It seems like an important question that I haven't seen great analysis on is: if we stop or heavily pace frontier training runs and development, how much of the benefits (measured either in economics/welfare or straight capabilities) will we lose per unit time.
This is sort of vague and I would also like some help operationalizing this question. What I'm imagining is approximately as follows.
- We open source fable/astra weights (or imagine in a year some open source model gets to that level)
- You can still finetune the models (probably misusing this but let's say a "team" can use 1/100th or some other factor of the compute used to train them to update the weights per year)
- You can still update harnesses
- You can still update APIs
- You can still build new company structures or workflows
- You can still optimize caching, datacenter compute use, etc.
How much would this
- Lower gdp growth
- Lower capabilities growth
over (1,5,n) years.
