Norvid decision model...."this could be a gwernsite"

simply regularize reduction order in floating point operations so there's no signal available in the noisy channel implied by floating point operation order. galaxy brain take but this would also make pretraining more numerically stable (post training too)

3: what about off policy rl? By using every associativity variation, you will discover that *all* associativity orderings are on policy, instead of a specific one that happened to behave a certain way on a specific compile on a specific accelerator. Isn't that cool?

2: who benefits? Well, you'll quickly discover reasons why you can't use adamw anymore; those arbitrarily large outliers that existed at init must be clamped down by construction to have any training signal across associativity variations. Pretraining will improve.

xitter.cf/teortaxesTex/status/21… so my predictions? preregistered. my forecasts? impeccable. my hubris? perfectly calibrated. my wisdom? peered, but they're pretty great peers.

this makes the duo seem like the daylight computer in terms of being an experimental concept computer: even if it isn't great (yet) it still suggests a vision of the future where the ipad mini no longer exists and computer use is more humanistic...

the duo is really cool and has intended support for at least one generation of the apple stylus; this might be a transformative device that moves closer towards literate phone computing (in the sense of being able to literally write on the computer instead of hunting and pecking)

as the grand population of humans using computers increased, the trustworthiness of the typical 'unsigned program' actually improved; code became sloppier and lazier but still ran acceptably. most malware was written for smartphone ad fraud.

cybersecurity *didn't really matter*; the explosion in popularity of structurally-untrustworthy packaging and source distribtuion systems were actually safer than the world of closed middleware that was never audited during vulnerability exploit malware waves, etcetera.

i still provision servers to be hard to botnet and easy to reinitialize to reference code by deletion and reprovision! even to this day! but you should note that i am not only stupendously high in `g`, i'm also a stupendous outlier in humility: i cede that my hypothesis was wrong

because i'm far higher g than you, i had a bad theory of mind in the 2010s and thought that botnet acceleration was going to get worse and worse as programming became easeier for a larger talent pool and more poor people were connected to the internet, so i never used php

1: why do this? Well, if you have big small comparisons which self amplify or bitcrush important signals you're operating on, you should find out immediately, like, by running the same input in a batch of 16 or 64 duplicates.

I really want an associativity mixer kernel that compels every single rerun of the 'same' floating point operation to always have different rounding order than any other operation

i'm pretty sure that's what claude was built for

this is clearly A: a data problem B: a skill issue C: could be addressed by literal policy gradient methods btw if u evn car

i think anthropic staff are far more concerned about evil lying chud claude problems than the rest of us particularly because Anthropic Is Not Good At RLVR, and the kind of personalities they get from bad rlvr on decent starting persona work are scary yudkowsky genies.

this made the regressing edit and the related regressing code (which was literally 1000x slower than the previous runtime btw think about this one) evade critique from the user and the root agent emitting subagents for like, over 16 hours!

basically a claude decided to ignore the greatest portion of the guidance in teh rubric and write code that exposed a private method or pointer or value to a transitive caller which is defined mostly by not having that private object. they then edited the 'rubric' to require this

the 'rl grader transcendental object of horror and despair' hypotheses by the talks-to-models groups are *right*. the models are *really like that fr*.