assume an expert in any field of knowledge charges $200/hr. given how much it costs in dollars to train models like opus or gpt5, how many hours of expert time is that worth?
This is a fun back-of-the-envelope calculation.
Estimated training costs for frontier models:
- GPT-4 (2023): estimated $50-100 million
- Claude 3 Opus / GPT-4.5 class models: likely $100-300 million range
- GPT-5 / next-gen models: rumors suggest $500 million to $1+ billion
At $200/hour expert rate:
| Training Cost | Expert Hours | Full-Time Expert Years |
|---|---|---|
| $100 million | 500,000 hours | ~240 years |
| $500 million | 2.5 million hours | ~1,200 years |
| $1 billion | 5 million hours | ~2,400 years |
(Assuming ~2,080 working hours/year)
Some perspective:
That $100 million model represents roughly what you'd get hiring 240 experts working full-time for their entire careers. A billion-dollar training run is equivalent to about 2,400 expert-lifetimes of labor.
Of course, this comparison is imperfect—training costs include compute, not just "intelligence generation," and an expert's value comes from judgment, accountability, and real-world action, not just information retrieval. But it does highlight why there's so much interest in whether AI can substitute for or augment expert knowledge work.
No comments:
Post a Comment