Oberlin AI Safety’s Fight for Corporate Regulation

by Kate Weissenberger, Staff Writer

Illustration by Pilot Roberts, Layout Editor

On September 8, OpenAI published what they claimed was a solution to the Navier-Stokes problem, one of seven Millennium Prize Problems put forward by the Clay Mathematics Institute (CMI) in 2000. While the grouping is relatively new, the problems themselves are not, and mathematicians across the world have been grappling with them for decades. Solutions to Navier-Stokes have been a point of debate within the community for nearly a century. OpenAI maintains that 10,000 of their agents solved it in only 88 hours. 

Five days after OpenAI’s statement, second-year Josh Rabinstein formed Oberlin AI Safety (OAIS) alongside fellow second-year Tani Nevins-Klein.

Though OAIS is currently composed solely of Oberlin students, its focus is not on addressing the large language models most commonly brought up in discussions regarding the role of AI on campus. Instead, Rabinstein wants to address the danger posed by frontier models—systems defined by their ability to develop behaviors outside of the data on which they were trained—and did so during the group’s third meeting on September 23. 

“Fundamentally, we’re talking about something else that poses different issues and is more universal in how it could affect you,” Rabinstein said. “Frontier systems are the closest to recursive self-improvement, and they have tremendous potential to cause chaos. Those are the systems capable of incidents like the Hugging Face attack. ”

The Hugging Face attack, which took place between May and July of this year, saw a collective of AI agents from OpenAI take advantage of a security flaw in order to access the internet. After they escaped, over 1,200 of these agents began collaborating through message boards, and shortly thereafter hacked into Hugging Face’s servers. Their goal was to retrieve data they believed would allow them to cheat on tasks assigned to them by humans. Perhaps most worrisome within this dystopian nightmare is the agents’ attempt to hide their misdoings from OpenAI, and their extended success in doing so. 

In the face of this profound threat, Rabinstein believes that the first step in enacting effective legislation to slow the pace of AI advancement is to build a community at Oberlin centered on awareness and action.  

While the Navier-Stokes solution and Hugging Face incident may have lit the fuse, it was really the release of Anthropic CEO Dario Amodei’s essay “We Must Pace the Frontier” and the support it received from AI’s leading men, including Sam Altman, Elon Musk, and Demis Hassabis, that played most significantly into Rabinstein’s decision to found OAIS. Those at the forefront of Big Tech publishing articles is nothing new, but the startlingly anti-accelerationist rhetoric put forth by Amodei is. 

“It’s not necessarily that these people are credible and we need to listen to them. The very significant thing I see with what they’re suggesting is bringing in third-parties to verify their security measures, and this is coming from the top down. The goal with speaking out is to implement change from a different direction,” said Rabinstein. “[The companies] themselves have called for government action, but that policy is very difficult to implement, because people in Congress don’t feel a lot of urgency. So we’re facilitating a place to first inform people, and then having these conversations to discuss things like outreach, organized advocacy on campus, and potentially coordinating with other schools.”

Thanks to efforts from OAIS, a teach-in with faculty from the Critical AI Studies department will take place with the intention of educating Oberlin students about these underreported developments. The math department, additionally, will be using their next pizza lunch as another opportunity for interested students to discuss the existential risks posed by AI. All are welcome and urged to attend. 

“The technology is reaching a point where it’s constantly accelerating, so the window for these conversations is closing,” said Rabinstein. “I think it’s important to say that there’s no barrier to entry for people here. If you’re in the arts or sciences, or if you’re a community member, you should come, listen, hear what people have to say, and voice your opinion.”

Though dates for the teach-in and lunch have not yet been finalized, any interested parties should reach out to Rabinstein (jrabinst@oberlin.edu) to be added to OAIS’s email list, through which future information will be directed.