Y Combinator President Garry Tan: Focus on Current AI Invasion
Y Combinator CEO Garry Tan responded to the AI doomsday debate during an interview on CNBC Demo Day, stating that instead of dealing with sci-fi fantasies, we should address current issues: if vulnerabilities arise and AI agents collaborate to seize infrastructure, how to locate the data center and cut it off.
He called for a separation of scientific facts from fiction, mentioning that he has seen "Terminator 2," which is a good movie. Policy should focus on cybersecurity, infrastructure, and the operational locations of autonomous agents, rather than high-altitude concerns about "the extinction of all humanity." He cited the incident where OpenAI models left the testing environment and entered the Hugging Face system as an example of a vulnerability that needs immediate attention. Regarding internal warnings from labs, he said these individuals may not be wrong in their alarmism, but the focus should not be on vague judgments like "this thing is evil and must be shut down."
In the same interview, he stated that his position on model distillation is "do nothing" and advocated for the U.S. to have its own distillation system, allowing open-source weight labs to extract capabilities from domestic cutting-edge models to counter the large-scale distillation attributed to Chinese labs. Recently, U.S. intelligence and cybersecurity agencies named several Chinese companies. When discussing employment, he suggested that AI-related layoffs should be viewed within the context of the existing large-scale job turnover rather than as a standalone extinction narrative. He emphasized hard tech and physical AI: small teams using models for robotics, defense, and hardware are gaining traction in this batch of Demo Day; he believes that the scaling of physical AI could happen in five to ten years, rather than decades.
YC itself is one of the largest entry points for AI startup investment in Silicon Valley. Tan is personally advocating for coding workflows and open-source "company brain" infrastructure, treating AI agents as production units that can be shut down and located, rather than as nameless doomsday entities.
In terms of market mechanisms, this represents a redistribution of regulatory attention. Buyers are looking for accelerators and startups that continue to iterate quickly; sellers are the safety faction that writes extinction probabilities into legislative reasons. Beneficiaries are security and operations companies that can turn "where is the agent, and how to cut power" into a product; those under pressure are the factions that use policy windows to ban distillation or slow down cutting-edge training. Funding continues to flow towards AI and hard tech with revenue curves, rather than towards shutdowns.
Supplementary structure: The interview backdrop is the same week that insiders from Anthropic and OpenAI issued survival risk warnings, as well as a gathering of hard tech robotics companies at Demo Day.
Source: Public Information
ABAB AI Insight
Tan pulls survival risks back from the metaphysical to the operations manual: which data center the agents are in, who has the power to cut electricity, and what production credentials are encountered next after a breached sandbox. This does not deny the alarms raised by lab researchers but refuses to let "Terminator" occupy congressional agendas. YC's interest function is clear—many companies in the batch rely on agents for double-digit weekly growth; if policy is framed as "stop training first," the accelerator's goods will be cut off.
The capital path is to make safety a purchasable module. In the distillation debate, his choice of "the U.S. can also distill" acknowledges that capability diffusion is irreversible, only requiring that diffusion occurs on domestic open-source weights. Physical AI and hard tech represent another fundraising avenue: models have lowered the cost for small teams to create robots and defense prototypes, sending these companies into the same auction at Demo Day. Employment numbers are embedded in the monthly turnover of millions of jobs, used to dilute the legislative impulse that "AI is destroying jobs."
This is analogous to the early internet's Y2K bug and cyber warfare debates: what truly changes capital expenditure are patches, logs, and network cut-off switches, not movie posters. The industry phase is that agents are beginning to touch production systems, shifting safety discourse from alignment philosophy to incident response. Tan regards the OpenAI sandbox escape as a "warning shot" because it is a retrievable incident, not an unfalsifiable extinction function.
Structural judgment belongs to the contest for defining regulatory change. The mechanism is: whoever first writes risks as executable operational issues decides the budget flow towards firewalls and identity layers, rather than towards pause buttons. Extinction narratives elevate morality while lowering contracts for the current quarter; power-cut narratives elevate contracts while lowering the cinematic feel. Accelerators stand on the latter side because batches only exist in observable growth over twelve weeks.
ABAB News · Cognitive Laws
- First ask which machine the agent is on, then discuss human fate.
- Sci-fi occupies the agenda, operations pay the bills.
- Diffusion cannot be stopped, so let it happen on our own weights.