Nadella warns AI companies train models on customers' business secrets

Full text

Microsoft CEO Satya Nadella has warned AI companies are training their models on the business secrets of their customers

These secrets are then used to train new, more powerful models, that are sold to their customer's competitors

But, Nadella says there is a way to remain competitive without being locked in to one AI vendor

Microsoft CEO Satya Nadella has warned the big players in the AI industry are using their proprietary models to learn the business secrets of their customers, which they can then use to train and deploy more advanced AI models.

The crux of the issue, Nadella said in a blog post , is that, “You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!”

What Nadella is saying in essence, is that AI companies are harvesting sensitive business data from their customers, using it to make training their models cheaper, and then launching these models for use by their own customer’s competition.

“The kind of knowledge a competitor could never buy”

“Models learn from ‘exhaust,’ the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong. Every correction is distilled into institutional know-how,” Nadella explained.

Nadella also criticized how AI companies are increasingly complaining about how their models are being distilled by their own competition. For example, Anthropic accused retailer and e-commerce company Alibaba for using thousands of Claude prompts to distill their own models . By figuring out how a proprietary model works, you don’t have to spend the enormous amount of capital needed to source training data and create your own AI model.

This, for Nadella, is a major contradiction in how AI companies work. “While the great innovation that comes from model providers having fair use rights to train models on public data is needed, I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation,” he said.

It is also therefore hypocritical for AI companies to accuse other companies of distilling their own product, and then include within their AI usage contracts clauses that allow AI companies to “reserve the right to learn from customer usage and interaction data.”

“In consuming intelligence, you are creating intelligence. And what you create should belong to you,” Nadella added.

On-prem is back in fashion

Nadella’s fix for this growing problem? It’s time to move back to on-prem. Nadella encourages businesses to “retain ownership” of the data they feed AI models by switching to the use of “proprietary learning environments” built on the cloud.

The added benefit of moving to these environments is that they allow businesses to switch between different AI models provided by different companies using “orchestration layers” and AI gateways.

There is also a growing trend of businesses switching to using open source technologies, which goes hand in hand with businesses operating in the cloud. Businesses can train open source AI models using their data that is already available in cloud environments to do much of what the proprietary models do, for far cheaper — and without handing over that same sensitive data to be used by AI companies to train their own models.

The on-prem solution also has additional benefits. AI models operated on-site within manufacturing plants, stores, and other premises are far cheaper and require less specialized hardware. Businesses that operate using a centralized cloud are increasingly encountering issues with data egress fees, storage bloat, and idle specialized hardware.

Google Cloud recently released a report about these very issues , and also encouraged businesses to move towards using AI gateways and on-prem models to reduce latency, improve resilience, and cut per-token costs by switching to local, highly optimized models.

Via TechCrunch