TORONTO — As technology firms work to build businesses atop artificial intelligence, they are trying to figure out which of their innovations to make public—and which to hold back in pursuit of profits.

Executives and researchers say the explosion of commercial interest in machine learning applications has made some companies clam up, in what has traditionally been an open field. With industry rather than academia now driving the most advanced areas of AI, some worry the shift could slow the development of the technology and hinder safety initiatives. “This tightening has real implications for progress,” said Sara Hooker, vice-president of Cohere For AI, the research arm of Toronto-headquartered startup Cohere.

Computer scientists in university labs made critical early breakthroughs in machine learning, sharing their ideas in academic journals and at research conferences for others to build on. In the mid-2010s, U.S. tech firms hired many of those people into their own labs, but let them continue to publish.

In June 2017, Google researchers published a paper that made the current generative AI boom possible. OpenAI used the researchers’ “transformer” innovation—which changes the way AI systems read or write strings of words—to develop the large language models (LLMs) that would power ChatGPT. (One of the paper’s eight authors, Aidan Gomez, went on to co-found Cohere.) 

After OpenAI launched ChatGPT in November 2022, companies started rethinking their approach to sharing advancements. Before, “everyone was publishing, in all the major labs, all the time—it was a beautiful state,” said Richard Socher, CEO of You.com, which makes AI assistants for workers. But ChatGPT’s success showed that there was significant money to be made by selling models instead of publishing results in research papers. “OpenAI is kind of the end of some open AI research,” Socher said.

The leaders of some industrial labs say they must balance sharing significant advancements against protecting innovations that are important to their business. 

“Our goal is to produce the next generation of models,” said Hooker of Cohere For AI. “That’s how you leapfrog.” Her 30-person team works on systems that use multiple languages as well as formats beyond text like images, audio and video. It’s also trying to make models safer and more efficient.

In February, Cohere For AI released Aya, an LLM conversant in 101 languages, developed with thousands of independent researchers. The group open-sourced the model and datasets it trained on, allowing anyone to use or recreate it. But Cohere’s commercial offerings also benefited from what the team learned during the project. For example, Hooker notes the firm’s Command R+—a product which clients use in chatbots and assistants—performs particularly well in Arabic, Korean and Simplified Chinese. 

“Some of the state-of-art breakthroughs we’ve done could not have happened without this commitment to keeping collaboration open,” she said. (Cohere does hold back proprietary data used to train its models as well as some performance improvements, for both business and safety reasons.)

Santa Clara, Calif.-headquartered ServiceNow has a 55-person AI research group centred in Montreal, where the enterprise software firm acquired Element AI in November 2020. The lab is split between doing science for publication and working with the company’s engineering and product teams to add those advancements to the software it sells. 

“There is a lot of noise out there, and you cannot just pursue every little thing,” said Nicolas Chapados, vice-president of research. His research group lets ServiceNow filter out the most important new developments in AI to pursue, and prove to potential customers and hires that it’s serious about AI. It’s currently working on AI agents, which users can send to perform tasks around the web, and on systems that improve supply chain forecasts.

In February, ServiceNow released StarCoder2, a LLM that writes in programming languages. The firm built it with independent researchers and open-source site Hugging Face, and shared the model’s weights, which guide how it turns inputs into outputs. “We are very keen on offering open-source contributions,” Chapados said.

The firm is less likely to publish about technologies it’s building into products right now, and when it does, it replaces the proprietary datasets with public data. “We want to maintain our ability to operate as an AI vendor,” Chapados said. ServiceNow sometimes files for intellectual property protections before putting out research papers. 

Palo Alto, Calif.-based You.com has applied for a patent on the use of LLMs in search, a hot field right now. “Part of me thinks that patents are kind of silly,” said Socher. But he noted they can help head off lawsuits from competitors and generate revenue through licensing. Publishing research papers only makes sense once a startup is profitably generating revenue, Socher said. “We’re shipping product, and we’re focused on that.”

Industrial labs have used their resources and workforces to take the lead in advancing AI. Model-makers are spending huge sums on chips and processing power, compute to which academic scientists don’t have access. 

With those labs publishing less, the pace of AI breakthroughs could slow. Take multimodal models, which can take in and generate text, images and other formats. They debuted after company-controlled labs had started sharing less; as a consequence, proprietary ones significantly outperform open-source competitors, Hooker said. There’s less of a gap among LLMs, which the sector started working on when labs were publishing more. 

The protective instinct also poses challenges when it comes to safety. Policymakers could have a harder time monitoring and regulating AI developments if companies are keeping more of what they produce secret. Yet some researchers, including former University of Toronto professor Geoffrey Hinton, have called for a ban on open-sourcing the most powerful models to prevent bad actors from getting access to them.

Hooker said AI’s closed era could be particularly damaging because the pool of researchers in the field is still quite small and interconnected. “It has implications for when discoveries are made and where,” she said. “This can set back fields [for] decades.”