Keep pulling the thread on Leopold Aschenbrenner.
The automation of AI research by Artificial General Intelligences (AGIs) is predicted to compress a decade of algorithmic progress, equivalent to 5 or more orders of magnitude (OOMs) of effective compute gains, into one year or less.
The automation of AI research will cause a rapid transition from human-level AI to vastly superhuman AI systems.
By the time AGI is achieved, inference GPU fleets will be large enough to run millions of AGI instances, equivalent to 100 million human researchers.
Superintelligence will be applied to R&D in fields beyond machine learning, solving robotics and causing dramatic scientific and technological leaps within years.
Superintelligence is likely to provide a decisive military advantage to whoever controls it.
An extrapolation of current AI capabilities suggests that AI systems could reach or exceed the level of the best human AI researchers by the end of 2027.
Global GPU fleets are expected to reach the tens of millions by 2027.
By 2027, AI training clusters are expected to be approximately 3 orders of magnitude (OOMs) larger than today, reaching over 10 million A100-equivalent GPUs.
The transition from Artificial General Intelligence (AGI) to superintelligence is plausibly expected to occur in less than one year.
A plausible timeline for an intelligence explosion involves proto-automated researchers emerging in 2026/27, leading to a 10x acceleration in progress by 2028 and superintelligence by the end of the decade.
By the end of the decade, GPU fleets numbering in the hundreds of millions will be able to run billions of superintelligent agents that can think orders of magnitude faster than humans.
The advent of superintelligence and full automation could lead to economic growth rates of 30% per year or higher.