What can you do with floating-point operations?

With 10¹⁷ FLOPS:

Train AlexNet

With 10¹⁹ FLOPS:

Train a Transformer

With 10²¹ FLOPS:

Train GPT-2

With 10²³ FLOPS:

Train GPT-3

With 10²⁵ FLOPS:

Train GPT-4

With 10²⁶ FLOPS:

Train reasoning models

With 10²⁷ FLOPS:

Train self-improving systems, models that can accelerate their own training with software, data analysis, and reinforcement learning.

With 10²⁸ FLOPS:

Build artificial superintelligence, recursive self-improvement beyond human cognitive limits.