Why technology usually improves exponentially

post by lemonhope (lcmgcd) · 2025-02-27T03:55:25.888Z · LW · GW · 0 comments

Contents

No comments

(Mostly taken from my old comment.)

Memes and genes are obvious enough, but why is the rate of technological improvement[1] proportional to the current technological level (or basically zero)? Don't ideas get harder to find?

Well Big Ideas do get harder to find, but if you make a 1% improvement to the US's steel production, then you get an extra 800,000 tons of steel. That doesn't help you think up new improvements but it does mean that the next 1% improvement will yield 808,000 tons.

Basically, any cost reduction or speedup or quality improvement is on top of what you have. How would you save a silicon foundry $500,000 flat, without saving them more money as they expand? Maybe you could get a one-time government grant or a one-time supplier discount. You have to do a lot of one-time things like this for it to add up to anything significant.

Let's consider an improvement that appears to be linear. Say you change the voltage regulator within a particular circuit to use 1 microwatt less. There's two reasons this might end up compounding. First, the total power consumption reduction is proportional to the number of times that circuit is used across all chips/devices. Second, if someone later finds a 1% power save across all transistors, then your little circuit will probably get that improvement too. It ends up being like a deposit into a savings account with interest.

If your savings account doesn't have interest, then you probably will never be a millionaire from small deposits. If some branch of technology hasn't found a few sequential compounding improvements then it probably won't go anywhere soon. But if folks keep making steady advances, then expect an exponential plot for widgets/$ over time or performance/widget over time.

(Note: the exponential increase in the cost of frontier AI models says little either way about the performance/$ of AI stuff over time.)

Realizing this has made me lean a bit towards Hanson's view that technology is not very lumpy. So has the history of the nanogpt speedrun. (Like widgets/$, you want to think about runs/hour, not hours/run.)

  1. ^

    You can't use "technology" as a y-axis; pick some more specific measure like "number of nails i can buy with one inflation-adjusted dollar" or "number of int32 multiplies I can do with $1 of compute" or "joules of electricity per $1" or "average life expectancy in country X" or "average travel time between city A and city B" or "how many 2018-coding-jobs can a single software developer do".

0 comments

Comments sorted by top scores.