How to sell RL envs and data to AI labs: Interview with Sean Cai

What areas of data do you feel are

underserved by now um that labs want to

buy?

>> Uh certainly we're still in a huge darth

of long and unverifiable. Uh the domains

that matured the quickest

matured because they were much much more

easily verifiable with web 2.0

instruments. Coding had GitHub. We don't

have a GitHub for all other domains but

we ventured into finance healthcare and

law afterwards. Nowadays, biological and

cyber security is is is a craze but only

I think by the most sophisticated labs

namely anthropic and some of OEI.

>> You said bio and cyber.

>> Yes.

>> But in general I think incredibly long

horizon realistic is still very much in

demand. A lot of benchmarks are built as

this but not actually that. Does that

mean things that would like basically

serve making co-work and codeex better

for non-technical work?

>> Uh potentially. Yeah. I would classify

that as like back office ERP type tasks.

>> Yeah.

>> Which largely have to do with very

complicated search and retrieval

functions across quite convoluted data

links uh and environments.

>> Okay. Um what kinds of software programs

would people be using?

>> Uh even so Excel file systems.

um just like think very convoluted file

systems and applications with data

across multiple formats sometimes

tabular sometimes graphical even as well

>> um and it's an exercise in model tool

calling as well

>> cool also stuff like SAP or or not

[snorts] necessarily

>> uh yes but I think for maybe some of the

computer use circles computer use

continues to be sort of like a smaller

market but dominated by a few top RLM

companies there uh like a a version of

SAP would have gone for like 500K uh in

the computer use craze of like late 25

probably. Um I suspect a version of SAP

has been created by one of the RL

computer use companies at this point or

maybe by an internal anthropic and OI

but that's speculation. Yeah. The um bio

and cyber for the bio stuff is it uh bio

real world things or like purely digital

you know stuff that's pure computer?

>> Yeah it started from bioinformatics but

nowadays we're trying to make a lot of

these processes with one step in the

digital and one step in the physical

areable.

Uh naturally this gets really hard

because you think about the verification

mechanism for a lot of things in

biology. Anthropic just put out this

benchmark called mystery biobench which

I think enumerates a lot of the problems

pretty succinctly there. We don't even

know even amongst the top experts how to

verify something in biology because

they're literally denovo experiments,

right? Like we combine certain chemicals

like what happens? Um

so uh they're almost veering into

physics based models. Of course if you

get to physics based models then you

have like

uh semreal and robotics uh to to which

you could do RL and robotics and that's

an entirely different domain. But I I'd

say it's like on our slow march to make

sim to real uh and and generally model

much more things in the physical world

more accurately as verifiers and RL um

>> certain domains

>> fall in the middle of purely software

based work and like robotics

>> like biological workflows, chemistry

workflows, even scientific discovery

that make that pretty useful. What are

some of the pieces of software that uh

bioinformaticists might be using that

like they would have you know ends for?

>> Oh, I think I think like I'm not a

biologist but uh there's so many bespoke

tools and like bespoke processes within

a lab itself. I I I'm helping a wet lab

sort of digitize our processes here

doing this and a lot of this stuff

doesn't have purpose-built software for

it. It's

and it does just adds to the environment

complexity and bespoke tool build out

for those environments complexities as

well.

>> Okay. So it's more like general computer

use both both guey and bash.

>> Yeah, I would I I would say so that's

the primitive in which like kind of

everything is based right.

>> Yeah. the um

cyber stuff any particular subsets like

uh uh you know program analysis or um

pentesting or um you know what's the

what's the most underserved cyber

subsets do you think?

Yeah, for sure. Uh, I think cyber is

mostly being bought by Anthropic and

maybe some of OEI right now and then the

rest of the labs are are following

whatever they do. Uh, naturally

irregular security is probably the

company to fall into space who does a

lot of this type of stuff. Um,

I would say a lot of offensive cyber is

uh being modeled right now in sort of

interactive environments. So a lot of

stuff in code level security uh web app

exploitation CTF challenges and like

remediation is already covered by

existing benchmarks like um cyber gym

cybench these are pretty saturated

nowadays so long horizon wise as all

domains are tending towards you're

looking towards um

stuff like infrastructure exploits uh

and agent layer attacks. So there

there's a lot of stuff you can model out

there. Um because there are new zero

days every single day. It's like one of

the most dynamically changing fields. So

naturally you're going to expect there

to need to be uh real-time data streams

to translate these things into model

actionable formats.

>> Yep. Awesome. The um

the is the the sales process is roughly

um either do something in public such

that researchers reach out to you

already know researchers or kind of get

intros or do cold emails to researchers

to get a you know a pilot. Is that

roughly the first step?

Yeah, but you know the our our appetite

for data is voracious and expanding, but

there's still only 24 hours in a day and

a researcher's job is not to talk to

data vendors all the time. So the bar to

get I think a researcher's attention is

getting higher and higher. So those

without research sophistication, it's

just like that's simply not the best

move anymore. the human data supply

chain is expanding such that one can

meaningfully participate in it in it um

without interacting with an end

researcher

>> cross-selling partnerships you mean or

>> yeah protege is an example uh I think

there are companies out there which

almost train companies to produce good

post training data and then sell their

data to researchers themselves using

themselves as a stamp of approval

Yep. Yep. The um

the uh what are what are the most

important things that labs look for?

What are the main reasons that a lab

might uh not kind of renew or increase

their purchase volume? You know, after

they get the uh the first set of data.

>> Uh each lab has their own QC processes

and they run them internally and they

see. But I I you know there there are so

many reasons why

some data might be quite quite poor. Um

so

I think there are just a lot of very

small things that a researcher can look

at a data set and like maybe a data set

is a million lines and they notice

something that is off about one line and

they really start to question whether a

startup even has a QC process at all in

the first place or not. Um,

subsequently,

uh, it's it's it's I think the most

common case is just your tasks are

incredibly incredibly poorly designed,

um, in terms of they're easily reward

hackable. The props are vague, they're

not emblematic of real world tasks from

an obvious setting.

There's a lot of other smaller things

afterwards too like the model failures

you identify and the reasons for why the

model fails at them are not actually

genuine capability failures. So because

you designed the harness quite poorly.

Um it's because you ran it in a very

specific environment that isn't actually

how most users would have ran this task

or model in. Uh you don't do cross

harness testing. You don't do engram

contamination testing. So which is to

say you don't test whether a data set is

already in the pre-trained corpus

literature or not. Um there there's a

lot of uh QC checks that one can run

before delivering say OTS RL data that

would make it just uh that would make

DOM a great partner that researchers

would want to work with and iterating

the shape of posting data.

>> Yeah. the um

the and the QC is both uh researcher

flagged issues and then also just the

data teams as well reviewing things

themselves.

Yeah, I think it's uh

c certainly researcher needs are bespoke

as well for certain projects, but if you

think about

uh researchers exploring a net new

research question like how do we improve

taste and models and they're exploring

different OTS RL data sets to uh along

with maybe a data company provided

benchmark to exploit this question.

There there are just so many things that

they look for in a typical data delivery

that are really is really difficult for

you to know unless you worked in the in

the data industry yourself or you've

been a researcher and you understand

what good RL data is, right?

>> Good data taste, good research taste,

and perceived ability to scale quality

with quantity are the three things that

are necessary to building a a good human

data company out.

>> Good data taste, good research taste,

and what was the third one?

perceived ability to scale quality with

quantity.

>> Yeah. Cool. The um uh when does a

researcher like in a question like that

they're doing some new kind of area um

that's vague. How do they get the model

to do a certain thing?

Um

what do researchers do to first explore

off-the-shelf offerings before um will

they just hit up all their existing

vendors from their team or they go to

their you know data team people

internally and ask them to go and get a

a uh a set of options for them? What's

the what's the first steps that the

researcher takes?

Yeah, I'd say they do some bespoke reach

out themselves, especially for new

research team directions like OpenAI's

newest robotics VA direction which spun

out of Sora or not I wouldn't say Denovo

spun out of Sora but was sort of

combined with the remnants of Sora. they

will go out to real world data vendors

and reach out to get samples uh because

you got to remember their jobs are to

improve model capabilities and if that's

the bottleneck they won't go and solve

the bottleneck themselves but that's why

these labs have human data teams. It's

like one to procure the data necessary

and manage vendor relations but two is

like negotiate on price

>> and and all that all those things

associated. So it's a collaboration

between those two entities.

>> Cool. The um in the robotics data space

is there anything that your viewers

underserved

uh there? Obviously there's a lot of

things that are well served but what are

the ones you view as underserved in

robotics?

>> Data vendors who are genuinely research

first and running post training

experiments on their own data. if

they're trying to sell things like ego

and data up the training mix pyramid to

if you like companies

uh or

uh ju just running a lot of training

experiments on their data that match the

research direction of companies they're

trying to sell to in order to be a bit

more

>> selling egocentric data to VA companies

is underserved.

It is underserved in a sense of there

are not many

>> my naive view is like everyone is

selling egocentric data.

>> Well, there are very few egocentric data

vendors that actually know how the

downstream training is done.

>> So it's egocentric data vendors who are

doing their own postraining and

therefore like research have good

research taste.

>> Yeah. But um you you got to remember

what is being sold when you sell data in

the first place is just model capability

improvement, right? And data is just the

medium to do that. So if you're trying

to sell data, but you're not actually

cognizant of how model improvement is

achieved or you don't have an opinion

there and can't really help the

researcher with that, you are in a

losing battle and losing market uh and

you are going to be commoditized.

>> Yep.

>> Yep. The um

what is the uh what does the initial

meeting look like? the um someone talks

to researcher, the researcher maybe

requests some samples and the founder

sends it in Google Drive. Um how does

that uh what does that typically look

like? What's the formats people are

expecting?

>> For RL data, for the longest time, it's

literally just been a Docker container.

>> Yeah.

>> Uh a Docker container isolated

environment, all the tools on there, all

the verification mechanisms and rubrics.

One simply simply has to plug and play

their agent. Uh and then you get an eval

score and then you can use a multitude

of these software containers to run

rollouts for GRPO RL.

>> Yep.

>> Whatever other training me mechanisms

you employ

>> and labs do they have more kind of

sophisticated internal kind of you know

uh

setups for um running environments now

that need different formats.

>> Yes, they do. Anthropic notably has one

whose name I can't uh disclose but the

most sophisticated labs I would say are

like Anthropic, Open AI, Deep Mind and

then everybody else and then Chinese

labs in that order.

>> Yep. Yep. the um

the um and then in terms of the data

that the non

you know three frontier labs are buying

um are they buying more off-the-shelf

data that you know companies have

already sold to anthropic and open AI on

like a non-exclusive thing

>> uh I think OTS is a relatively new

phenomenon it is the mechanism with

which Serge has done business for a long

long time. Uh but that's because Serge

is a very fundamentally different

company than all the other ventureback

players. Um I I believe

>> how are by the way on that?

>> Oh, they genuinely started off as just

model capability caring about model

capability improvement, right? Not not

as a sort of data company and for the

longest time like mostly SFD data as

well. Um,

>> so

you're talking about exclusivity.

Certainly exclusivity reflects different

labs philosophies towards data vendors.

Anthropic is the only one who I think

really pushes for exclusivity. Open AI

at different points throughout its human

data turnover uh human data teams

turnovers because a lot of people shift

around in OAI a lot. Uh but Enthropic

genuinely views their data vendors as

research partners and if you think your

research partner is genuinely novel

research you probably want to get

exclusivity on that. Um

>> which is the approach that they've

employed with many of the data companies

they've worked with.

>> Yep. Do they have expiry clauses on the

exclusivity like 12 months or 24 or

something like this? I'd imagine they

are starting to think about that pretty

closely now. But I am aware of many many

companies who have recently just ended

anthropic exclusivity. Some of them had

the agreement that they would only have

it for a year. Some of them

for other strategic reasons they've

stopped exclusivity with. So

>> yeah.

>> Yep. the um

the

uh

almost all purchase decisions researcher

led at this point as opposed to you know

like the researcher pulls it in and then

the the data team kind of uh is effect

is like a form of pro procurement or um

is it different? You can imagine it's a

partnership of sorts,

but if you want to think about it from

from a economic buyer perspective, you

always want to be just in general B2B

sales dealing with the economic buyer

because if you can convince the economic

sorry, not the economic buyer, the the

end user, right? If you can convince the

end user of your product that there's

substantial value, the question is not

whether the org is going to buy it or

not. It's just how much are they going

to buy it for. Yep.

>> Um, and so if you're going to the guy

who's pricing it first, who doesn't know

how available it is, doubtlessly it's

going to be a harder sell than if you

had convinced the end user that it's

available first, right?

>> Um, Decagon, Sierra, and Ramp. Um, what

kinds of uh data are they buying

relative to the Frontier Labs?

>> Voice data. Uh, RAMP is not so much

buying data. Actually, the Ramp Labs

report came out um the other day, and I

was surprised at a couple things. One, I

really love the fact that you've got

really sophisticated elite engineering

or applier companies out there

post-raining their own small models for

their own use cases. But I was surprised

that they used a synthetic data set to

inform some of the environments in which

they were training like accounting level

transactions if they're an app layer

company that should have access to that

data themselves

>> which is

which one suggests that one could

sell data to them if they can't use

their own app layer data uh for for

these training environments. But uh two

uh also suggests that Apple companies

may be feasible buyers in the future if

there's a substantial systematic issue

that prevents them from using their own

users data. Certainly doesn't look like

it's been a problem with cursor though.

So I'm sure this is just a small uh

quirk.

>> Yeah. Do you think it's a it's a privacy

thing that they'll just figure out?

>> I think so. Privacy is not like data

privacy is very easy to figure out

nowadays for all these companies.

>> Yep. the um

uh do you have a certain view on you

know long-term

um

the

labs have their applications those

applications give them you know traces

that they can train on um

how it evolves where they still need to

buy data externally versus training on

the data from their users

Um

yeah, one would have thought that

Enthropic has so much data from claude

code and

work

>> right

>> that maybe they would not have needed to

procure from external vendors

but they still do. Um and and and this

reflects the fact that most external

data vendors that are succeeding with

sophisticated research labs and data

markets, they're mostly selling

capabilities that are N plus one of

current tier models, right?

>> Y

>> um Andon Labs, by the way, Andon is a

fantastic company in this regard in

terms of producing really hard realistic

benchmarks, but a bit too ahead of its

time, I think.

>> Yeah. Um, and on labs is a good example

of the fact that we're we we're going to

produce these really real world long

horizon benchmarks that are not going to

be saturated for a long time and that is

quite available to us.

>> Yep. So if it's already within the

capabilities of the model then they can

train on it from their traces but if

it's not and no user is going to attempt

it in the model then they don't have any

traces to train on. And this is from a

purely single axis performance-based

perspective, right? Whereas it's like

there's only one thing to help climb and

it's is perceived performance. Um cost

and latency are also big questions too.

An anthropic researcher I think told me

at some point our benchmarks are really

not going to index on performance and

that we'll have prohibitively expensive

AGI in some sense but like how much does

it cost and how fast does it take to do

something is going to be new

>> new new dimensions of benchmarks. So

then you expect that um end vendors will

uh start to do benchmarks that are

basically performance divided by price

rather than just performance essentially

>> perhaps. Yeah. And this expands greatly

the aperture of different niches that RN

companies can play in because if you

think about the enterprise world, right?

There are many use cases where I just

want a much much cheaper model at a

fixed level of intelligence

>> that is satisfactory for certain like

job functions, right? And then even in

ramp lab's recent implementation on

their Twitter post they showed that they

use a above head frontier model for

planning but they they collapse the

search and retrieval function to a small

model that they post trained just for

that purpose.

>> The um how many labs are spending at the

you know billion dollar plus per year

data level?

>> Seven or eight.

>> Mhm. the how much more than

like you know Anthropic talked about

their billion dollar number. Do you

think it's going to end up being like

closer to like you know three to four

kind of this year?

>> Yeah. I mean I'd say like honestly each

Frontier Lab if you're loose with your

definition of data like they spend

between 10 to 20 billion a year. I think

I posted about this a while back too.

>> 10 to 20 if you're loose with your

definition of data. Um yeah. Can you say

how so? Uh this is this shouldn't be a

surprise to anybody, right? Like three

things Hill climb model capabilities,

compute, data, and talent. And data

spend is still a drop in a bucket

compared to compute costs, right? Um I

I'd say we're generally still supply

constrained in that if you think about

RL data or just data in general, that

means the quality bar for these labs,

we're still very much still in demand of

that data.

>> Yep. So you're saying 10 to 20 billion

in aggregate?

No, per lab.

>> Per lab

with eight labs spending that much

>> uh sevenish.

uh I think for some labs

>> including this isn't like salaries of

data team people is included like how

does it get to

>> like literal data from external vendors

and and and by the way most of this

spent does not actually get satisfied

like I'm sure that there is a data

budget set aside whose upper limit is

not actually met because there's simp

just simply not enough good quality data

vendings. I've still seen I have still

never seen a data contract get turned

down by a top lab if it's good quality

data for budget reasons.

>> Yeah. What's the delta between the

billion dollar number versus the you

know 10 to 20 billion like what's

included in the latter that's not

included in the former?

>> Uh I would say body shop type data

labeling that's very emblematic of scale

type what what scale used to do uh and

what many people still think the data

industry is which is just manual manual

data labeling for pre-training data. Um

>> so then that would be like you know 70

billion plus in aggregate. Um

what's what's what's like the ballpark

of like surge scale

annual revenue

>> surge is between two to three bill

runway rate I'm pretty sure

>> what's the yeah where's the where's the

gap come from like if Serge is you know

leading provider they're doing two to

three 70 billion aggregate spend

>> um there are so many companies that

participate in data markets that you

would have never even expected just a

big massive long tail basically.

>> Yeah, it's an it's a very massive long

tail. Yes. Uh also

>> yeah staffing agencies as well. It's

like uh uh and this encompasses a lot of

the spend that OEI and anthropic

directly have like acquiring companies

from the real world too just for data

assets.

>> Yeah.

>> Uh which certainly I don't know why like

is happening a lot more and more and

people are not discussing this very

closely. M

>> um

>> this is like acquiring little like

little wet labs and that kind of stuff

>> like app layer companies in certain

domains that they're they're interested

in building products in right

>> uh I I I can't name them specifically.

Um

>> enterprise software type small app

player companies.

>> Yeah. Yeah, you could say that with like

network effects from like having I don't

know 10 to 15 years worth of user

activity like a stack overflow type type

thing. Uh so the the data markets as

exemplified by like Merur and these

companies they represent like the tip of

the iceberg in terms of like the the

entire long tale of companies where data

procured actually comes from.

>> Cool. As a last question, um the

what makes inference providers and

neoclouds a good fit uh to acquire RLM

codes is that they basically act as

implementers to the enterprise partners

that are their customers.

>> They they are the compute they are the

compute providers for our labs as well.

It naturally makes sense that they want

to do horizontal product expansion and

bring post-training infrastructure.

>> Yeah.

>> Uh and tooling alongside their product

offering to labs. B 10 actually I think

it was B 10 made an RLM's acquisition

>> uh like in December January time that

very few people are talking about so

there's precedent and I think um some of

the sophisticated RLM targets are very

good acquisition targets for this

>> both help themselves to enterprises and

to labs

>> uh to build out their uh post training

infrastructure product suite

>> the the and the end customer the post

training infra is um mostly like non-top

three frontier labs just like other

enterprises.

>> Yeah. Yeah. Like app layer companies

too. Y

>> um like for a while while you know

Perplexity and Cursor were more than 50%

of fireworks revenue for example.