in hindsight it's obvious that you could build a system to make a distribution move itself to a target with simulated data and better-than-garbage selection

exactly so

I've had a few like this. The only thing that's worked is to let them dig themselves into a hole so that we can have a frank conversation about learning. Not very satisfying

> that's my secret
> cap

🧐

huh. now I wonder if the cloud is symmetric wrt the dead tokens. like are they getting moved into a particular corner?

this kind of behavior makes sense to me in ai/ml, where fomo appears to be growing faster than aggregate prose output and much faster than actual experience or taste can develop

much of this feels downstream of academic impulses to collect stars and defer to authority. as evidenced by my peer-reviewed position paper on arxiv, my opinion has been blessed and is more scientifically valid than your ~/user/position.md

I see... is there a meaningful distinction between the dead tokens moving vs everything else moving?

"We never needed LLMs to generate an exponentially growing stream of boring prose." no notes

is that the same cohort of tokens, or does the composition change as some receive updates?

makes sense

show, 2 seasons. originally paramount+ I think. decent schlock

I am tbqh shocked their gear wasn't stowed

"interesting question!" cortana is AI confirmed

halsey is a standard kind of twisted genius. cortana so far is very passive. the characterization overall is a bit thin. decent schlock

john is an idiot