US Launches 'AI Manhattan Project' Aiming for Massive Computing Leap by 2027
Summary
The US government initiates an 'AI Manhattan Project' consolidating private sector resources and investing a fraction of GDP, aiming for a 2e29 FLOP training run by 2027, around 500 times larger than the most compute-intensive model to date and 10,000 times larger than GPT-4.
Key Points
- A US government-initiated 'AI Manhattan Project' consolidates private sector AI resources and invests a fraction of GDP similar to the original Manhattan Project or Apollo program.
- Such a project could yield a 2e29 FLOP training run by the end of 2027, around 500 times larger than the most compute-intensive model to date and 10,000 times larger than GPT-4.
- While ambitious, this scale appears feasible given recent evidence on supercomputer and energy scaling, with power needs potentially met through accelerated gas plant construction or requisitioning existing capacity.