Silicon Valley has been embroiled in a pitched battle over how artificial intelligence software should be created. Some have said that A.I. models should be “closed” because they are too powerful and dangerous to be developed openly. Others have said that A.I. models must be “open source” for people to further develop the technologies and build new businesses.
Amid the debate, technologists have battled over how much openness is enough, citing “open weights.” What are open weights? We’ll explain it.
First, what is open-source software?
When software developers decide to make their technology open source, they publish all of the underlying computer code. They also release the code under an “open license,” which lets other developers use it without paying fees or royalties.
There are several advantages to making code publicly available.
One is security. Private code can be reviewed by a company’s employees for flaws, but public code can be reviewed by everyone so that many more technologists weed out mistakes that could lead to cybersecurity incidents.
Another is faster innovation. Many companies have leaned into open-source software to bootstrap new products more quickly, instead of writing code from scratch. Today, open-source software undergirds much of the modern internet.
What are “open weights” in A.I.?
A.I. models learn their skills by pinpointing patterns in massive amounts of digital data. By finding patterns in books, news stories and other online documents, an A.I. model can learn to generate English on its own. The model stores these patterns as numerical calculations called “weights.” These values determine how an A.I. model “thinks” as it responds to prompts from users.
When technologists refer to open weights, they mean that the calculations, or rules, governing the A.I. system are public. People who use the system can tweak those rules to adjust it to their own needs — perhaps focusing an A.I. system on a particular area like medical research or cybersecurity, or making it prioritize certain information over all else. If the weights aren’t public, people have to use the A.I. system just the way its makers designed it.
Open weights do not go so far as to disclose all of the code that makes the system work. Some A.I. companies have released their underlying code, while others have published only the weights. And other companies have not disclosed any details about the inner workings of their A.I. models.
Why do we care about this now?
It boils down to the A.I. race and who holds power over the technology.
Anthropic and OpenAI, the leading A.I. labs, have spent billions of dollars on developing cutting-edge A.I. models that are mostly closed, which means the companies do not share anything about how those models were created. These models have been widely adopted around the world.
But this concentration of A.I. power has raised concerns among tech giants such as Nvidia, Microsoft, Meta and others, who are also competing in this race. The sharing of open weights models, many of these companies argue, could allow for other ventures to be created atop the technology that would broaden the playing field.
There is also a safety component. Releasing open weights models allows for wider scrutiny, which can help companies fix vulnerabilities before those flaws can be exploited, some companies argue.
Anthropic and OpenAI claim that because some of the technology is so advanced, giving all of the code away freely could be more dangerous, especially if the models are used by those with bad intentions. (It also happens to be very bad for their business.)
Releasing A.I. model weights is something of a compromise between open- and closed-source software. Hardcore proponents of open-source software say that open weight models still do not go far enough in sharing information.
(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
How much is China a factor?
China is a major factor in this debate.
China’s tech industry has embraced open source as a tool to catch up in A.I. development. Last year, the Chinese start-up DeepSeek shocked the tech industry when it released an A.I. model that performed as well as the best in the world and published much of the underlying code as open source.
Since then, China has become the source of most of the world’s leading open-source models, including from companies like Alibaba, Z.ai and Moonshot AI. This month, Xi Jinping, China’s leader, also gave a speech portraying the country as a global champion of an open approach to A.I. development.
China’s strength in open-source A.I. has appeared to pose a threat to U.S. competitiveness in the technology, spurring the debate in Silicon Valley about how to move forward. Washington has also been watching the back-and-forth as officials debate how or whether to regulate A.I.
The post What Is Open-Weight A.I.? appeared first on New York Times.
