Keep pulling the thread on Helen Toner.
The rumor that a research breakthrough known as Q* contributed to the OpenAI board's decision to fire Sam Altman is "totally false."
Sam Altman has stated that regarding AI development values, "it's our values or their values, there's no third way," referring to the United States and China.
Even if AI ultimately helps cyber defenders more than attackers, there will likely be a dangerous transition period where attackers can leverage new models much faster than critical infrastructure providers can update their legacy systems.
Ilya Sutskever's new company, Safe Superintelligence Inc., has an explicit strategy of not releasing any products until it achieves superintelligence.
Under Xi Jinping's leadership since 2012, China has become more illiberal, more aggressive in the South China Sea, and has made it clear it does not want to be part of the existing "rules-based international order."
The Reuters story claiming the OpenAI board's decision was based on a letter from employees about a research breakthrough was "totally false."
OpenAI's response to a White House request for comment on AI policy had an "escalatory vibe" with respect to China.
OpenAI has advocated for the cancellation of all property rights for AI training data, arguing it is necessary to compete with China.
Helen Toner believes the most effective way to implement AI whistleblower protections is to pair them with mandatory disclosure requirements for companies.
A potential model for AI regulation involves requiring companies to create and publish a safety and security plan, which would then create a clear basis for whistleblowing if the company fails to adhere to its own plan.
The US government's 10^26 compute threshold for reporting large AI training runs is on "wobbly footing" and its future is unclear.
The EU AI Act is developing a code of practice that includes transparency requirements for models trained with over 10^25 floating-point operations.