Canadian AI Startup Cohere Shifts Focus to Custom Models

By Krystal Hu

Canadian AI startup Cohere, last valued at $5.5 billion, will pivot towards creating tailored models for enterprise users rather than focusing on larger foundation models, as reported by Reuters.

This shift in strategy was detailed in a letter to investors on Thursday, coinciding with many companies continuing to navigate the incorporation of large language models since ChatGPT became popular two years ago.

> “What we're hearing from customers is that they don't just need bigger models to be good at everything. They need models that are actually built for their specific use cases,” said Nick Frosst, co-founder of Cohere, in an interview.

Cohere, viewed as a competitor to AI labs like OpenAI and Anthropic, will maintain its development of foundation models but will concentrate on different training techniques to enhance model performance without increasing their size. While the sale of Application Programming Interfaces (APIs) will remain a minor part of its business, the priority will be on deploying customized models.

The competitive landscape of AI has led to significant investment from startups and tech giants alike. Companies like OpenAI, Anthropic, and xAI have secured billions to support the intense resource requirements for developing state-of-the-art AI models.

Cohere operates out of Toronto and San Francisco and has raised over $900 million from investors such as Nvidia, Cisco, and Innovia Capital.

Positioning itself as an enterprise-centric AI company, Cohere has been collaborating with clients like Oracle and Fujitsu to create models tailored to their specific needs.

Cohere's new emphasis on customization aligns with a broader industry trend where previously successful scaling of computation power and model sizes is yielding diminishing returns. AI labs are experiencing delays in training the latest generation of large language models. Ilya Sutskever, co-founder of AI labs Safe Superintelligence and OpenAI, noted that advancements from scaling pre-training efforts have plateaued.

Frosst emphasized that merely enlarging model size doesn't guarantee improved outcomes. The focus on customization is expected to enable Cohere to operate more efficiently, minimizing the need for extensive computational resources. The company is not aiming for artificial general intelligence (AGI) like OpenAI.

> “We're going to work with an enterprise to figure out how we can make the model perfect at their use case, tailor it to that specific needs and get to production, not bank on the AGI future is coming next year,” Frosst remarked.