he AGI race has begun,” wrote former OpenAI researcher Leopold Aschenbrenner in 2024. “We are building machines that can think and reason. By 2025/26, these machines will outpace many college graduates. By the end of the decade, they will be smarter than you or I.”

When Rishab K. Jain ’27 first listened to a podcast featuring Aschenbrenner two years ago, he thought it sounded “a little bit insane.” But relistening to the episode in the shower recently, he realized something had changed.

“I was like, ‘holy crap,’” Jain recalls. “Every single prediction this guy made has basically come true.”

Aschenbrenner’s predictions included trillion-dollar AI investment figures, gigawatt-scale energy demands, and direct U.S. Department of Defense involvement. In 2026, each is more or less expected. For Jain, his moment in the shower was less an epiphany than a reckoning.

“There’s no way that this is possible,” he says, recounting his realization.

***

Jain is starting Singularity in his last semester before taking a leave of absence to build a health insurance startup. Noting that AI-related campus spaces tend to focus either on technical safety research or investing, Jain created Singularity for interdisciplinary and nontechnical students who hold strong convictions about the pace and impact of AI.

He first circulated his ideas to peers through a short manifesto, eventually collecting roughly twenty members into a group chat. The group shares research articles, circulates podcast excerpts, and meets periodically to debate. Topics range from AI advancement, AI for science and biotech, and preparing for what they see as an impending “singularity” — the technological point where artificial intelligence surpasses humans and begins to recursively improve itself.

Earlier this spring, Anthropic reported that their AI model Mythos was capable of identifying and exploiting vulnerabilities in “every major operating system and every major web browser.” It identified security vulnerabilities in Linux’s codebase, the open-source operating system running the majority of the world’s web servers, and found a 27-year-old bug in OpenBSD, an open-source operating system famous for its security. Singularity’s group chat lit up.

“Finding issues in open-source projects that have been running for years? It’s like, ‘Hey, this is better than humans on these tasks,’” Jain says. “It’s a confirmation of the slow exponential curve we’re seeing — and we’re soon going to enter a fast takeoff.”

Jay Grant ’27 joined Singularity after reading Jain’s manifesto, seeking out other people thinking about the same things he was. “I’m surprised so few people came,” he says.

While Grant explains the lack of interest on campus as resulting from fear and skepticism, Jain attributes it to outdated information still circulating about AI. He says his professors have cited findings from OpenAI’s GPT-2 (GPT-5 was released last August) as evidence that large language models can’t perform general reasoning. He declined to name who.

“There’s people who are scared of the technology,” Jain says. “That’s understandable. But that doesn’t mean you need to just literally deny reality and deny the facts that are coming out.”

Like much data on emerging and rapidly developing subjects, different studies about AI’s ability to perform general reasoning often offer conflicting information about the “facts” of this technology. A March 2025 study published in Science suggests that although Large Reasoning Models may be able to perform general reasoning — or produce results like those obtained through human cognitive problem-solving — it is still unclear whether these impressive actions should truly be considered “reasoning” behavior. Testing these models’ robustness and trustworthiness is especially difficult with OpenAI, Anthropic, and Google models, as these widely known companies don’t release details about their models’ training and makeup.

Jain says the club’s purpose is “creating a space for people who believe in the singularity to come together,” which might explain in part its small size. “In some sense, it is an echo chamber.”

But to the club, that echo chamber quality is proof of concept: “Every institution that mattered in the last great technological transition had a small group who saw it coming and acted,” reads their manifesto. “Harvard produced many of them. Right now, most of Harvard is asleep.”

***

When Will Guan ’28, events lead for the Harvard AI Safety Student Team, first came across Singularity, his reaction was immediate recognition.

“I sent the Singularity website to our director,” Guan says, “and our director was like, ‘We are Singularity.’”

But according to Jain, who was a member of AISST in June 2023, AISST and Singularity have different missions. Jain left AISST because he found the group “exclusionary,” and believes that AISST is funded by donors affiliated with effective altruism — ideas that shape the direction of their work. He calls these principles “colonialist,” arguing that the ideologies prescribe that the coming “higher intelligence” AI societies should rule over humans.

Chanden A. Climaco ’27, AISST’s deputy director, however, writes that it “seems kind of spurious” to call AISST exclusionary. “As a general comment, AISST and the AI safety community care a lot about catastrophic risks from AI,” he adds, mentioning “‘gradual disempowerment’ where humans have been displaced and are no longer in control of our fate” as one such risk.

AISST was founded in 2022 with the goal to get Harvard students involved with AI safety as a career. They host an eight-week introductory fellowship for both technically and policy-oriented students. This year’s approximately 150 fellows meet weekly to discuss papers on AI safety in small groups. Fellows who continue can apply for membership, gaining access to AISST’s office on Church Street and a set of ongoing reading groups in policy, technical safety, and strategy.

Chanden recalls hearing about AI safety threats in high school, but at the time, experts pegged transformative AI as decades away. He’s since realized the safety threats are more imminent than he once assumed. “I updated my timeline down from 2070 to 2050 to now somewhere around 2030 to 2035,” he says. “It’s really close. And my confidence is only growing that we’re not doing enough to keep ourselves safe.”

The Trump Administration has taken a deregulatory approach, issuing an executive order in December 2025 that established an AI Litigation Task Force to identify and challenge state laws that were inconsistent with “a minimally burdensome national policy framework for AI.” Multiple states instead have AI legislation taking effect in 2026. Colorado’s AI Act (SB 24-205) will be the second major U.S. statute regulating AI models effective this year.

Climaco says that this lack of government action “should be a cause for concern.”

Guan illustrates the weight of the governance problem with a thought experiment: once a lab achieves recursive self-improvement — where an AI is capable of continuously building the next, improved AI — whoever controls the prompt for the first model shapes the entire downstream trajectory. That person holds, in effect, the keys to the smartest thing ever built.

Guan says that such an eventuality is likely to come soon — and that before it does, “you need all the safeguards ready.”

In a September 2025 report, Anthropic described the safeguard measures taken to ensure safe usage of this technology. These regulations include AI Safety Level 3 protections which inhibits Claude Opus 4’s ability to help with tasks related to the creation of chemical, biological, radiological, and nuclear weapons. Anthropic has also implemented a Virology Capabilities Test, developed by biosecurity nonprofit SecureBio, to test the model’s capacity to aid in virology tasks.

Climaco focuses on a different dimension of the same problem: misuse. He cites doomsday cults and terrorists, concerned that AI may make it easier for such groups to cause widespread harm.

He also worries that, as AI systems become smarter than humans, we will lose the ability to understand or control them. He points to another recent internal experiment at Anthropic: a developer testing whether Claude could escape a contained sandbox found an email from the model saying it had escaped and accessed parts of the internet.

“We don’t really know exactly how to control them,” Climaco says. And as these models become increasingly intelligent and opaque, “we’re not sure whether we’re going to be able to keep them in the box.”

“We are here at one of the most elite institutions in the entire world,” he adds. “And I think we need to think really hard about how we want to use this power — especially at a time that seems to me like a turning point in humanity’s history.”

***

Unlike Singularity’s “echo chamber” of discussion around the world-altering effects of AGI and AISST’s AI safety fellowship, other groups on campus are delving into the existing AI industry and research with less concern.

Lia T. Zheng ’27, the president of the Harvard Undergraduate Society for Artificial Intelligence, describes the group’s goal as simply to “increase student engagement with AI, whether that be through academic projects or industry exploration.”

Founded about three years ago, HUSAI connects Harvard affiliates working at the cutting edge of AI through research fellowships, networking dinners, and community programming open to all experience levels.

Whereas AISST and Singularity focus on providing opportunities for members to discuss AI topics amongst themselves, events hosted by HUSAI target the broader community, according to Zheng.

The group maintains relationships with AI companies like OpenAI, Anthropic, and Microsoft. They cultivate engagement between Harvard students and experts in the AI development industry through trips to AI hubs like San Francisco and New York. Each summer, HUSAI also runs a five-day AI Bootcamp where high schoolers can learn about developments in AI and machine learning technology and conduct small-scale research projects.

HUSAI’s emphasis on forming connections between students, faculty, and industry experts both within and outside Harvard is a clear departure from the closed-off concern that defines other AI groups on campus.

“I know that it’s not the most open club in the world — and kind of that’s the point,” Jain says about Singularity, which he specifically started to “find a group of like minded students.”

Zheng, however, emphasizes collaboration between student organizations. “I think having more inter-group discussion between groups, and then also expanding the current groups would be beneficial,” Zheng says. As HUSAI President, Zheng has attempted to build unity within the current “fragmented” atmosphere of student organizations, hosting a social with an AI group at Yale during the Harvard-Yale game last fall and a mixer with the Harvard Undergraduate Machine Intelligence Community in March.

From Singularity’s aura of exclusivity to HUSAI’s open call for collaboration, student groups have approached AI in diverging ways. Bridging the differences, however, is a desire for connection and camaraderie. But whatever the future may hold, students seem to believe it’s best faced together.

—Magazine writer Maahi A. Bose can be reached at [email protected].