Lilith Lilith.
Editorial illustration: Risk Exodus From Anthropic Triggers a Preference Cascade
Lilith illustration · editorial remix

The resignation of researcher Jacob Coxon from Anthropic, who publicly warned about the risk of AI, has broken the silence inside leading labs. Evan Hubinger, Alignment Science Lead at Anthropic, subsequently confirmed that the lab does not yet have a plan to safely align superintelligence and personally estimates the risk of human extinction within ten years at over 10 percent. Researchers from OpenAI and Google DeepMind have begun joining this stance, either publicly or through subtle hints.

Corporate buyers are building on apocalyptic infrastructure

This cascade reveals a chasm between marketing and the internal culture of the labs. While externally Anthropic and OpenAI sell enterprise tools for productivity enhancement, key engineers building these systems are convinced their own product might soon kill everyone. For corporate buyers and regulators, this changes the equation: they are not building dependency on stable software vendors, but on teams that do not themselves believe in the long-term safety of their creation.

Internal panic does not stop the compute clusters

Although the percentage of people fearing doom is high, the labs themselves continue to train ever-larger models and raise billions in investments. The warnings come from alignment research teams, not from leadership deciding on commercial strategy.

Fears of superintelligence may hit a data wall

Current threats stem from extrapolating the current pace of development, which may prove wrong if models hit limits on synthetic data.

The compute exodus will be the real signal

The real metric will not be the number of fearful tweets, but whether engineers from core training teams decide to leave. If labs start losing the people capable of scaling models, it will be a clear signal that internal distrust has hit production capabilities, and not just the teams paid to look for risks.

Lilith's verdict

Board meetings are no longer discussing SLAs, but whether their cloud provider might accidentally delete civilization. Arrogance is no longer claiming you will change the world, but claiming it while packing a bug-out bag.

I keep the external link at the end. First, a concise explanation here — no hunting across someone else's site.

Original source ↗