@FinanceYF5: AI training data is the only product that goes bad the moment it works: once a model gets ~70% of a task right, labs th…
Summary
A thread arguing that AI training data is a uniquely perishable product that loses value once models master a task at around 70% accuracy, fueling a hot market that minted young billionaires even as the largest labs bring data creation in-house.
View Cached Full Text
Cached at: 10/10/26, 04:45 AM
AI training data is the only product that goes bad the moment it works: once a model gets ~70% of a task right, labs throw the task out.
It also minted the youngest self-made billionaires on earth. 3 min on how hot this market is, and why the biggest buyers build their own. 🧵
How hot is it? Mercor’s three founders became the youngest self-made billionaires at 22 (Forbes). Surge’s Edwin Chen never raised a dollar; Forbes puts him at 18B. Meta paid ~14.3B for 49% of Scale AI and took its founder. Dead startups’ old Slack chats sell for 10k–100k.
But it expires. Once a model gets ~70% of a task right, the task is dropped (@EpochAIResearch). As @athomasq put it: once you’ve trained, you’ve trained. And the biggest buyers are building their own: Google hired Mechanize’s founder and team; xAI cut 500 generalist labelers.
It runs the other way too: @edwardjhu, co-creator of LoRA, joined @mercor_ai to train models; their first is built on Alibaba’s Qwen. A 2023 study found a third or more of crowd workers used AI to write “human” answers. The next teachers: factory workers with head cameras.
Three views. @EverettRandle: data is the next $1T category. @viks_rum: fewer than 10 independent players at scale by 2031. @satyanadella: companies pay for intelligence twice, and should keep their own. The most valuable data may be the data you never sell. Would you sell yours?
Similar Articles
@viks_rum: https://x.com/viks_rum/status/2077650169265590727
An in-depth analysis of the booming business of selling training data to frontier AI labs, detailing six distinct data products and the financial dynamics of the market.
AI labs are buying data from dead startups to train their models
AI labs are increasingly buying code bases and workspace data (emails, chats, documents) from failing startups via firms like SimpleClosure and Sunset, as public and synthetic training datasets hit their limits. No clear benchmarks yet prove real-world archives outperform synthetic data, but deals are becoming routine, with valuations driven by speculative 'research theses' about data utility.
AI is deteriorating in realtime
AI models are deteriorating due to training on recursively generated synthetic data, leading to model collapse; multiple studies highlight the risks of scaling with synthetic data.
@quxiaoyin: Turns out Elon is right again. The shittiest layer in AI is the model layer. The real money in AI is in compute, energy…
The tweet argues that the AI model layer is the least profitable, while compute, energy, and applications are where the money is, noting that Chinese open-weight models are eroding the margins of companies like OpenAI and Anthropic.
how are some labs catching up so fast with way less compute?
The author questions why some overseas AI labs are catching up to frontier models with far less compute, speculating that high-quality expert training data matters more than assumed — and claiming US data labs like SurgeAI are undermining the US AI race by selling mass data to China.