Ovnius


251219

Don't Fear the Socialist AI

from Seb Krier

If you believe that capitalism is better than socialism because a centrally-run government lacks access to the knowledge to successfully manage the economy, you needn't be afraid that a single AI will take over the world.

A super intelligent AI would still have the same problem as a socialist government — the knowledge necessary to be effective doesn't exist in a single place.

So, such an AI would not be capable of suddenly running everything in its own fashion.


251218

AI and the March of Nines

In the Dwarkesh Patel podcast, Andrej Karpathy says that AI development is a "march of nines".

If it takes 1 year to get to 90%, it will take 1 year to get to 99% and another full year to get to 99.9%.

I had not heard of this but it's useful in a lot of contexts, even just a simple programming task.


Developing Pwika, my SVG-based website builder, went like this and it always mystified me.

So it might be helpful, when begining a project, to declare how many nines will be targeted.


It might be helpful for thinking about how fast AI will arrive more generally.

My impression is that we're at about 60% — if I ask ChatGPT to help me with something, it frequently fails.

So we're still not even at 90%. Perhaps it will take five more years to get to 90% — so let's say 15 years for 90%

Then we can expect another 15 years to go from 90% to 99%. A very slow takeoff.


Think about self-driving cars — I was driving near my house the other day, and there was construction.

I arrived at a T intersection:

This kind of situation, at least superficially, requires understanding what you're seeing, and I have to imagine that a car that can automatically navigate it is still far in the future.