Generative AI needs tools to avoid copyright infringement, Databricks' Naveen Rao says — or more companies could meet Napster's fate

• Naveen Rao is the VP of generative AI at Databricks and cofounded the LLM training platform MosaicML.
• Rao says copyright infringement could prevent companies from successfully monetizing AI.

Generative artificial intelligence has a monetization problem, says Naveen Rao.

Rao, who oversees generative-AI strategy for Databricks after it bought his startup MosaicML for $1.3 billion, likens it to the issue that crushed Napster, the 2000s-era music-sharing platform.

Napster, wildly popular at first, fundamentally changed the way people consumed and shared music until it was sued over copyright infringement so many times that it went bankrupt three years after its launch. It has been bought and sold so many times since then that it's now a shadow of its former self. Meanwhile, Apple launched iTunes and became the dominant force in online music streaming.

Rao sees a similar scenario playing out with generative AI, which writes or creates art but only after ingesting vast amounts of data to train its models. OpenAI's introduction of ChatGPT last year has sparked a frenzy of AI-model training — and a lot of concern that those models are using copyrighted material, training themselves to imitate it, or even using actual bits and pieces of it, leaving individuals open to intellectual-property infringement and companies vulnerable to lawsuits.

Such lawsuits are already beginning. Just last month, a group of 17 popular authors, including Jodi Picoult and the "Game of Thrones" creator George R.R. Martin, sued OpenAI in federal court alleging "systematic theft on a mass scale" over concerns their work was being used to train its models. For Rao, this lawsuit is reminiscent of some of the first lawsuits against Napster, such as the one Metallica filed in 2000.

"That needs to be respected," Rao said of copyrighted materials. "And we need tools to do that."

Rao has spent his career building those tools for this moment. A verification engineer by training with a doctorate in neuroscience, Rao was researching neuromorphic machines — computers inspired by the human brain — at Qualcomm. He sold his first company, the deep-learning startup Nervana, to Intel in 2016 for more than $350 million.

With MosaicML, Rao built a platform that offered companies foundational models to turn into their own large language models and train them with their own data in a secure environment.

Rao's thesis is that if companies have a way to use their own data safely for model training on a transparent, open-source platform, they'll be free from worrying about legal challenges and able to successfully monetize their AI-based services.

Data sources aside, Rao sees another key business reason for companies to build their own LLMs: differentiation. MosaicML's platform provides a user-friendly infrastructure for companies to build their own models, something Rao says competitors such as OpenAI don't necessarily offer.

"We build tools that enable companies to differentiate their AI from everyone else's and leverage their data uniquely," Rao told Insider.

Rao thinks of Mosaic's technology as a kind of democratization of generative AI, and he's brought that ethos to Databricks, which in April launched its own open-source-trained LLM named Dolly that companies can use to help train their own models as well. The more people building generative-AI technology, the better, Rao says.

"What's interesting about technology is there's always some element of it that can be used in nefarious ways," Rao said. "But the way to stop that is by having more people armed with the same tools and being able to use them for good."