site banner

Forecasting Newsletter August 2022

forecasting.substack.com

Highlights

I'd prefer comments or questions here on account of themotte.org site being pretty young. Long live The Motte!

2
Jump in the discussion.

No email address required.

A lot of stuff in the 30-60% range which means few have any idea

the "steelman" here is that it's evidence of some form that either it's up to chance or it's hard to predict.

They'll also claim that a 30% and a 70% really are different, and that's useful

afaict forecasting sucks because the reasons for believing things are complicated, and those reasons and understanding the complexities involved is much more worthwhile than trading for numbers. A SSC post >> a forecast! Forecasters aren't gonna succeed at guessing 'will the big image generation models be based on diffusion models or VAE or GAN' (and will fail much worse at 'what architecture will image generation models be using in 5 years' five years ago, or even at caring about image generation) because ... well, the people working on it didn't know either. Essentially making good predictions is as hard as understanding the thing being predicted, so sticking a layer of 'vote for number' on top of it gets you essentially nothing versus 'talking to smart people or people involved' or 'reading and writing papers' or 'doing research'