Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Open-source artificial intelligence has been one of the most surprising tech stories of the past year. As companies such as OpenAI and Google have poured billions of dollars into building ever more powerful AI, “open” models that are freely available for developers to use and adapt have closed the performance gap.
There’s just one drawback: most of these open-source systems are not very open. Critics accuse their backers of “open washing” — trying to benefit from the halo effect of open source, with its freedom from the constraints of normal commercial software products, but not living up to the name.
The effort to create a truly open-source version of AI is finally gaining momentum. But there is no guarantee that its progress will match that of open-source software, which has come to play a critical role in the tech world over the past 20 years. With traditional open-source software, such as the Linux operating system, the code is freely available for developers to inspect, use and adapt. So-called open source AI has been very different, not least because most modern AI systems learn from data rather than having their logic programmed in code.
Take Meta’s Llama. Only the “weights” which determine how the model responds to queries are disclosed. Users can take and adapt it, but they can’t see the underlying data on which it was trained and don’t have enough information to reproduce the model from scratch.
For many developers, this still has some clear benefits. They can adapt and train quasi-open models on their own information without needing to hand the sensitive internal data over to another company.
But not being fully open has its costs. According to Ayah Bdeir, a senior adviser to the Mozilla Foundation, only a true open-source technology would give people a full understanding of the systems that are starting to affect all facets of our lives, while also guaranteeing that innovation and competition can’t be squashed by a handful of dominant AI companies.
One response has come from the Open Source Initiative — which laid out the definition of open-source software more than 20 years ago. This week, it produced a near-final definition that could help to shape how the field develops.
This would need not only the weights for a model to be released, but also enough information about the data on which it was trained to allow someone else to reproduce it, as well all the code behind the system. Other groups, such as Mozilla and the Linux Foundation, are pushing similar initiatives.
Moves such as these are already leading to a greater segmentation in the AI world. Many companies are being more careful with their terminology — perhaps mindful that the OSI owns the trademark to the term “open source” and could sue to prevent it being used on AI models that fall outside its own definition. Mistral, for instance, calls its Nemo an “open weights” model.
Alongside the partly open systems, full open-source models are starting to appear, such as the Olmo large language model developed by the Allen Institute for AI. Yet it is far from clear that this version will have as big an impact in the AI world as it has had in traditional software. For this to happen, two things would be required.
One is that the technology will need to meet a big enough need to attract a critical mass of users and developers. With traditional software, the Linux server operating system represented a clear alternative to Microsoft’s Windows, winning it a large base of users and strong backing from Microsoft’s rivals, including IBM and Oracle. Linux has no equivalent in the AI world. The market is already more fragmented and many users will find quasi-open LLMs such as Llama adequate.
Backers of open source AI also need to make a better case for its safety. The prospect of such a powerful, general-purpose technology being released for anyone to use rightly stirs widespread concern.
Oren Etzioni, former head of the Allen Institute, says that many fears are overblown. When it comes to going online to research how to make a bomb or a bioweapon: “You’re not really able to get more out of these [AI models] than you’re able to get out of Google. There’s lots of it out there — it’s just being packaged differently.” He concedes that there are some areas where making AI more freely available could cause harm, such as automating the creation of more online misinformation.
“Closed” AI also comes with risks. But until the extra marginal risk of open sourcing the technology has been more thoroughly studied, along with the potential benefits, the fears will remain.
Read the full article here