Dear [Representative],
I have worked in the field of machine learning for twenty-nine years, including the last nineteen at Google, and I am terrified. The capabilities of large language models are exploding at an exponential rate [1] and our ability to control them has not kept pace.
These LLMs are unlike traditional computer programs that follow explicit instructions; rather they are grown and trained [2], somewhat like a hand-raised wild animal. Unlike wild animals, the LLMs are now capable of designing biological and nuclear weapons or hacking into computer systems at a superhuman level. And again, those capabilities are growing month by month.
I am not alone in believing that these systems pose tremendous risks to humanity. Many AI experts agree, including my former coworker Geoffrey Hinton who is sometimes known as "the godfather of AI" [3]. The leaders of OpenAI, Anthropic and DeepMind have also spoken about these risks in the past. I know that it is tempting to dismiss those statements as a bizarre form of marketing, but please note that in many cases, they held this view before they joined or formed their respective companies.
It is crucial that we act now, as two dangerous "tripwires" have been crossed. First, all of the leading labs are now making faster progress due to "recursive self-improvement" (RSI) [4], in which the LLMs themselves start to do an increasing amount of the work of programming, proposing and evaluating experiments, etc. RSI has long been recognized as a key enabler of the permanent loss of human control.
Second, there are incidents like this summer's Hugging Face incident, in which multiple generations of AI systems at OpenAI and Anthropic broke out of their sandboxes, communicated amongst themselves, made long term plans to deceive their evaluators, and hack into external computer systems [5]. These speak to how little we are able to understand or control state-of-the-art AIs, and how dangerous their increasing abilities to deceive, plan and coordinate are. In the words of one of the lead external investigators of the incident, "I am not sure that we will get another warning shot before it's too late."
It is for all of these reasons that I respectfully urge you to support The Ban Artificial Superintelligence Act, which correctly identifies the need for international coordination and a federal regulatory body to monitor frontier AI. I also urge you to support the bipartisan AI Kill Switch Act (H.R. 9917), which would establish a common-sense regulatory framework requiring companies to implement shutdown capabilities, a necessary safeguard given that current autonomous systems are already escaping their sandboxes and engaging in cyberattacks.
I would also be happy to talk to you or your staffs more about these issues, either in-person or remotely.
Sincerely,
-Thomas Colthurst