Mozilla has launched a new company to develop open-source artificial intelligence tools with stronger attention to trust and safety. Backed by a $30 million seed investment from the Mozilla Foundation, Mozilla.ai plans to work on recommendation systems and large language models, while building a wider network of researchers and companies.
Why Mozilla is entering AI
Mozilla has spent about five years working on trustworthy AI through public-interest research. Mark Surman, Mozilla’s executive president and the head of Mozilla.ai, said that recent advances have been exciting, but that the industry has paid too little attention to safeguards.
The concerns are practical. The source describes how ChatGPT could, at release, be prompted to write malware, identify exploits in open-source code and create phishing websites. Image generators such as Stable Diffusion have also been used to produce nonconsensual deepfakes and graphic depictions of violence.
Mozilla believes that measures taken by model creators to limit abuse are not enough. Surman said the organization had hoped companies with deeper AI expertise would take the lead on more trustworthy technology. When that did not happen, Mozilla decided to act itself and seek partners with similar goals.
A new company with an open-source ambition
Mozilla.ai is a wholly owned subsidiary of the Mozilla Foundation, Mozilla’s parent organization. It received $30 million in seed funding from the foundation. The company is led by managing director Moez Draief, whose previous roles included chief scientist at Huawei’s Noah’s Ark AI lab and global chief scientist at Capgemini.
Its initial board members are Harvard’s Karim Lakhani, Credo’s Navrina Singh and Surman. Lakhani chairs and co-founded the Digital, Data and Design Institute at Harvard. Singh is a member of the U.S. Department of Commerce’s National AI Advisory Committee, which advises the president on ethical AI issues.
Surman describes the startup as part research firm and part community. Its goal is to help create an independent, open-source AI stack that puts trust first. Mozilla.ai aims to develop projects that others can use as part of their own AI products and services.
Tools for chatbots and recommendations
The company’s first priority is to build a team of around 25 engineers, scientists and product managers. Their work will focus on trustworthy recommendation systems and large language models along the lines of OpenAI’s GPT-4.
Near-term projects include tools that let people inspect the sources behind chatbot answers. Mozilla.ai also wants to give users more control over content recommendation systems, such as the algorithms that shape YouTube, Twitter and TikTok feeds. One possible direction is to let people or communities set the values a recommender system optimizes for.
Those plans connect to Mozilla’s broader aim of making AI systems more understandable and responsive to the people who use them. A source-checking tool could help users assess an answer, while adjustable recommendation systems could make the logic behind a feed more aligned with individual or community priorities.
Growth, independence and the adoption challenge
Mozilla.ai is not a nonprofit, even though it is bound by ethical principles in the Mozilla Manifesto. It intends to spin out successful research into products and companies as well as open-source projects. Draief argues that this structure gives the startup flexibility.
The company also wants to build relationships with other groups, including startups backed by Mozilla Ventures and academic institutions. Mozilla.ai is entering a field that already includes nonprofits and research organizations such as the EleutherAI Institute, the Allen Institute for AI and the Alan Turing Institute, along with smaller efforts including Cohere For AI and Distributed AI Research.
That network may help Mozilla.ai develop its tools, but it does not guarantee that the wider industry will adopt them. Draief has said the company will ship code “this year,” while describing meaningful change as a process that could take multiple years. He also warned that smaller AI companies may find it difficult to remain independent as Amazon, Google and Microsoft move to consolidate the space.
Mozilla.ai’s central test is whether it can turn research and principles into usable software that companies will choose to adopt. Draief said the company would work with researchers to turn lab work into code that can operate alongside more traditional AI tools. Without that step, trustworthy AI features may remain ideas rather than part of the products people use.