Startups develop tools to verify US-China AI agreements — South China Morning Post
Ahead of the planned meeting in Washington, US, between US President Donald Trump and Chinese President Xi Jinping, researchers and startups are developing technologies that could help verify compliance with potential agreements on restraining artificial intelligence development. Such tools are intended to allow governments and third parties to verify certain information about the training and use of AI models, possibly without direct access to the models themselves or computing infrastructure, the South China Morning Post reports.
What they propose to verify
Tom Milton, co-founder of British consulting company Amodo Design, identified two key questions for a possible international agreement between Washington and Beijing. These concern verifying whether a data center is training a new model or merely serving an already created one, as well as confirming that the deployed model matches the one declared.
According to Milton, serving an already created model, or inference, differs from training a new model. Governments may also seek to verify whether another country's data center is using a model considered safe rather than a potentially dangerous system.
The authors of an open letter representing 20 countries called for the creation of an international governing body that could carry out AI verification. At the same time, the letter did not specify exactly how this should be done.
More current news is available on the UA.News Telegram channel Telegram.
Testing the technologies
One proposed method is called recomputation. It involves checking an AI result by running the same calculation on another system with the same inputs and settings. According to the Machine Intelligence Research Institute, equipping data centers with such tools may cost only a small fraction of a facility's initial costs, while computing expenses can be reduced by checking a random sample of activity.
Lucid Computing began an 18-month test of its solution earlier this month on a cluster of 16 Nvidia H100 chips in Sweden, provided by the state research institute RISE. Amodo Design is in talks about scaling the trials up to a cluster of 128 graphics processing units. Milton acknowledged that the developments still need to be tested at the scale of cutting-edge data centers, where more than one million graphics processing units may be installed.
Cryptographic approach
Israeli startup Attestable, which raised $20 million in seed funding last month, uses zero-knowledge proofs. The company proposes using this cryptographic technology to verify that a response to a request was provided by a specific AI model, without disclosing information about the underlying model or data.
Lucid Computing, Amodo Design and the Singapore AI Safety Hub said they intend to publish the results of their research. According to Lucid's head of policy and strategy, Connor Dunlop, around 50 people worldwide are working on AI verification technologies.
