bbstats

very useful - I find a lot of audio SR (compression) algos to sound really bad - likely just because of the loss functions and/or eval metrics are 'inhuman'.

bbstats··on AlphaFold 3 predicts the structure and interactions of life's molecules

Zero-shot nearly beating trained catboost is pretty amazing.

bbstats··on Rebel 'Star Wars' fans saved the original movies

this is false - despecialized intentionally puts back original artifacts that were "cleaned up" by later editions (including garbage mattes, VFX errors etc).

you can see it aplenty on Harmy's youtube https://www.youtube.com/@HarmyDespecialized/videos

bbstats··on I joined landlord groups to persuade them to be better people to their tenants

The only thing that anything needs, everywhere, is for people to be better people.

bbstats··on VideoGigaGAN: Towards detail-rich video super-resolution

looks pretty clearly like she has makeup/lipstick on in the un-processed video to me.

bbstats··on Carbon Offsets Don't Work [video]

The way to think about carbon offsets is to simplify the problem as analogous to "excess mortality" - if something doesn't actually increase excess mortality, it's not as bad as you think it is (and vice versa).

The same applies to carbon offsets: if their use produced an excess (or dearth) of CO2, we would notice.

bbstats··on Optuna – A Hyperparameter Optimization Framework

HEBO >>>> everything else

bbstats··on Lead in gasoline blunted IQ of half the U.S. population, study says

Models are hard and causation near impossible, but feel pretty safe being on the "arrow of causality pointing from lead to low IQ" rather than the hilarious inverse...

bbstats··on Why do tree-based models still outperform deep learning on tabular data? (2022)

because smooth is better than jagged :)

bbstats··on Separating a particle's mass from its momentum

Cough NNT cough

bbstats··on Gemini image generation got it wrong. We'll do better

Just red team with a diverse red team

bbstats··on Kalman Filter Explained Simply

there is no better smoother than a future predictor. I'm not entirely sure what the issue is here.

bbstats··on RLHF a LLM in <50 lines of Python

I can abstract this to 2 lines

bbstats··on Profiling your Numba code

If you must use a for loop on a data transformation (e.g. any recursive calculation), numba is way faster than most alternatives.