ponder

ponder

@ponder.ooo

director of research @ponder.technology

3.7K followers266 following23.2K posts

ponder
@ponder.ooo

openai should invite nathan fielder and a camera crew to live in their office and document everything they do

ponder
@ponder.ooo

"now im not a... computer person.. but uh. if it's connected to 'artifactory' and that's connected to the internet, then isn't it, uh, connected to the internet?" "no. absolutely not. no internet. artifactory mediates the connection securely. we patched the vuln the models were using to get past it"

ponder
@ponder.ooo

it's also weird that pop explainers of math act like zfc has already succeeded at this and mathematical work always proceeds from the axioms of zfc. but also now are sometimes acting like lean (and thereby univalence or hott or whatever) has done so, that a lean cert is "the gold standard" in math

ponder@ponder.ooo

it's cool to have ppl working on foundations in math, encoding as much of the body of math as they can in whatever system they like. but if any such project succeeded in bringing about a future where everyone else doing math always had to operate in one foundational system, that would be horrible

ponder
@ponder.ooo

it's cool to have ppl working on foundations in math, encoding as much of the body of math as they can in whatever system they like. but if any such project succeeded in bringing about a future where everyone else doing math always had to operate in one foundational system, that would be horrible

ponder
@ponder.ooo

very much dislike openai's view that when results are obtained via processes involving sufficiently large proprietary RL models, the academic norm should be to leave all details of the process except the involvement of a proprietary model undisclosed and to leave most researchers involved unnamed

View full thread
ponder
@ponder.ooo

also the lean certs are the main results, not the generated preprints, especially not the preprints that not only lack lean certs but that nobody has meaningfully reviewed. those "preprints" should be released as attachments/supporting material alongside an actually-well-written writeup

ponder
@ponder.ooo

or rather *writeups*, of their full process for creating their internal model as well as their full process for obtaining the mathematical results they obtained with it

ponder
@ponder.ooo

very much dislike openai's view that when results are obtained via processes involving sufficiently large proprietary RL models, the academic norm should be to leave all details of the process except the involvement of a proprietary model undisclosed and to leave most researchers involved unnamed

View full thread
ponder
@ponder.ooo

they have insane amounts of compute, proprietary data, and human research talent. once the model capabilities presently only available to them become generally available, their norm of assigning all credit to the llm will become increasingly obviously untenable and deranged

ponder
@ponder.ooo

ofc by that time they will have even better models than that so they'll just keep doing the same shit, trying to pass off a massive material advantage in compute resources, data, & highly skilled intellectual laborers as something fundamentally different in kind from all other research

ponder
@ponder.ooo

very much dislike openai's view that when results are obtained via processes involving sufficiently large proprietary RL models, the academic norm should be to leave all details of the process except the involvement of a proprietary model undisclosed and to leave most researchers involved unnamed

ponder
@ponder.ooo

they should publish with full transparency about the full processes by which they achieve their results, and their researchers should be putting their names on that

ponder
@ponder.ooo

unbelievable how well this works. "correcting" unbounded orbits as soon as they exit the unit interval by just modding out by 1

ponder
@ponder.ooo

this seems like the "natural" extension in the +/+ quadrant but it obliterates some structure in the other 3 quadrants so i think the idea needs more work

ponder
@ponder.ooo

watching chess experts commentate ai chess games is funny bc they'll be rattling off all these ideas and consequences of why some move is clever or some other option isn't viable &c and then every once in a while they're just like "i dont rly get what this move is doing" and then just move on

ponder
@ponder.ooo

it's not enough to just have cameras and guards everywhere they have to insult u with the stupidest gaslighting imaginable. "this is for you actually, it's to improve your shopping experience, you love it, we wouldn't be surveilling you if it didn't make your experience so much better :) say thanks"

ponder
@ponder.ooo

traffic: owning everything and extracting rent is a wildly successful business model

ponder
@ponder.ooo

traffic: owning everything and extracting rent is a wildly successful business model

ponder
@ponder.ooo

fucking how is gboard interpreting my swipes correctly not a solved problem yet

ponder
@ponder.ooo

genre of post that's like "one thing i wanna bring up while everybody's talking about fred," and u have no idea who fred is and havent seen anybody talking about fred

ponder
@ponder.ooo

the human mind is a battleground of base drives; for sex, food, status &c &c. but the most primal and powerful drive is the drive to be miserable

ponder
@ponder.ooo

meal prep photo but instead of meals it's water. water prep. photo of a kitchen counter with a week's worth of mason jars filled with tap water on it

ponder
@ponder.ooo

pr review is when you paste the pr into chatgpt and then paste the model's paranoid hallucinations into the pr review comments

ponder
@ponder.ooo

"please confirm that [condition that if it didn't hold would literally make the project, which compiles, not fucking compile]"

ponder
@ponder.ooo

"please add a test to ensure that [condition confirmed by one of the tests already introduced in this pr]"

ponder
@ponder.ooo

our product poses an unprecedented danger to all life and the only way we can protect you from it is by making the most money anybody has ever made

ponder
@ponder.ooo

momentarily misread doritos "FLAMAS" as "HAMAS" and had to do a double take

ponder
@ponder.ooo

"to install, just curl some shit into sh" no fuck you publish an actual package

ponder
@ponder.ooo

the biggest problem with trying to make new systems programming languages is that instruction sets are a nightmare hellscape

ponder
@ponder.ooo

every presently existing programming language is a cautionary tale for language designers