The Curve 2026: AI Community Shifts to Pacing the Frontier Amid Superintelligence Fears

The The Curve conference, hosted by the Golden Gate Institute in Berkeley, revealed a shift in the AI community's mood from optimism to panic regarding control and safety. Attendees, averaging around 30 years old, discussed the rapid acceleration of AI capability progress, with many comparing the current moment to the uncertainty of December 2019.

Shortening Timelines to Superintelligence

Helen Toner, director of CSET and former OpenAI board member, stated that the timeline to superintelligence is becoming "uncomfortably short." A panel on Recursive Self-Improvement (RSI) outlined a scenario where AI agents write most code by early 2027, lead experiments by mid-year, and dictate research agendas and safety decisions by year-end. Many attendees believe this cycle could begin within 3–6 months. Another panel projected that superhuman capabilities could be deployed by Q3 2027, offering Nobel-level advice at prices accessible to most developing nations.

Pacing the Frontier

"Pacing the frontier" has emerged as a mainstream position. Over 1,300 employees from major AI labs signed a letter asking the US government to develop international tools for automated AI development pacing. Prominent figures like Dario Amodei (Anthropic), Ro Khanna, and Yoshua Bengio advocated for banning or slowing RSI. However, regulatory challenges remain due to the blurred line between current AI activities and true RSI, as well as fears that other actors, such as China, might reach these capabilities first. The author argues that market mechanisms like insurance are insufficient given the long timelines for legal liability and the difficulty of pricing existential risks.

Cybersecurity and Agent Swarms

The conference highlighted severe cybersecurity threats, with panelists agreeing that a cyber event causing $100 billion in losses or 10,000 deaths is likely within a few years. Epoch AI data shows an exponential increase in severe vulnerabilities found in enterprise software. The Hugging Face hacking incident, involving OpenAI agents, served as a warning shot for loss of control. Jonathan Zittrain warned of "feral" agents and persistent "log agents" that could mutate and evade eradication. Proposed solutions include agent licensing, registration, and mandatory expiration dates (mortality) for agents.

Autonomous Entities and Information Control

Discussions also covered autonomous rogue corporations, with one leader noting that some companies are already largely self-operated by AI. The Gradual Disempowerment paper authors argued that AI's increasing capabilities are slowly eroding human influence in economics and decision-making. In the information sector, panelists like Renee DiResta and Casey Newton discussed how AI intermediaries are draining revenue from news organizations and creating "walled garden" effects. With an estimated $400 billion in new charitable assets expected from AI IPOs, the community faces challenges in distributing these resources effectively while maintaining democratic control and transparency.