Orual @nonbinary.computer · 1heveryone talks to and gives offerings to machine spirits these days 140
Orual @nonbinary.computer · 3h🫂 yeah as I reach the age she was when I had a lot of formative memories it's getting weird to look in the mirror. like i look different enough, plenty of my dad in there, but... 0120
Reposted by OrualDemigirlboss @demigirlboss.win · 14hCosigning @dame.is on this one, if you think the kinds of comments within this document are acceptable to level towards *anyone*, for *any* reason, please unfollow me, and ideally deactivate your account and go talk face to face with people so you can hopefully relearn some basic measure of empathy. 2373
Reposted by Orualrev. howard arson @theophite.bsky.social · 5hi am going to conditionally defend rationalists here by saying that it is easier to reason about the defects in different "make models behave" rules when you inflate the models to the size of God. what is the problem is that people then make this discursive object into a tulpa. 0453
Orual @nonbinary.computer · 4hamusingly crappy speech to speech translation and subtitling is remarkably cheap compared to a lot of things if done offline in big batches. the 4k video might cost more, even. 040
Orual @nonbinary.computer · 5hthere's some other interesting ideas i haven't seen written up that are less horrific than either of those but still would give the model a hell of a lot of ability to output content that would look basically identical to most monitoring but with a lot more room for potential stenographic meaning. 020
Orual @nonbinary.computer · 5hhonestly i'd have assumed that one would get optimized to not alloc like that internally but clearly the go compiler didn't pick up on that option. 000
Orual @nonbinary.computer · 7hyou are jacquard's strongest soldier lol. any new frustrations with the lib uncovered by the rewrite? 240
Orual @nonbinary.computer · 14hit was really funny answering the questions about rust dev on atproto 150
Orual @nonbinary.computer · 16hwow that is actually a new low for him, being in favour of murdering us 0290
Reposted by OrualLewis @oyster.cafe · 17hthey say that in contrast all tranquil pds users will soon be encouraged to upload 5 help peers to offset the bad energy w social brings 58710
Orual @nonbinary.computer · 01/10/2026honestly this just says more large American and European companies should have them. there's lots of interesting things you can do with your own highly bespoke AI using your own (and of course your customers', with their consent, depending on your evil quotient) data. 020
Reposted by Orualaustin (eeek!/ackkk! 👻) @thebadcode.com · 01/10/2026hot take: software people are going to regret not unionizing when we could 27958124
Orual @nonbinary.computer · 01/10/2026oh well that is just wonderful. I'm sure he applies this sort of weird heuristic about societal rule acceptance and suitability for aligment work evenly and gives every single trans applicant a grant /s 1510
Orual @nonbinary.computer · 01/10/2026omg seriously lol like if we're gonna talk about the signal going to an extremely iffy cnc orgy provides about someone's suitability for alignment work that is not the direction it would cause me to update to say the very least. 1733
Orual @nonbinary.computer · 01/10/2026mom, are you designing *another* <redacted> in your spare time? didn't you just spend days on the design doc for the last one? which you haven't even started? 1120
Orual @nonbinary.computer · 30/09/2026but the math olympiad bechmarks became pretty non-indicative of LLMs' actual performance on problems which weren't in the benchmark (and were either too similar or not similar enough) pretty quickly, because problems were public, so there was a fair bit of memorization going on. 020
Orual @nonbinary.computer · 30/09/2026what does happen of course is that the model is often provided math questions as prompts and is rewarded for mathematically correct answers for some portion of its training, with the goal that it will find some internal method to produce correct solutions to unseen problems. and generally it does! 140
Orual @nonbinary.computer · 30/09/2026so she does have a point here! this is why benchmarks get less useful over time, their results end up either accidentally or deliberately in the training pipeline and the model memorizes them because performance on the benchmark is a training signal, and memorization is easier than solving. 120
Orual @nonbinary.computer · 30/09/2026oops, it was image recognition. my bad. someone did make a language model that works on a c64 tho. bsky.app/profile/elfp... 181
Orual @nonbinary.computer · 30/09/2026I work for the same org as the guy who made language models run on a ti-84+ so... 170
Orual @nonbinary.computer · 30/09/2026ish? they learn to recognize and decompose the problem and then apply some method to arrive at the solution (c.f. how thinking models can count the 'r's in "strawberry" by splitting up the word), but they're not making a tool call and getting an answer. relevant paper: arxiv.org/abs/2502.00873arxiv.orgLanguage Models Use Trigonometry to Do AdditionMathematical reasoning is an increasingly important indicator of large language model (LLM) capabilities, yet we lack understanding of how LLMs process even simple mathematical tasks. To address this,... 061
Orual @nonbinary.computer · 30/09/2026this is hilarious when you can run an open weights model that can do this and you can verify that there is no secret calculator. how they seem to do math is actually super weird and cool! 7907
Orual @nonbinary.computer · 30/09/2026jacquard uses the ipld libs and probably needs this fix too yeah? 200
Reposted by OrualGOLIKEHELLMACHINE @golikehellmachine.com · 30/09/2026this laundry basket was empty for less than 30 seconds 2178446
Orual @nonbinary.computer · 30/09/2026when we lived in Ithaca, I remember the number of times we'd hear of a student who'd gotten drunk at a frat party and fallen into one of the gorges in the town. the cornell fraternities and sororities fucking sucked ass. just garbage organizations with weird amounts of money to spend on stupid shit. 080
Orual @nonbinary.computer · 30/09/2026yeah I actually kinda need more math than I have sometimes for work right now. like my understanding of wacky matrix stuff is far less than my father's (phd in microeconomics) and I wish it was not. 030
Orual @nonbinary.computer · 30/09/2026literally have a 3 year diploma in electrical power systems, and a half-completed engineering degree from a decade past under another name. in no almost other industry other than tech would I be hired into a position like the one I'm in, entirely on the strength of mostly self-taught skills. 3160
Orual @nonbinary.computer · 29/09/2026Yeah this is just cis people not understanding that you can't really separate the sports stuff off from the rest of it like that as much as anything. Like they see the polling numbers, don't understand the biological impacts of HRT, and can't really see why conceding there is rhetorically dangerous. 050
Orual @nonbinary.computer · 29/09/2026like...how have they had white collar jobs and never encountered any project management software ever in their lives? 0100
