TToday best AI models, like OpenAI’s ChatGPT and Anthropic’s Claude, come with conditions attached: their creators control the conditions under which they are accessible to prevent them from being used in harmful ways. This contrasts with “open” models, which can be downloaded, modified, and used by anyone for almost any purpose. A new report by a non-profit research organization AI era found that the open models available today are about a year behind the best closed models.

“The best open model today is comparable in performance to closed models, but with a lag of about a year,” says Ben Cottier, lead researcher of the report.

The meta Llama 3.1 The 405B, an open model released in July, took about 16 months to reach the capabilities of the first version of GPT-4. If Meta’s next-generation AI, Llama 4, is released as an open model, as is widely expected, this gap could narrow further. These findings come as policymakers grapple with how to manage increasingly powerful AI systems, which have already been tested. remodel information environments ahead of elections around the world, and which some experts fear may one day be able to design pandemicsperforming sophisticated tasks cyberattacksand causing other harm to humans.

Researchers at AI era analysis hundreds notable models released since 2018. To arrive at their results, they measured the performance of the best models on technical criteria – standardized tests that measure an AI’s ability to handle tasks such as solving math problems, answering general knowledge questions and demonstrate logical reasoning. They also looked at the amount of computing, or computing, power used to train them, as this has historically been a good indicator of capabilities, although open models can sometimes perform as well as closed models while using less calculation, thanks to the progress made in the field. the effectiveness of AI algorithms. “The gap between open and closed models provides policymakers and AI labs with a window to assess cutting-edge capabilities before they become available in open models,” Epoch researchers write in the report.

Learn more: The researcher tries to glimpse the future of AI

But the distinction between “open” and “closed” AI models is not as simple as it seems. While Meta describe its Llama models being open-source, it does not respond to new definition released last month by the Open Source Initiative, which has historically set the industry standard for what constitutes open source. The new definition requires companies to share not only the model itself, but also the data and code used to train it. Although Meta publishes its model “weights” (long lists of numbers that allow users to download and modify the model), it does not publish the training data or the code used to train the models. Before downloading a template, users must agree to a Acceptable Use Policy which prohibits military use and other harmful or illegal activities, although once the models are downloaded these restrictions are difficult to enforce in practice.

Meta disagrees with the Open Source Initiative’s new definition. “There is no single definition of open source AI, and defining it is challenging because previous open source definitions do not encompass the complexity of today’s rapidly evolving AI models,” said a spokesperson. Meta’s word to TIME in an emailed statement. “We make Llama free and openly available, and our license and Acceptable Use Policy help keep people safe by putting certain restrictions in place. We will continue to work with OSI and other industry groups to make AI more accessible and responsibly free, regardless of technical definitions.

Making AI models open is widely seen as beneficial because it democratizes access to the technology and drives innovation and competition. “One of the key things that open communities do is they bring together a larger, more geographically dispersed and more diverse community involved in the development of AI,” says Elizabeth Seger, director of digital policy at Demosa UK-based think tank. Open communities, which include academic researchers, independent developers, and nonprofit AI labs, also drive innovation through collaboration, including making technical processes more efficient. “They don’t have the same resources as the big tech companies. So it’s very important to be able to do a lot more with a lot less,” says Seger. In India, for example, “AI embedded in public service delivery is almost entirely based on open source models,” she says.

Open models also allow for greater transparency and accountability. “There needs to be an open version of any model that becomes a basic infrastructure for society, because we need to know where the problems come from,” says Yacine Jernite, head of machine learning and society at Hugging Face , a company that manages the digital infrastructure where many open models are hosted. He cites the example of Stable Diffusion 2, an open image generation model that allowed researchers and reviewers to examine its training data and combat possible bias or copyright violations – this which is impossible with closed models like that of OpenAI. SLAB. “You can do it a lot easier when you have the receipts and the records,” he says.

Learn more: The heated debate over who should control access to AI

However, the fact that open models can be used by anyone creates inherent risks, as people with malicious intentions can use them for harmful purposes, for example to produce child pornography, or even be used by rival states. Last week, Reuters reported that Chinese research institutes linked to the People’s Liberation Army had used an older version of Meta’s Llama model to develop an AI tool for military use, highlighting the fact that once a model was made public, he can no longer be recalled. Chinese companies like Alibaba have also developed their own open models, would have competitive with their American counterparts.

Monday, Meta announcement it would make its Llama models available to U.S. government agencies, including those working on defense and national security applications, and to private companies supporting government work, such as Locked Martin, AndurilAnd Palantir. The company says U.S. leadership in open source AI is both economically beneficial and crucial to global security.

Closed proprietary models present their own challenges. While they are more secure, because access is controlled by their developers, they are also more opaque. Third parties cannot inspect the data the models are trained on for bias, copyrighted material, and other issues. Organizations using AI to process sensitive data may choose to avoid closed models for privacy reasons. And while these models come with stronger guardrails to prevent misuse, many people have found ways to protect them. ‘jailbreak’ them, effectively circumventing these safeguards.

Governance challenges

Currently, the safety of closed models is primarily in the hands of private companies, although government institutions such as the U.S. AI Safety Institute (AISI) are playing an increasingly important role. role in security testing models before release. In August, the US AISI signed formal agreements with Anthropic to enable “formal collaboration on AI research, testing and safety evaluation”.

Due to the lack of centralized control, open models present distinct governance challenges, particularly regarding the most important areas. extreme the risks that future AI systems could pose, such as the empowerment of bioterrorists or the intensification of cyberattacks. How policymakers should respond depends on whether the capability gap between open and closed models narrows or widens. “If the gap continues to widen, then when we talk about border AI security, we don’t have to worry as much about open ecosystems, because everything we see will happen with models first closed, and these are easier to regulate,” says Seger “However, if this gap is to narrow, then we need to think much more seriously about whether, how and when to regulate the development of open models. , which is a whole other can of worms, because there is no central entity that can be regulated.”

For companies like OpenAI and Anthropic, selling access to their models is central to their business model. “A key difference between Meta and closed model providers is that selling access to AI models is not our business model,” Meta CEO Mark Zuckerberg wrote in a statement. open letter in July. “We expect future Llama models to become the most advanced in the industry. But even before that, Llama is already a leader in openness, modifiability and cost-effectiveness.

Measuring the capabilities of AI systems is not simple. “Capabilities is not a defined term one way or another, which makes it a terrible thing to discuss without a common vocabulary,” Jernite explains. “There are a lot of things you can do with open models that you can’t do with closed models,” he says, emphasizing that open models can be adapted to a range of use cases and that they can outperform closed models when trained to do so. specific tasks.

Ethan MollickWharton professor and popular technology commentator, argues that even if there were no further advances in AI, it would likely be years before these systems were fully integrated into our world. With new capabilities being added to AI systems at a steady pace, in October, frontier AI lab Anthropic introduced the ability of its model to directly control a computer, still in beta: the complexity of governing this technology will only increase.

In response, Seger says it’s critical to determine exactly what risks are at play. “We need to establish very clear threat models describing what harm is and how we expect openness to lead to the realization of this harm, and then determine the best point of intervention among these individual threat models. »