Get your free copy of Editor's Digest
FT editor Roula Khalaf picks her favourite stories in this weekly newsletter.
Open source artificial intelligence was one of the most surprising tech stories of the past year. As companies like OpenAI and Google poured billions of dollars into building ever more powerful AI, “open” models that developers can freely use and adapt have filled the performance gap.
There's just one drawback: Most of these open source systems aren't very open. Critics accuse their proponents of “open washing” — that is, trying to reap the benefits of the open source halo effect, free from the constraints of typical commercial software products, without actually living up to the name.
Efforts to create a truly open-source version of AI are finally gaining momentum. But there's no guarantee that progress will match the progress of open-source software, which has played a key role in the tech world over the past two decades. Traditional open-source software, such as the Linux operating system, allows developers to freely explore, use, and adapt the code. So-called open-source AI is also very different in that most modern AI systems learn from data, rather than programming their logic in code.
Take Meta's Llama for example: only the “weights” that determine how the model responds to queries are exposed – users can take those and adapt them, but they can't see the underlying data it was trained on, and don't have enough information to recreate the model from scratch.
For many developers, this still has a clear benefit: they can adapt and train semi-open models based on their own information without having to hand over sensitive internal data to other companies.
But not being fully open comes at a cost. According to Ayah Bdeir, a senior advisor at the Mozilla Foundation, only truly open source technology can give people a full understanding of a system that is beginning to affect every aspect of our lives, while also ensuring that innovation and competition are not stifled by a few dominant AI companies.
One answer comes from the Open Source Initiative, which developed the definition of open source software more than 20 years ago. This week, the group released a near-final definition that could help set the tone for how the field develops.
This would require not just publishing the model weights, but also enough information about the data the model was trained on to allow others to reproduce it, as well as all the code behind the system. Other groups, including Mozilla and the Linux Foundation, are pursuing similar efforts.
Such moves are already leading to increased fragmentation in the AI world. Many companies are being cautious about using terminology, perhaps mindful that OSI owns the trademark on the term “open source” and could sue to block its use in AI models that don't fit the OSI definition. For example, Mistral calls Nemo an “open weight” model.
Alongside partially open systems, fully open source models are beginning to emerge, such as the Olmo large-scale language model developed by the Allen Institute for AI. But it's not yet clear whether this version will have as profound an impact on the world of AI as it has on traditional software. For this to happen, two things need to happen:
For one, the technology needs to meet enough demand to attract a sufficient number of users and developers. In traditional software, the Linux server operating system has become a clear alternative to Microsoft's Windows, attracting a large user base and strong support from Microsoft rivals such as IBM and Oracle. There is no equivalent to Linux in the AI world; the market is already fragmented, and many users will find a semi-open LLM such as Llama to be sufficient.
Proponents of open-source AI also need to make a more compelling case for its safety: The prospect of opening up such powerful, general-purpose technology for anyone to use understandably raises widespread concerns.
Oren Etzioni, former president of the Allen Institute, says many concerns are overblown. When it comes to searching the internet for instructions on how to make a bomb or biological weapon, “these[AI models]don't give you any more information than you can get from Google. There are a lot of AI models out there, they're just packaged differently.” Etzioni acknowledges that giving AI more freedom could do harm in some areas, such as the automated generation of misinformation online.
“Closed” AI does carry risks, but until the further marginal risks and potential benefits of open sourcing the technology are more thoroughly studied, concerns will remain.
Video: AI power demand could stifle industry growth | FT Energy Source
Source link
