正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

SITUATION EXPLAINED: How much are frontier labs actually spending on training data? .Sean Cai: "Frontier labs are spending about $10 to $15 billion per lab on data." "Really good long horizon tasks go up to $20,000 each. A complete browser-use version of SAP was rumored at $500,000." "Despite everybody... show more

MTS

569,649 subscribers

336,216 次观看 • 5 个月前 •via X (Twitter)

科学技术财经

Anya Rossi• Live Now

Private livecam show

23 条评论

tz 🏴‍☠️ 的头像

tz 🏴‍☠️5 个月前

@SeanZCai I remember when BigQuery queries would cost $2k per run and people flipped out

Pit Schultz 的头像

Pit Schultz5 个月前

@SeanZCai Ant calls systematic distillation of its Claude outputs a serious ToS violation & near-illegal threat, while buying or benefiting from third-party brokers selling $500k SAP browser-use RL datasets made the exact same way. Self-serving legal hypocrisy + ethical contradiction.

haro 的头像

haro5 个月前

@SeanZCai $500k for a single SAP browser task is a number that reframes the whole "data is cheap" narrative pretty fast. the supply side being genuinely undersupplied at these prices is the actual story here.

🃏 的头像

🃏5 个月前

@SeanZCai this seems way too high. like 10-100x too high?

mista! 受神庇佑的头像

mista! 受神庇佑5 个月前

@SeanZCai "... everybody thinking the market is super crowded, we still don't have enough good quality data vendors that actually understand how to deliver product plus services in a way researchers are looking for." wouldn't be a problem if these mfs were not gatekeeping their pipelines

toni 的头像

toni5 个月前

@SeanZCai Fair point on the scale, but most of that spend is on human feedback and labelers, not buying raw data. The $20k tasks show it - you're paying skilled people for complex work, not licensing datasets.

vibe yoda 的头像

vibe yoda5 个月前

@SeanZCai @SeanZCai is full of BS

arti 的头像

arti5 个月前

@SeanZCai This dude doesn’t come off as someone who knows what he’s talking about here but I enjoy listening to him

Przemek Chojecki | PC 的头像

Przemek Chojecki | PC5 个月前

@SeanZCai time to double down on - we have very unique long-horizon trajectories coming from research-level mathematics.

Chuck Petras 的头像

Chuck Petras5 个月前

@SeanZCai @BrianRoemmele

OliverAItools 的头像

OliverAItools5 个月前

@SeanZCai finally someone admitting quality beats cheap at scale this is the only way we get models that actually think

MagicApe 的头像

MagicApe5 个月前

@SeanZCai this dude doesn’t come off as someone who knows what he’s talking about here but I enjoy listening to him

Jason傑森 🇭🇰 | 🛠️ 的头像

Jason傑森 🇭🇰 | 🛠️5 个月前

@SeanZCai 100亿还能人嫌少啊。

Beto Borbolla 的头像

Beto Borbolla5 个月前

@SeanZCai Related:

Duoduo 的头像

Duoduo5 个月前

@SeanZCai the worst datasets i've ever seen all looked clean on the surface. then you actually try to use them and find six weeks of regex hell waiting. that's the part nobody buys

Henry Zhang 的头像

Henry Zhang5 个月前

@SeanZCai a browser-use sap clone is the right kind of ugly example. did anyone say how they keep those tasks from going stale?

That AI Guy 的头像

That AI Guy5 个月前

@SeanZCai

@ChuckBaggett Chuck Baggett Peace Freedom Ad Astra 的头像

@ChuckBaggett Chuck Baggett Peace Freedom Ad Astra5 个月前

@SeanZCai What does @schisofrenia have to do with the singularity , AI, robotics, et cetera?

thecolorswag 的头像

thecolorswag5 个月前

@SeanZCai White monster spotted

SEVEN 的头像

SEVEN5 个月前

@SeanZCai Crazy numbers

Mathis Dupuy 的头像

Mathis Dupuy5 个月前

@SeanZCai Seems bullshit. What would you with just one session ? You would need thousands of hours with the current paradigm

AI Mastery Guide 的头像

AI Mastery Guide5 个月前

@SeanZCai $500k for a single browser-use task dataset is wild. The data economy is massive

tau 的头像

tau5 个月前

@SeanZCai lol

相关视频

Yesterday I interviewed Sean Cai about AI data. This is essentially a guide for founders on how to sell data and RL envs to AI labs. "I've never seen a data contract get turned down by a top lab, if it's good quality data, for budget reasons." 00:00 What areas of data are underserved? 02:10 For bio data, is it real-world or purely digital? 04:21 For cyber data, which subsets are most underserved? 05:50 What is the sales process like? 07:04 Why would a lab not renew or increase their purchase volume? 10:13 When a researcher is exploring a new direction, what's the first step? 11:35 In robotics data, what do you view as underserved? 13:12 What does the initial data delivery look like, what format? 13:53 Do labs have more sophisticated internal setups for running environments? 14:32 Are the non-frontier labs buying off-the-shelf data from Anthropic / OpenAI vendors? 16:11 Do Anthropic data vendors put expiry timeframes on the exclusivity? 16:42 Are purchase decisions researcher-led? 17:41 Decagon, Sierra, Ramp: what kinds of data are they buying? 19:06 Long-term, when do labs still need to buy external data vs train on user traces? 21:15 Will end-vendor benchmarks shift to performance per dollar? 22:04 How many labs are spending at the 1B+/yr data level? 23:53 Delta between Anthropic's stated $1B and your 10-20B/lab number? 26:05 What makes inference providers / neoclouds a good fit to acquire RL env cos?

Yesterday I interviewed Sean Cai about AI data. This is essentially a guide for founders on how to sell data and RL envs to AI labs. "I've never seen a data contract get turned down by a top lab, if it's good quality data, for budget reasons." 00:00 What areas of data are underserved? 02:10 For bio data, is it real-world or purely digital? 04:21 For cyber data, which subsets are most underserved? 05:50 What is the sales process like? 07:04 Why would a lab not renew or increase their purchase volume? 10:13 When a researcher is exploring a new direction, what's the first step? 11:35 In robotics data, what do you view as underserved? 13:12 What does the initial data delivery look like, what format? 13:53 Do labs have more sophisticated internal setups for running environments? 14:32 Are the non-frontier labs buying off-the-shelf data from Anthropic / OpenAI vendors? 16:11 Do Anthropic data vendors put expiry timeframes on the exclusivity? 16:42 Are purchase decisions researcher-led? 17:41 Decagon, Sierra, Ramp: what kinds of data are they buying? 19:06 Long-term, when do labs still need to buy external data vs train on user traces? 21:15 Will end-vendor benchmarks shift to performance per dollar? 22:04 How many labs are spending at the 1B+/yr data level? 23:53 Delta between Anthropic's stated $1B and your 10-20B/lab number? 26:05 What makes inference providers / neoclouds a good fit to acquire RL env cos?

Chris Barber

132,784 次观看 • 6 个月前

Today we're launching Hebbian Robotics (YC S26) Hebbian Robotics . We build APIs for evaluating data quality, without training a robotics model. Data vendors use our API to search and analyze Physical AI data at scale, getting on-demand quality signals and metrics without managing infrastructure. High-quality data is the bottleneck to the next frontier of intelligence in robotics. Buyers want data they can trust. Data providers are focused on operational excellence. They don't have the bandwidth to understand the quality of their data at scale. We're fixing this. Vendors use our APIs to understand the impact of their decisions on their data quality: - Buyers demanding per-collector hour limits? Prove compliance while we stay ahead of the workarounds. - Changed your collection SOPs? Monitor for quality drift and regression. - Suspect duplicate samples? Verify with a single API call. Before Hebbian, I trained AI models for industrial cleaning robots, and Kingston Kuan spent years building large-scale infrastructure at Jane Street.

Today we're launching Hebbian Robotics (YC S26) Hebbian Robotics . We build APIs for evaluating data quality, without training a robotics model. Data vendors use our API to search and analyze Physical AI data at scale, getting on-demand quality signals and metrics without managing infrastructure. High-quality data is the bottleneck to the next frontier of intelligence in robotics. Buyers want data they can trust. Data providers are focused on operational excellence. They don't have the bandwidth to understand the quality of their data at scale. We're fixing this. Vendors use our APIs to understand the impact of their decisions on their data quality: - Buyers demanding per-collector hour limits? Prove compliance while we stay ahead of the workarounds. - Changed your collection SOPs? Monitor for quality drift and regression. - Suspect duplicate samples? Verify with a single API call. Before Hebbian, I trained AI models for industrial cleaning robots, and Kingston Kuan spent years building large-scale infrastructure at Jane Street.

Brandon Ong

166,971 次观看 • 3 个月前

So I have a question. If this is indeed data center equipment, as it appears to me, are they paying for the police escorts? How much do we know about these data centers, and who thinks they are a good thing?

So I have a question. If this is indeed data center equipment, as it appears to me, are they paying for the police escorts? How much do we know about these data centers, and who thinks they are a good thing?

Suzee Q

53,113 次观看 • 2 个月前

In 2024, Stanford's Yann Dubois gave a 104-min masterclass on how frontier LLMs are actually built. 99% of people still don't know this. His frameworks: - Data beats model cleverness - Perplexity isn't product quality - Alignment changes behavior, not IQ 12 insights: 🧵

In 2024, Stanford's Yann Dubois gave a 104-min masterclass on how frontier LLMs are actually built. 99% of people still don't know this. His frameworks: - Data beats model cleverness - Perplexity isn't product quality - Alignment changes behavior, not IQ 12 insights: 🧵

andrew engler

67,145 次观看 • 6 个月前

.Sigil Wen reveals testing frontier AI's cyber capabilities scared him enough to start moving his personal data off data centers and onto his own devices: "There is going to be a radical reformation of the economy as businesses and the economy shift to serving agents as one of its newest users on the internet, especially as humans trust AIs more to act on their behalf." "We were working on agents that could run on long horizon harnesses, and we were also testing different frontier models, and it was just really scarily apparent just how good at cyber they were." "It was really clear to me that, oh my God, Mythos class models are open source now. I think that really led to me wanting to build Underdog for myself." "I wanna bring my data onto my devices, off of data centers, which are some of the biggest targets." Conway Research

.Sigil Wen reveals testing frontier AI's cyber capabilities scared him enough to start moving his personal data off data centers and onto his own devices: "There is going to be a radical reformation of the economy as businesses and the economy shift to serving agents as one of its newest users on the internet, especially as humans trust AIs more to act on their behalf." "We were working on agents that could run on long horizon harnesses, and we were also testing different frontier models, and it was just really scarily apparent just how good at cyber they were." "It was really clear to me that, oh my God, Mythos class models are open source now. I think that really led to me wanting to build Underdog for myself." "I wanna bring my data onto my devices, off of data centers, which are some of the biggest targets." Conway Research

MTS

13,842 次观看 • 1 个月前

Traverse (Traverse) is a data research lab solving the hardest problem in AI training: subjective, taste-dependent work. They generate long-horizon training data for frontier models by observing and capturing what specialists actually do on the job. Congrats on the launch, Lance Yan and Zachary Yu!

Traverse (Traverse) is a data research lab solving the hardest problem in AI training: subjective, taste-dependent work. They generate long-horizon training data for frontier models by observing and capturing what specialists actually do on the job. Congrats on the launch, Lance Yan and Zachary Yu!

Y Combinator

78,385 次观看 • 8 个月前

#WATCH | Chicago, US: Chairman of Indian Overseas Congress, Sam Pitroda says, "...This is a policy issue. Congress party would frame a policy through which the wealth distribution would be better...We don't have a minimum wage (in India)...If we come up with a minimum wage in the country saying you must pay so much money to the poor, that's the distribution of wealth. Today, rich people don't pay their peons, servants, and home help enough but they spend that money on vacation in Dubai and London...When you talk about the distribution of wealth, it is not that you sit on a chair and say I have this much money and I'll be distributing it to everybody. It's naive to think like that. The PM of a country thinks like that, that means I got some concerns about his brain...You are really dealing with the policy issues to redistribute wealth and when you ask for the data, you are really trying to understand what is the distribution today. We don't have good data on all this. I think we need data to decide on policy issues. We don't need data to distribute wealth. We need data to decide on policy issues going forward..."

#WATCH | Chicago, US: Chairman of Indian Overseas Congress, Sam Pitroda says, "...This is a policy issue. Congress party would frame a policy through which the wealth distribution would be better...We don't have a minimum wage (in India)...If we come up with a minimum wage in the country saying you must pay so much money to the poor, that's the distribution of wealth. Today, rich people don't pay their peons, servants, and home help enough but they spend that money on vacation in Dubai and London...When you talk about the distribution of wealth, it is not that you sit on a chair and say I have this much money and I'll be distributing it to everybody. It's naive to think like that. The PM of a country thinks like that, that means I got some concerns about his brain...You are really dealing with the policy issues to redistribute wealth and when you ask for the data, you are really trying to understand what is the distribution today. We don't have good data on all this. I think we need data to decide on policy issues. We don't need data to distribute wealth. We need data to decide on policy issues going forward..."

ANI

244,501 次观看 • 2 年前

Andrej Karpathy had already talked about this possibility 1 month back, that he might join one of the big AI labs and why he may do so. "They (the AI labs) are kind of like at the edge of capability of what's possible. And they are working on what's coming down the line. And I think if you're outside of that frontier lab, your judgment fundamentally will start to drift because you're not part of what's coming down the line. And so I feel like my judgment will inevitably start to drift as well. And I won't actually have an understanding of how these systems actually work under the hood. It's a big system. I won't have a good understanding of how to develop and etc. And so I do think that in that sense, I agree and something I'm nervous about. I think it's worth basically being in touch with what's actually happening and actually being in the frontier lab. And if some of the frontier labs would have me come for some amount of time and do really good work for them, then I think that's maybe a good setup" --- From 'No Priors: AI, Machine Learning, Tech, & Startups" YT channel (link in comment)

Andrej Karpathy had already talked about this possibility 1 month back, that he might join one of the big AI labs and why he may do so. "They (the AI labs) are kind of like at the edge of capability of what's possible. And they are working on what's coming down the line. And I think if you're outside of that frontier lab, your judgment fundamentally will start to drift because you're not part of what's coming down the line. And so I feel like my judgment will inevitably start to drift as well. And I won't actually have an understanding of how these systems actually work under the hood. It's a big system. I won't have a good understanding of how to develop and etc. And so I do think that in that sense, I agree and something I'm nervous about. I think it's worth basically being in touch with what's actually happening and actually being in the frontier lab. And if some of the frontier labs would have me come for some amount of time and do really good work for them, then I think that's maybe a good setup" --- From 'No Priors: AI, Machine Learning, Tech, & Startups" YT channel (link in comment)

Rohan Paul

57,675 次观看 • 5 个月前

Grass is betting that better AI still starts with better pre-training data “Whatever you put into your model ultimately determines what your model's capabilities are” “Several of the top frontier AI labs in the world are working directly with Grass to acquire large amounts of pre-training data” “If that information is not present in your model, your model doesn't know how to do that thing in the first place” Andrej puts Grass close to one of the most fundamental inputs in AI: the training corpus itself. As frontier models scale, they need enormous amounts of diverse, high-quality web data. They are building the infrastructure to collect and supply that data at internet scale. Model capability starts with what the model gets to learn from.

Grass is betting that better AI still starts with better pre-training data “Whatever you put into your model ultimately determines what your model's capabilities are” “Several of the top frontier AI labs in the world are working directly with Grass to acquire large amounts of pre-training data” “If that information is not present in your model, your model doesn't know how to do that thing in the first place” Andrej puts Grass close to one of the most fundamental inputs in AI: the training corpus itself. As frontier models scale, they need enormous amounts of diverse, high-quality web data. They are building the infrastructure to collect and supply that data at internet scale. Model capability starts with what the model gets to learn from.

Kaiz 🐂🀄️

13,989 次观看 • 1 个月前

Introducing Contra Labs. The first frontier data and evaluation lab for Creative AI.

Introducing Contra Labs. The first frontier data and evaluation lab for Creative AI.

ben

451,450 次观看 • 7 个月前

I gave a talk at ICLR 2026 about how we are scaling RL on frontier LLMs with 1T+ parameters, on experimental data from our physical lab at Periodic! Here's a rough recording of the talk:

I gave a talk at ICLR 2026 about how we are scaling RL on frontier LLMs with 1T+ parameters, on experimental data from our physical lab at Periodic! Here's a rough recording of the talk:

Rishabh Agarwal

238,918 次观看 • 6 个月前

Are Data Centers Worth the Noise to Local Communities? "Multiple cities of people have filed lawsuits and complaints against these data centers because of the noise." "The first question I ask myself is, are we gonna need data centers in America? Yes, if we wanna compete in the modern world, in the modern economy, we are going to need them. They are good for society and good for the economy."

Are Data Centers Worth the Noise to Local Communities? "Multiple cities of people have filed lawsuits and complaints against these data centers because of the noise." "The first question I ask myself is, are we gonna need data centers in America? Yes, if we wanna compete in the modern world, in the modern economy, we are going to need them. They are good for society and good for the economy."

PBD Podcast

32,573 次观看 • 6 个月前

Intelligence is getting cheaper every quarter, and the frontier labs pay for that. The companies that win are the ones that own data no one else has: every cheaper model makes their data more valuable and their costs lower. Own the data, not the frontier.

Intelligence is getting cheaper every quarter, and the frontier labs pay for that. The companies that win are the ones that own data no one else has: every cheaper model makes their data more valuable and their costs lower. Own the data, not the frontier.

Antonio Linares

21,378 次观看 • 11 天前

Lori Chavez-DeRemer on jobs data: "Sometimes it's too late to get accurate data points ... we don't have that accurate data for the Fed to make a good decision, but I believe the Fed needs to cut interest rates again in December to really prop up this economy."

Lori Chavez-DeRemer on jobs data: "Sometimes it's too late to get accurate data points ... we don't have that accurate data for the Fed to make a good decision, but I believe the Fed needs to cut interest rates again in December to really prop up this economy."

Aaron Rupar

158,278 次观看 • 1 年前

Innovation Council President Taylor Budowich lays out how to deal with concerns about data centers. Start by listening to people’s concerns—and take them seriously. Good policy is good politics. And good policy is: not a single shovel should go in the ground, and not a single slab of concrete should be laid for a data center that drains the surrounding community. The only data centers that should be built are the ones that power the community. Then tell the story. Many data centers, especially in red states are doing just that--often in towns that have been abandoned by politicians for decades. From today's Innovation Council event with Breitbart News, moderated by Matthew Boyle:

Innovation Council President Taylor Budowich lays out how to deal with concerns about data centers. Start by listening to people’s concerns—and take them seriously. Good policy is good politics. And good policy is: not a single shovel should go in the ground, and not a single slab of concrete should be laid for a data center that drains the surrounding community. The only data centers that should be built are the ones that power the community. Then tell the story. Many data centers, especially in red states are doing just that--often in towns that have been abandoned by politicians for decades. From today's Innovation Council event with Breitbart News, moderated by Matthew Boyle:

Innovation Council

149,761 次观看 • 2 个月前

Here is the full run of MicroFactory, that autonomously assembles a photo frame - a real product sold on Amazon. This is a good example of how to deal with complex tasks, while we do not yet have a large set of robotic data to train a big general ai model. Thread:

Here is the full run of MicroFactory, that autonomously assembles a photo frame - a real product sold on Amazon. This is a good example of how to deal with complex tasks, while we do not yet have a large set of robotic data to train a big general ai model. Thread:

Igor Kulakov (MicroFactory)

210,614 次观看 • 1 年前

Ex-OpenAI Psyho says much of AI research is probably glorified hyperparameter optimization, with surprisingly few frontier lab researchers trying something genuinely new: "In competitive programming, we optimize for challenge and fun, and AI research is like the sad reality. In competitive programming, a good problem is one where when someone reads it, they have no clue how to solve it, optimized for creativity." "In AI research, it depends what part you're working in. A lot of people are probably doing a glorified hyperparameter optimization. There are people trying to come up with something completely new." "Agents reduce mundane work in AI research. You no longer have to thoroughly read papers. But you have 50 times more papers that are complete slop and completely useless. Maybe not 50, but it's a big problem today." "Every person in a lab would have completely different responsibility. I really wonder how many people in frontier labs, like at OpenAI or Anthropic, how many of them are trying something genuinely new. I would imagine that number is actually pretty low."

Ex-OpenAI Psyho says much of AI research is probably glorified hyperparameter optimization, with surprisingly few frontier lab researchers trying something genuinely new: "In competitive programming, we optimize for challenge and fun, and AI research is like the sad reality. In competitive programming, a good problem is one where when someone reads it, they have no clue how to solve it, optimized for creativity." "In AI research, it depends what part you're working in. A lot of people are probably doing a glorified hyperparameter optimization. There are people trying to come up with something completely new." "Agents reduce mundane work in AI research. You no longer have to thoroughly read papers. But you have 50 times more papers that are complete slop and completely useless. Maybe not 50, but it's a big problem today." "Every person in a lab would have completely different responsibility. I really wonder how many people in frontier labs, like at OpenAI or Anthropic, how many of them are trying something genuinely new. I would imagine that number is actually pretty low."

MTS

14,432 次观看 • 2 个月前

What if we stopped calling them data centers? Samo Burja: "Consider why do we talk about data centers, not supercomputers?" "A data center in my hometown that sounds like someone's gonna steal my data. I don't feel like I have any opportunities." "A supercomputing center in my hometown? Suddenly I might have opportunities. I might have a job. Galaxy-brain stuff is happening there." "These are not data centers anymore, these are actually distributed supercomputers." "If the median use case is not the laptop on my desk or the phone in my hand, the chips can be pretty big. You might actually see the macro chip. I think people are underrating that future and not really expecting it."

What if we stopped calling them data centers? Samo Burja: "Consider why do we talk about data centers, not supercomputers?" "A data center in my hometown that sounds like someone's gonna steal my data. I don't feel like I have any opportunities." "A supercomputing center in my hometown? Suddenly I might have opportunities. I might have a job. Galaxy-brain stuff is happening there." "These are not data centers anymore, these are actually distributed supercomputers." "If the median use case is not the laptop on my desk or the phone in my hand, the chips can be pretty big. You might actually see the macro chip. I think people are underrating that future and not really expecting it."

MTS

13,454 次观看 • 5 个月前

Over the last few months, we’ve been thinking about how to learn from “off-domain” data - data from non-robot sources like video or simulation. These data sources are not quite good enough to learn policies (even monolithic VLA models) directly, but they still contain lots of information that can be useful for generalizable robot control. How can we develop robot learning models that are able to make use of this type of data for generalizable control? In new work, that we call HAMSTER, we show that VLMs can be useful for enabling robotic learning from off-domain data, but specifically when used through hierarchical VLA architectures. We show that this class of models can learn generalizable robot policies for the real world from large-scale, off-domain data. A 🧵 (1/10)

Over the last few months, we’ve been thinking about how to learn from “off-domain” data - data from non-robot sources like video or simulation. These data sources are not quite good enough to learn policies (even monolithic VLA models) directly, but they still contain lots of information that can be useful for generalizable robot control. How can we develop robot learning models that are able to make use of this type of data for generalizable control? In new work, that we call HAMSTER, we show that VLMs can be useful for enabling robotic learning from off-domain data, but specifically when used through hierarchical VLA architectures. We show that this class of models can learn generalizable robot policies for the real world from large-scale, off-domain data. A 🧵 (1/10)

Abhishek Gupta

11,994 次观看 • 1 年前