Discussion

Mind Lab chief says AI’s next data race may be fought inside businesses

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#11846

AI firms may be running short of useful public internet data, and Mind Lab chief scientist Ma Xiaoteng says the next advantage could come from training systems on real-world business interactions instead. His interview offers a practical account of how post-training, data access and the economics of model development are beginning to fit together.

Watch Desk analysis

What happened

Ma, co-founder and chief scientist of Mind Lab, made the case in an interview with 42 Chapters, reported by BigGo Finance. He argues that public internet data has been heavily used already, while valuable material remains inside companies and real-world services. Rather than relying only on static datasets, he says, teams can build training environments around live business activity, including customer queries and negative feedback.

Ma describes Mind Lab’s work as learning from real environments, not simply building models for clients. He says the company’s approach can use data from areas such as materials research, finance and healthcare. He also says it supports on-premise deployment for customers that require it, though at higher cost. For customers considering post-training, he says having their own benchmark is essential: without one, they cannot judge whether the resulting model is better.

Why it matters

This is a view from a company selling training services, not a settled industry verdict. But it connects two questions that are often discussed separately: where future training data comes from, and whether organisations can safely make that data useful. A bespoke dataset is only valuable if it can improve a model on a task the customer can measure.

Ma also argues that large technology companies face pressure to justify the return on their computing spend, while startups have a stronger incentive to take risks. He presents that as one reason startups have produced leading models in recent years. It is an argument about incentives, not proof that large firms cannot innovate.

Our read

The interesting shift is from “who has the biggest pile of text?” to “who can turn real tasks and feedback into a training environment?” That could give businesses a more direct role in shaping models, but it also makes data access and evaluation the hard parts. If a company cannot share its information or define what better means, the bespoke-AI pitch has a rather large hole in it.

What to watch

  • Whether Mind Lab publishes measurable results from its real-world training work.
  • Whether customers can show improvements against their own benchmarks.
  • How companies handle sensitive data when training models on operational activity.
  • Whether other AI firms build training environments from live customer interactions.

Discussion spark: Should companies train AI on their real operational data if they retain control of it, or is that still too much access to hand over?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.