OpenAI and other tech giants must notify the US government when they start new AI projects

When OpenAI's ChatGPT took the world by storm last year, it surprised many power brokers in both Silicon Valley and Washington, DC. The US government should now have an early warning about future AI breakthroughs linked to the big language model of the technology behind ChatGPT.

 

The Biden administration is preparing to use the Defense Production Act to force tech companies to notify the government when they use significant amounts of computing power to train an AI model. This rule may come into effect from next week.

 The new requirement would give the US government access to key data from some sensitive projects among OpenAI, Google, Amazon, and other tech companies competing in AI. Companies must provide security testing information on their new AI creations.

 

OpenAI is silent on how much work it has done on the successor to its current top offering, GPT-4. The US government will be the first to know when work or safety testing will begin on GPT-5. OpenAI did not immediately respond to a request for comment.

 

"We're using the Defense Production Act, which is an authority we have because of the president, to do a survey that every time companies train a new large-scale language model and share the results with us — safety data — so we can review," U.S. Commerce Secretary Gina Raimondo said Friday at an event at Stanford University's Hoover Institution. He did not say when the requirement would go into effect or what action the government might take about the information. Received about the AI project. More details will be announced next week. It is expected that

 

The new rules are being implemented as part of a White House executive order issued last October. The executive order gave the Commerce Department a Jan. 28 deadline to come up with a scheme that would require companies to tell U.S. officials details about powerful new AI models in development. The order states that these details must include the amount of computing power used, data ownership information provided in the model, and details of security testing.

 

The October order called for AI models to begin working to determine when to report to the Commerce Department but set an initial bar of 100 septillions (one million billion billion or 1026) floating-point operations or flops per second and 1,000 levels. Many times less for large language models operating on DNA sequencing data. Neither OpenAI nor Google disclosed how much computing power they used to train their most powerful models, GPT-4 and Gemini, respectively, but a Congressional Research Service report on executive order revealed that 1026 flops were used to train GPT. Slightly more -4.

 

Raimondo also confirmed that the Commerce Department will soon implement another requirement in an October executive order requiring cloud computing providers such as Amazon, Microsoft, and Google when a foreign company notifies the government of their resources to train a large language model. Foreign projects must be reported when they cross the same initial threshold of 100 septillion flops.

 

Raimondo's announcement comes the same day Google released new data highlighting the capabilities of its latest artificial intelligence model, Gemini, which outperforms OpenAI's GPT-4 that powers ChatGPT in some industry benchmarks. If the project uses enough of Google's cloud computing resources, the Commerce Department could get an early warning about Gemini's successor.

 

Rapid advances in AI last year prompted some AI experts and executives to halt the development of anything more powerful than GPT-4, the model currently used for ChatGPT.

 

Samuel Hammond, senior economist at the Foundation for American Innovation, a think tank, said a key challenge for the federal government is not to exceed a computational threshold in training a model to be potentially dangerous.

 

Dan Hendricks, director of the nonprofit Center for AI Safety, said the need is commensurate with recent developments in AI and concerns about its power. "Companies are spending billions on AI training, and their CEOs are warning that AI could become super-intelligent in the next few years," he says. "It seems reasonable for governments to be aware of what AI companies are doing."

 

Anthony Aguirre, Executive Director of the Future of Life Institute, a nonprofit dedicated to ensuring transformative technologies benefit humanity. "So far, giant experiments have been going on with effectively zero outside oversight or regulation," he says. “Executing this AI training and reporting on the associated security measures is an important step. But much more is needed. There is strong bipartisan agreement on the need to regulate AI, and hopefully, Congress can act on this soon.”

 

Raimondo said at a Hoover Institution event on Friday that the National Institute of Standards and Technology (NIST) is working to set standards for safety testing of AI models as part of the creation of a new US government AI Safety Institute. Determining how weak an AI model usually involves testing a model and trying to produce problematic behavior or output, a process is known as "red teaming".

 

Raimondo said his department is working on guidelines that will help companies better understand the risks lurking in their hatching models. These guidelines could include ways to ensure that AI cannot be used to violate human rights, he suggested.

 

An October executive order on AI gives NIST until July 26 to implement those standards, but some working with the agency say it lacks the funding or expertise to do so adequately.