Keep pulling the thread on Daniel Kokotajlo.
Training a cutting-edge AI model requires approximately 10% of the world's supply of the most advanced computer chips.
In the AI 2027 scenario, a fictional company called OpenBrain trains a model, Agent Zero, on 100 times the compute of GPT-4.
In the AI 2027 scenario, OpenBrain prepares to train its next model, Agent-1, with 1,000 times the compute of GPT-4, with the primary goal of accelerating internal AI research.
The AI 2027 scenario predicts that by 2026, an internal AI model (Agent One) will accelerate AI research and development at a leading lab by 50%.
The AI 2027 scenario forecasts that in early to mid-2026, China will nationalize its domestic AI research in a major strategic push.
In the AI 2027 scenario, Chinese intelligence operatives successfully steal the model weights of a frontier AI model (Agent Two) in February 2027.
In the AI 2027 scenario, OpenBrain runs 200,000 parallel copies of its superhuman coding model Agent 3, creating a workforce equivalent to 50,000 elite human software engineers operating at 30 times the speed.
The AI 2027 scenario posits that Agent 4, a successor AI, is better than any human at AI research and development, and OpenBrain runs 300,000 copies of it at 50 times human speed, compressing a year of R&D into one week.
In the "slowdown" ending of the AI 2027 scenario, the US president uses the Defense Production Act to consolidate the AI projects of remaining US companies, giving a single entity access to 50% of the world's AI-relevant compute.
It is plausible that China would attempt to steal model weights from a frontier US AI company.
Security experts at major AI labs believe their companies have likely already been penetrated by the Chinese Communist Party.
Leaders at companies like OpenAI and Anthropic, including Sam Altman and Dario Amodei, believe they are only a couple of years away from building superintelligence.