEAKER_03: from a violence or gore perspective will vary widely.
26:01
SPEAKER_03: And so from our perspective,
26:02
given our focus on providing that more broader infrastructure
26:06
layer for all sorts
26:07
SPEAKER_03: of clients,
26:08
SPEAKER_03: we're focused on
26:10
just removing the stuff that is clearly illegal and then
26:13
provide guard routes and
26:15
moderation standards for our developers who will then be able
26:19
to kind of make it
26:21
SPEAKER_03: more bespoke for their customers.
26:23
SPEAKER_01: It seems like part of this is a way that you have developed kind of a
26:28
company culture company identity around combating the idea that
26:32
openness is unsafe
26:33
it seems like you're basically your your sort of way to flip
26:36
that is to say well if
26:38
we provide openness but we also provide measurable mitigations
26:41
we evaluate
26:42
ourselves we try to encourage downstream norms we can get to a
26:46
better place where
26:47
SPEAKER_01: we raise the bar on these types of issues is that a fair characterization
26:50
SPEAKER_02: Yeah,
26:50
SPEAKER_02: I'll just say I think part of this exercise is also helping policymakers,
26:55
SPEAKER_02: civil society,
26:56
SPEAKER_02: the wider public,
26:57
SPEAKER_02: think about open technology in a different way.
27:00
SPEAKER_02: Again,
27:00
it's not access to open technology that is our primary choke
27:04
point and our primary
27:05
SPEAKER_02: intervention.
27:06
SPEAKER_02: It is how do we mitigate these residual risks right across the ecosystem,
27:11
SPEAKER_02: knowing that that technology is out there.
27:13
SPEAKER_02: We've done this with open software.
27:14
SPEAKER_02: We've done this with the open internet.
27:16
SPEAKER_02: And I think fundamentally the same is going to be true of AI.
27:20
SPEAKER_02: Now,
27:20
SPEAKER_02: when I say that,
27:20
SPEAKER_02: I don't want to diminish that there are some very acute,
27:23
SPEAKER_02: very concrete risks associated with this technology.
27:26
SPEAKER_02: And the uplift with AI and with visual AI is significant.
27:31
meaningful right it's significant over over the baseline in
27:34
some areas but again i
27:36
think we need to think in a more joined up way about how these
27:39
mitigations come
27:39
together right across the tech stack and to adam's point around
27:43
moderation if we
27:44
can make available the tools and resources that downstream
27:48
developers need to
27:49
implement this technology safely then that's an important way
27:52
of mitigating some of
27:53
SPEAKER_01: these risks as well so
27:55
SPEAKER_01: I think for lots of people who do a policy for a living,
27:59
SPEAKER_01: you spend a lot of your day reviewing various different policy proposals,
28:02
SPEAKER_01: lots of bills at the state level,
28:03
SPEAKER_01: lots of bills at the federal level.
28:05
SPEAKER_01: You guys are an international company, strong European presence.
28:08
SPEAKER_01: So you're reading bills internationally as well.
28:13
There's a lot of discussion about what the right disclosure
28:15
model is now for
28:16
SPEAKER_01: companies and for particularly for small companies.
28:18
It seems like just based on your behavior,
28:21
you're not saying no disclosure is the best course.
28:25
What's the delta between what you guys are doing now and what
28:29
then you see when
28:30
you're reading through disclosure mandate proposals from
28:33
various different
28:34
SPEAKER_01: governments?
28:37
SPEAKER_02: I'll be clear.
28:38
I think a lot of transparency proposals that I've seen are
28:41
actually trending in the
28:42
SPEAKER_02: right direction.
28:43
SPEAKER_02: I think the devil's in the detail, right?
28:46
SPEAKER_02: There's a lot of challenges and there are a lot of challenges that can
28:48
SPEAKER_02: disproportionately affect startups compared to big tech.
28:52
SPEAKER_02: As Adam says, we have a small compliance team.
28:54
SPEAKER_02: You go to Google or Meta,
28:55
SPEAKER_02: these compliance teams are much larger,
28:59
but I think the proposals that have been most challenging have
29:02
been on the
29:02
SPEAKER_02: mitigation side.
29:03
SPEAKER_02: So proposals that try to draw a line in the sand about acceptable risk.
29:07
SPEAKER_02: They don't just say disclose what valuations you've performed,
29:11
SPEAKER_02: but they try to actually stipulate what acceptable risk looks like.
29:15
And the challenge with that approach is that you create a world
29:18
where it's
29:18
SPEAKER_02: relatively easy for a closed source developer to comply.
29:22
You make it much more difficult for an open source or an open
29:25
way developer to comply.
29:26
You're not really comparing apples to apples,
29:28
you're comparing apples to oranges.
29:30
SPEAKER_02: And so most of the proposals that we've seen
29:33
SPEAKER_02: become challenging when they veer out of pure disclosure,
29:36
SPEAKER_02: pure transparency territory,
29:38
and they start moving towards either overt or implicit
29:41
restrictions on the release
29:43
SPEAKER_02: of the model itself.
29:44
SPEAKER_02: A good example of this is actually California.
29:47
SPEAKER_02: I think if you look at where SB 1047 was back in the day,
29:52
and you look at where SB 53 and some of these proposals are
29:54
going in the future,
29:56
SPEAKER_02: the trajectory is the right one.
29:57
which is don't stipulate secondary liability for model
30:01
developers in relation to
30:02
SPEAKER_02: some exotic and fuzzy ill-defined harms.
30:06
But there may be a well-designed targeted intervention focused
30:10
on disclosure of
30:12
SPEAKER_02: your risk assessments,
30:13
SPEAKER_02: disclosure of the evaluations that you've performed.
30:15
And without endorsing one bill or the other,
30:18
I think that is really the right direction.
30:20
SPEAKER_01: So that's helpful in terms of understanding like on substance where you
30:25
come down i'm curious as well just how you deal with the
30:29
jurisdictional complexity
30:30
so ben you mentioned california and i know you have written in
30:34
the past about 1047
30:35
and various different transparency models there um adam again
30:39
i'm like seized by
30:39
this four-person team image in my head that's a small team for
30:44
navigating like a
30:46
SPEAKER_01: compliance framework in europe a compliance framework
30:49
that's being set for the United States and Washington and then
30:54
navigating all
30:55
SPEAKER_01: various different state compliance frameworks,
30:58
SPEAKER_01: including California,
30:59
SPEAKER_01: which has passed a lot of AI legislation in the last couple of years.
31:02
SPEAKER_01: So how does that,
31:03
SPEAKER_01: like,
31:03
SPEAKER_01: even if you're seeing a positive direction of travel,
31:06
SPEAKER_01: how do you guys navigate that
31:09
SPEAKER_01: state, federal, international patchwork?
31:11
SPEAKER_03: I think maybe just addressing this from a more operational level,
31:14
SPEAKER_03: number one is hire really great people.
31:16
So I think we're very fortunate to have people like Ben and
31:20
others kind of join us
31:21
and agree to kind of tackle this crazy hodgepodge and emerging
31:26
AI regulations all
31:27
SPEAKER_03: at the international,
31:28
SPEAKER_03: federal,
31:29
SPEAKER_03: European,
31:30
SPEAKER_03: state level.
31:31
SPEAKER_03: That is quite complex and interlayered.
31:34
SPEAKER_03: And more specifically,
31:34
SPEAKER_03: though,
31:35
I think trying to figure out the trends in which you can take
31:39
one set of actions,
31:40
let's say disclose one certain way on the transparency side and
31:44
allow you to kind
31:46
of meet regulations across the board that oftentimes will have
31:50
disproportionate
31:50
SPEAKER_03: results as well.
31:51
SPEAKER_03: which will be great and something the team looks to do.
31:54
SPEAKER_03: I think,
31:55
SPEAKER_03: though,
31:55
there's probably a larger question of if there's a lot of
31:59
regulations that conflict
32:01
SPEAKER_03: with each other,
32:02
SPEAKER_03: what do you actually do?
32:04
SPEAKER_03: And, you know, when that comes to that, I think we'll
32:07
SPEAKER_03: take on those issues when they come up.
32:09
SPEAKER_03: But I think those are hard,
32:10
SPEAKER_03: hard questions,
32:11
which the team is well equipped to tackle given kind of the
32:14
experience we have.
32:15
I think as a small team,
32:16
it's really important to triage in a really disciplined way.
32:20
SPEAKER_02: We are not a big tech advocacy operation.
32:23
We need to be very careful about which battles we pick and
32:26
where we can have an impact.
32:28
SPEAKER_02: From our perspective,
32:30
we look at the world through the lens of how can we best
32:32
promote and protect open
32:33
SPEAKER_02: innovation and visual intelligence.
32:35
Fundamentally,
32:35
that means researchers must be able to share the technology
32:38
widely.
32:38
It means the developers must be able to build on the technology
32:40
responsibly.
32:41
SPEAKER_02: And so when we boil the ocean on
32:44
SPEAKER_02: 1,800 state and federal bills just this year alone.
32:47
SPEAKER_02: And we look at 700 pages of the AI Act during the trial process,
32:51
SPEAKER_02: for example,
32:51
SPEAKER_02: and 100 pages of codes of practice.
32:54
SPEAKER_02: We look at it through that lens.
32:56
SPEAKER_02: We say, where is this going to set back that mission?
32:59
SPEAKER_02: We'll talk with anyone anywhere,
33:00
SPEAKER_02: to be clear,
33:01
SPEAKER_02: in any jurisdiction,
33:03
but we're not interested in photo ops and roundtables and naval
33:07
gazing just for the
33:08
SPEAKER_02: sake of it.
33:08
Everything we do from an advocacy perspective on law reform has
33:12
to serve that
33:13
SPEAKER_02: overriding mission.
33:14
SPEAKER_02: And when you're that disciplined,
33:16
SPEAKER_02: You'll find a lot of work kind of falls away.
33:18
SPEAKER_02: There's a lot of busy work out there in the policy world.
33:21
And we're,
33:22
as I say,
33:22
ruthlessly focused on what really matters to researchers and
33:26
developers.
33:26
SPEAKER_03: Where can we have outsized impact?
33:28
That's really fundamentally one of the central questions we
33:31
always ask ourselves,
33:32
SPEAKER_03: whether it's a policy team,
33:34
SPEAKER_03: legal team,
33:35
SPEAKER_03: and a research team,
33:37
SPEAKER_03: and engineering team,
33:38
SPEAKER_03: and so on.
33:38
SPEAKER_01: A year ago, Adam, you didn't have a policy team.
33:40
And then you decided that you would start one and that you
33:43
bring Ben in to lead it.
33:46
Can you explain a little bit about that decision and why you
33:47
decide that this was
33:49
SPEAKER_01: the time that you needed to have a policy contract?
33:51
I think one of the central things too is when you find
33:54
exceptional talent like Ben
33:55
SPEAKER_03: is,
33:56
SPEAKER_03: then you kind of try to bring them on board no matter what they do.
33:58
I think also more practically,
34:01
AI regulations is a central part of any foundation model lab.
34:07
It will be something that any foundation model lab will have to
34:10
tackle throughout
34:11
SPEAKER_03: its journey from a startup
34:14
to a larger company,
34:15
or if at a big tech company,
34:18
the AI labs within them.
34:20
SPEAKER_03: And I think we recognize that early on,
34:22
SPEAKER_03: that this would be such a unique thing for BFL on its journey,
34:26
that kind of tackling the problem ahead of time would give us
34:30
some outsized
34:30
SPEAKER_03: returns.
34:31
SPEAKER_03: And I think we're seeing good results there too.
34:33
SPEAKER_03: I think engaging with
34:34
policymakers and regulators early on makes it so that we're not
34:39
just some strange
34:41
random lab operating out of the middle of Germany that no one
34:43
knows about and is
34:45
SPEAKER_03: looked at more suspiciously and more fearfully by regulators.
34:49
SPEAKER_03: Instead, there's somebody that regulators and others can interact with.
34:52
We can engage on these issues and we can try to work through
34:55
thoughtful solutions.
34:57
SPEAKER_01: Ben,
34:58
SPEAKER_01: as the founding member,
34:59
SPEAKER_01: the lead of this policy team,
35:02
how are you thinking about the function and potential growth of
35:05
your team?
35:05
I just don't think that the big tech playbook translates very
35:09
well into startups
35:11
SPEAKER_02: and potentially even into model development.
35:12
SPEAKER_02: I think we have to think differently about how we do advocacy.
35:17
SPEAKER_02: our growing team,
35:18
I think we need to be fundamentally committed to the idea that
35:20
meaningful public
35:21
SPEAKER_02: oversight is compatible with open innovation.
35:24
SPEAKER_02: Our job is to explain the upside of this technology to policymakers,
35:28
SPEAKER_02: help to mitigate the downside in concrete ways,
35:30
and help policymakers to reconcile the two through any more
35:34
legislative or
35:34
SPEAKER_02: regulatory reforms.
35:35
SPEAKER_02: From my perspective,
35:37
SPEAKER_02: with the number of jurisdictions we have,
35:39
SPEAKER_02: the outsized impact that we have as a business,
35:41
SPEAKER_02: I think it's really important that we,
35:43
SPEAKER_02: as I say,
35:43
SPEAKER_02: relentlessly focus on meaningful,
35:46
SPEAKER_02: lasting impact in the regulatory system.
35:49
SPEAKER_02: And we don't just create noise for its own sake.
35:51
Like I am interested in targeted reforms and helping
35:54
policymakers together.
35:55
I'm not interested in generating PDFs and having photo ops and
35:58
staging round tables
35:59
SPEAKER_02: just for the sake of it.
36:00
And I think if you're a small team,
36:02
it's really important to have that discipline.
36:04
SPEAKER_02: We don't want to have the kind of
36:07
hyper-specialization that I think you see in some of the larger
36:10
organizations.
36:11
We need to stay focused and we need to be very clear about what
36:14
matters to us and
36:15
SPEAKER_02: our team and our community and what maybe doesn't matter.
36:18
SPEAKER_03: And the focus as well, or the point of view that we bring
36:22
SPEAKER_03: as a small startup is somewhat unique,
36:25
SPEAKER_03: I think,
36:25
from the big tech and policy team and then with a few points of
36:29
view that that can
36:29
SPEAKER_03: bring.
36:30
And I think we can see some outsized influence there or some
36:33
disproportionate
36:34
SPEAKER_03: impact from making those voices heard.
36:36
We believe that it's a voice that can articulate kind of a
36:38
little tech point of view.
36:40
is important to ensuring that policymakers have a more robust
36:43
understanding of the
36:44
SPEAKER_03: entire market and what sort of regulations they may be contemplating,
36:48
SPEAKER_03: what impact that could have on innovation and competition.
36:51
SPEAKER_01: Lots of people speak about what that point of view is.
36:53
SPEAKER_01: We're a firm that does.
36:54
SPEAKER_01: We have a little tech policy agenda.
36:56
SPEAKER_01: We've talked about it extensively.
36:57
SPEAKER_01: How would each of you characterize what the little tech perspective is?
37:02
And another way to say that is kind of what's the competitive
37:04
advantage that you
37:06
SPEAKER_01: bring as a little tech player to the policy ecosystem?
37:08
You're going up against these policy teams that have hundreds
37:11
of employees,
37:12
SPEAKER_01: sometimes thousands of employees.
37:14
SPEAKER_01: What is the little tech voice that you're bringing to the table?
37:18
SPEAKER_02: Well,
37:18
I think the value of a small team punching above its weight is
37:21
that we have the
37:22
agility to take some really bold positions on critical issues
37:26
or major reforms in a
37:27
SPEAKER_02: way that might be more difficult for a larger team.
37:30
We can sit down more as a thought partner with policymakers and
37:33
work through the
37:33
SPEAKER_02: menu of options.
37:35
SPEAKER_02: I think one of the challenges with larger organizations,
37:38
especially those that have come out of the kind of traditional
37:41
internet platform
37:42
SPEAKER_02: space,
37:43
SPEAKER_02: is everything is so colored by that Section 230 view of the world.
37:47
SPEAKER_02: We need to maintain immunity on the one hand,
37:48
SPEAKER_02: but don't worry,
37:49
SPEAKER_02: we've got these sort of privatised regulators internally,
37:51
SPEAKER_02: these trust and safety teams who will take care of it for us.
37:55
SPEAKER_02: And I think that's important for all sorts of reasons,
37:57
but it's fundamentally a different model of advocacy and a
38:00
different set of risks
38:01
SPEAKER_02: and options than you have.
38:03
SPEAKER_02: in something like mold development.
38:06
SPEAKER_02: So I think being able to take a bold view,
38:08
SPEAKER_02: sit down with policymakers,
38:10
SPEAKER_02: have a considered dialogue about the menu of options.
38:13
I think that's where small teams can have as much and sometimes
38:16
more impact than
38:18
SPEAKER_02: large organizations.
38:19
SPEAKER_01: Of course, there are downsides too.
38:21
SPEAKER_01: Are there policy debates you've been in so far where you've thought,
38:24
SPEAKER_01: oh,
38:24
SPEAKER_01: we're significantly disadvantaged relative to larger players?
38:29
SPEAKER_02: I think larger players are sometimes...
38:32
perhaps blind to how these proposals impact small teams doing
38:37
frontier work,
38:38
SPEAKER_02: teams in our position.
38:39
You have proposals out there that are very straightforward for
38:43
a vertically
38:44
integrated provider and very difficult for a provider who just
38:47
forms one part of
38:48
SPEAKER_02: the tech stack.
38:49
You have proposals that are a relatively minor compliance
38:52
overhead for a big tech
38:54
firm and an absolute massive compliance ordeal for a small
38:58
business in our
38:58
SPEAKER_02: position.
38:59
SPEAKER_02: And as I've said, you've got
39:01
SPEAKER_02: proposals out there,
39:02
SPEAKER_02: but relatively straightforward for a closed source provider and very,
39:06
SPEAKER_02: very difficult for an open source or open weight developer.
39:09
And so I think sometimes larger organizations have so many
39:12
competing interests in
39:13
equities internally that they sometimes don't understand the
39:16
second and third order
39:17
SPEAKER_02: consequences of these proposals.
39:19
And as a result,
39:21
the effect of these proposals on small teams like ours gets
39:24
lost in the noise.
39:25
SPEAKER_02: So one thing I think has been promising is that a lot of policymakers,
39:29
SPEAKER_02: including many who we might disagree with on principle,
39:32
are realizing that there is a more diverse community of
39:35
developers out there than
39:37
SPEAKER_02: just the big Fang or Mango companies,
39:39
SPEAKER_02: and they're willing to
39:40
sit down and listen to these smaller teams and see how they can
39:43
adjust these
39:44
SPEAKER_02: proposals in more targeted,
39:45
SPEAKER_03: more precise ways.
39:46
SPEAKER_03: And overall,
39:47
we're very heartened by the fact that policymakers are aware
39:50
that they often listen
39:51
to a lot of policy discussions or have a lot of policy
39:55
discussions with big tech
39:57
policy teams and have less of them with smaller tech and
40:00
smaller companies just
40:01
SPEAKER_03: because of a pure resourcing issue.
40:04
And I think the fact that they're willing to talk to us and
40:06
really willing to take
40:08
kind of our point of view and a consideration means that these
40:11
discussions do,
40:11
SPEAKER_03: you know,
40:12
SPEAKER_03: they do have an audience and it does have an impact.
40:15
SPEAKER_03: Providing that point of view,
40:16
especially as to how some sort of policy could have an onerous
40:19
impact on much
40:20
SPEAKER_03: smaller teams,
40:21
SPEAKER_03: I think is quite critical too,
40:23
SPEAKER_03: right?
40:23
SPEAKER_03: We often saw and heard when GDPR was put into place,
40:27
SPEAKER_03: as well as when it was implemented,
40:28
that GDPR could actually kind of slow down big tech expansion
40:32
within Europe when
40:34
SPEAKER_03: really after it was implemented,
40:35
SPEAKER_03: the exact opposite was the case because of how many lawyers
40:38
SPEAKER_03: the big tech companies could throw at this.
40:40
Being able to articulate this point of view to policymakers,
40:43
I think that's really powerful.
40:45
SPEAKER_01: Ben, this is a great conversation.
40:46
SPEAKER_01: Thanks so much for coming on the AI Policy Brief.
40:48
SPEAKER_01: Thanks, Matt.
40:48
SPEAKER_01: Appreciate it.
40:50
SPEAKER_03: Matt, it was such a pleasure.
40:51
SPEAKER_03: Thank you so much.