Hacker News new | ask | show | jobs
by ThrowawayR2 1114 days ago
They are. Patterson & Hennessy at their Turing Award lecture back in 2017, when they started talking about where future opportunities for increased speed were going to come from, the very first thing they took a (polite) dump on was how software was written very inefficiently: https://www.youtube.com/watch?v=3LVeEjsn8Ts&t=2183s

The old canard that programmer time was worth more than processing time that bad coders quote is less and less true by the year; SWEs who are able to wring better performance out of less hardware resources are going to become more valuable in the future.
1 comments
oofnik 1114 days ago
> SWEs who are able to wring better performance out of less hardware resources are going to become more valuable in the future.

As much as I'd like for this prediction to come true, I think it's wishful thinking. Compute capacity continues to increase exponentially while (human) cognitive resources, for all practical purposes, do not, which means less efficient code will continue to be the more economically viable option as far as I can tell.

Even with AI, where we're essentially trading compute (used for running code) for extended cognitive resources (used for writing code), what incentive exists to channel those resources into writing faster code rather than writing the same code faster?

link
sfpotter 1114 days ago
It’s a literal waste of energy. LLMs draw a tremendous amount of power. They enable problems to be solved in a low effort, careless sort of way, but this neglects the power efficiency of our own brains. We can think through and implement better solutions consuming only a tiny fraction of the power. In a world where we’re running up against climate change and finite resource constraints, that is valuable, it’s just not self-evident yet. We’ll first have to crash headlong into a wall to realize it.
link
oofnik 1114 days ago
We can use AI to help us design and build a functional fusion plant, which we can then use to power more datacenters running ever bigger LLMs to help us design more efficient datacenters to run bigger LLMs to help us build ... yeah this is going to be one hell of a century.
link