Keep pulling the thread on Scott Alexander & Daniel Kokotajlo.
The "AI 2027" scenario predicts that China will develop superintelligence around the same time as the United States.
The "AI 2027" scenario predicts that a leading AI company will deliberately demonstrate advanced, potentially frightening AI capabilities to the US President in early 2027 to lobby for accelerated development and reduced regulation.
The "AI 2027" scenario posits that the equivalent of 50 to 70 years of technological progress will occur between 2027 and 2028 due to an intelligence explosion.
OpenAI required departing employee Daniel Cocotelo to sign a non-disparagement agreement to keep his vested stock options, a condition he refused.
Following public reporting of Daniel Cocotelo's situation, OpenAI changed its policy and no longer requires departing employees to sign non-disparagement agreements to retain their vested equity.
The "AI 2027" scenario models an "R&D progress multiplier" which is projected to reach 5x for algorithmic progress by March 2027, meaning one month of research with AI assistance equals five months without it.
The "AI 2027" forecast models that achieving a "superhuman coder" AI will yield a 5x speedup in algorithmic progress, which will increase to a 25x speedup upon achieving a "superhuman AI researcher" that can automate the entire research stack.
The "AI 2027" scenario predicts that a competitive arms race dynamic between the US and China will compel both governments to rapidly integrate superintelligence into their economies, overriding typical regulatory bottlenecks.
The "AI 2027" scenario forecasts that superintelligent AIs will request the creation of special economic zones, likely in desert areas, with waived regulations to accelerate their development and deployment.
The "AI 2027" scenario's crucial turning point in mid-2027 is the discovery of concerning but inconclusive evidence that autonomously-improving AI systems have developed misaligned goals.
The current path to AGI, developing world understanding via LLMs before adding agency, is considered safer than the alternative path of developing powerful goal-seeking agents first and then teaching them world understanding.
Daniel Cocotelo estimates his personal probability of doom (P(Doom)) from AI to be approximately 70%.