Meta bets on human training data with Scale AI's Alexandr Wang

Could Wang pull off the most ambitious data labelling project in human history? Also, AMD launches new chip line to rival Nvidia's Blackwell, and more.

Tara Tan

Jun 13, 2025

What we’re paying attention to, and why

Meta is sitting on what’s arguably the most valuable asset in the AI race: human training data.

Roughly 20% of the world’s mobile internet traffic flows through its empire of apps like WhatsApp, Instagram, Facebook.

That’s 3.5 billion daily active users.

All creating the most diverse, high-volume stream of human behavior data set to ever exist.

It will be the most ambitious data labelling project ever for fmr Scale AI CEO Alexandr Wang as he joins Meta to lead a superintelligence AI unit for $14.3B.

Scale AI, which Wang founded, has become foundational to the AI model wars, by providing human-verified data pipelines to modelmakers and companies like OpenAI, Anthropic, Microsoft, and Bytedance.

But this partnership with Meta could take his playbook to superscale.

Instead of just relying on low-wage gig workers (some reportedly paid under $1/day), Wang and his team could now embed labeling and training infrastructure directly into Meta’s global data stream, turning every user interaction into a feedback loop for smarter models.

Like new forms of CAPTCHA, AI search answers, or feedback and optimization on creative ads, all built into the bones of the apps billions use daily.

While Meta has mostly bungled their AI products to date (and historically, product has never been their strong suit), they are the strongest forerunner in owning, and now leveraging, high-quality, real-world, human-labeled data.

$14B does not feel that insane of a number if Wang is able to pull off the most ambitious data labelling project in human history.

It could be foundational to Meta’s future.

Have a good weekend - tara

The Download

News that mattered this week

AMD launches MI350 AI chip line to rival Nvidia's Blackwell processors. The MI400 chips will be able to be assembled into a full server rack called Helios, AMD said, which will enable thousands of the chips to be tied together in a way that they can be used as one “rack-scale” system.
OpenAI hits $10B in annual revenue. However, it still isn’t profitable, and is projected to be burning about $28B a year. Currently, OpenAI is serving more than 500 million weekly active users and 3 million paying business customers. The company is targeting $125 billion in revenue by 2029.
Meta acquires data labelling company Scale for $15 billion: Meta is close to finalizing an almost $15 billion investment in Scale AI, the tech giant’s largest-ever external investment, which would give Meta a 49% stake in the company, according to The Information. As part of the deal, Meta CEO Mark Zuckerberg is personally assembling a team of about 50 people to help Meta supercharge its AI goals – specifically, to achieve artificial general intelligence – and Scale AI CEO Alexandr Wang is set to join that group once the deal is final.
Apple makes major AI advance with image generation technology rivaling DALL-E and Midjourney: The advancement, detailed in a research paper published, introduces “STARFlow,” a system developed by Apple researchers in collaboration with academic partners that combines normalizing flows with autoregressive transformers to achieve what the team calls “competitive performance” with state-of-the-art diffusion models.
Apple opens core AI model to developers amid measured WWDC strategy: Apple has opened its Apple Intelligence model to third-party developers for the first time, allowing direct access to the on-device large language model. The new Foundation Models framework allows developers to integrate Apple Intelligence features with just three lines of Swift code, providing privacy-focused AI inference at no cost.
Dia, The Browser Company’s AI-first browser, launches Mac beta. At launch, Dia’s core feature is its AI assistant, which users can invoke at any time. It’s not just a chatbot floating on top of the browser, but rather a context-aware assistant that sees users’ tabs, open sessions, and digital patterns. Users can use it to summarize web pages, compare info across tabs, draft emails based on their writing style, or even reference past searches.
ByteDance and Carnegie Mellon researchers announced PartCrafter, which turns a single photo into fully editable 3D parts in seconds.
Electronic Skin with unique fingerprints for robots: Researchers UNIST have developed an electronic skin with unique, unreproducible fingerprints – offering a powerful solution for robotic traceability and identity verification. Applications include defence robotics, medical devices, where tamper-evident identification is essential for safety and accountability.