Charles Rollet is a rockstar tech reporter who recently joined Business Insider. He’s knocked out several scoops about Scale AI lately. This is the startup that Meta is investing in. Founder Alexandr Wang is also joining Meta.
If you want to understand what Mark Zuckerberg is buying, Charles’s coverage is a must-read. He broke news on how Google used Scale AI to chase ChatGPT and led exclusive reporting on how big Scale AI customers are deserting the startup. There’s also an astounding story on how Scale AI left sensitive data about Google, xAI, and Meta open for anyone on the web to see.
I asked Charles what he thought of this peculiar deal.
Why is Zuck paying so much?
If you want a cutting-edge AI model, one way to do that is by having humans rewrite millions of chatbot responses and do lots of other little tricks to painstakingly improve each AI answer. Our reporting shows that Scale AI specializes in this kind of work: it’s what Scale AI did to help Google catch up with ChatGPT, for example. This kind of expertise is valuable for Zuck as he seeks to release new models that blow everyone away. The lukewarm reception to Llama 4 clearly got to him, and he’s taking aggressive action.
What’s the future of Scale AI now, and how can it help Meta and Zuck’s AI ambitions? Or is this just a massive acquihire of Alexandr Wang?
It was stunning to see Google pull Scale AI projects hours after the Meta investment was announced. It’s an open question whether Scale AI can regain that kind of lost revenue. But I don’t think Meta will just let Scale AI peter out and die. Meta now has much more direct access to Scale AI for training its latest models. Zuck needs Llama models to be at, or very near, the top of AI benchmarks. And Scale AI has a good idea of what works and what doesn’t, along with the platform to make things happen. This deal wasn’t solely about hiring Wang — no offense to Alex, but there were probably cheaper ways to do that.
What has your Scale AI reporting revealed about how AI models are developed?
Training AI models requires a ton of human labor. A key way to improve models is through Reinforcement Learning from Human Feedback, or RLHF. Scale AI’s biggest platform, Remotasks, has more than 240,000 human contributors doing this stuff. Ironically, automation is something of a problem here: It turns out that these humans often prefer to copy and paste answers from ChatGPT instead of giving genuine human feedback. We noticed many examples of humans being banned from Scale AI’s platforms for doing this.
What else should BI readers know about this situation?
Zuck made a huge bet with Scale AI. People have been comparing this to the Metaverse, which burned billions and hasn’t panned out yet. But if Meta starts to climb AI rankings and gets Llama models performing much better, this could be seen as a wise move — closer to Facebook buying Instagram.
Sign up for BI’s Tech Memo newsletter here. Reach out to me via email at abarr@businessinsider.com.
Read the full article here
