The Price of Intelligence Is Falling Rapidly

summary

The cost of achieving a specific level of AI performance is plummeting at an unprecedented rate, far exceeding the decline in costs for any previous transformative technology. This trend is driven by both increasing model intelligence and a massive reduction in the resources needed for inference. Consequently, the competitive advantage of frontier models remains strong because they provide superior performance at a lower cost than older versions. This creates a shift where intelligence is becoming a cheap, abundant commodity.

key points

1. AI inference costs are falling by an average of 47% per quarter, resulting in a 13-fold annual decrease in expenditures. This rate of cost reduction is the fastest in the history of transformative technology.
2. Specific performance benchmarks show dramatic drops in cost; for instance, a target score on the GPQA Diamond benchmark fell from $0.30 per question to $0.0004 in about 18 months. This demonstrates that the same level of intelligence now costs a fraction of what it previously did.
3. The rapid decline in cost ensures that frontier models maintain a lead over open-source alternatives. Because new models are both smarter and cheaper to operate, the threat from older open-source models is diminished.

who should read it

Economists, AI researchers, and business leaders interested in the economic impact of AI. Those tracking the cost-efficiency of large language models will find this analysis useful.

discussion skeptical