రామయ్య • /räːmajːɐ/

@nBJGvolvAdjzSHt @miniapeur What's cringe about Eugene Khutoryansky except his choice of animation software? Also why is ElectroBOOM not in the PopSci half?

True?

@miniapeur Not true. I just like 3B1B for his intuitive animations. I don't consider myself good at math. Veritasium neatly links math to everyday stuff, which is what I like about it. Khan academy is for high school.

@hanssium @bknshittpsyid This is interesting. I felt this t vs d apicality distinction when I was listening to "Mbah Dukun" by Alam. Didn't listen to a lot of Javanese audio to notice the pattern yet.

@generalusername @bknshittpsyid The mapping of foreign (incl. Indonesian) alveolar stops in Javanese is a bit uncommon: t → t̪ d → ɖ also my Indo /t/ is laminar and almost dental, while my Indo /d/ is always apical; and I think this holds true for most of Java(nese speakers) as well, idrk the others

ꦱꦗꦤ꧀ꦤꦺꦥꦸꦤꦏꦮꦤ꧀ꦲꦶꦏꦸꦲꦶꦱꦶꦤꦺ꧇ ꦱꦼꦩꦂ ꦒꦫꦺꦁ ꦥꦺꦠꦿꦶꦏ꧀ ꦧꦒꦺꦴꦁ

So İstanbul is Mickleyard or *Mahāgr̥dʰaḥ

@Fettermans_Iump @Vokabre @azforeman Sweden is Ruotsi/Rootsi and Russia is Venäja/Venemaa.

@cobbaltt Based on the usage of tānu/tāmu in Telugu, I feel like it's a calque from IA usage. tānam itself is a word for soul, another hint at ātmán. Tāmu/tamaru is largely used poetically and colloquially it usually stops at 2° which is mīru, & like Kannada, tamaru has memetic use now.

@generalusername That's the origin of āp, of course, but āp in modern H-U doesn't have the same level of honoricity that tāngaḷ does.

Very similar in shape Telugu: tamaru “yourselves” < taʜn(?) “self” ⊗ p “ᴘʟ” ⊗ a “ᴀᴅᴊᴢ” ⊗ ar “ʜᴜᴍ.ᴘʟ” Gujarati: tamārũ “yours(ᴘʟ)” < tuṣmākaṁ “your” ⊗ kr̥tyakam(?) “made-thing” Unlike Telugu tāru which is also pl. (from tamaru ultimately), Gujarati tārũ is sg.

Stanford has introduced Jain-friendly dining options, which means students can no longer pretend to be Jain to get out of the meal plan.

@KShrikaanth Right, in Telugu, with all variants included, we have: 1°: nuvvu/nūvu, nīvu, īvu 2°: mīru, īru 3° (ātmán): tāmu, tamaru, tāru The third set isn't restricted to the second person in usage, but is used in the general sense of any self, including 3rd person (Singular tānu).

Three levels in South also Kannada- nInu, nivu, tAvu Tamil- nI, nIngaL, tAm/tAngaL

A Katar with ten avatars of Vishnu engraved on its handle 17th century

@tlonic @hanssium + discourse into picture. I don't see any conclusive evidence proving syntax and semantics can be fully separated in LLMs

@tlonic @hanssium One could argue that at embedding level, things should have been normalised to a point where the exact textual representation of something doesn't matter as much as the semantics of it, but I think at some level, text may be related to syntax, and brings syntax vs semantics +

it pains me to see that complex scripts will always be discriminated in tech against simple, typographically uncomplicated scripts Aksara Jawa will always be "waste of horizontal & vertical space", Mongolian Hudum will always be "too difficult to implement" etc

@tlonic @hanssium + for instance, क्रौ is a single syllabic unit in brahmic abugidas but invariably gets broken down to ⟨k‌[zwnj]rau⟩ or something similar in its internal representation, rather than being treated as a basic unit

@tlonic @hanssium I didn't use the right term, but what I meant is that you no longer have the advantage of a syllabic abugida, since things are broken down to resemble alphabetic sequences +

@SagirAhmed98 @Saatvata I was checking translations for *numbers* in Gujarati the other day, and it got those wrong too.

GT needs to work on colloquial Assamese. It uses literal translations having heavily Sanskritised Assamese with English-based syntax which nobody speaks, but used in a formal register. It should be বঢ়িয়া। আমি তাইক পিছতকৈ কফি দোকানলৈ নিব লাগিব।

@straceX How is the position of each camera known to the system? Or does it figure things out itself?

Someone just built a fully open source motion capture system that works with regular webcams. It’s called FreeMoCap, and it can track human movement in 3D without any markers or specialized suits. You just use multiple cameras, and it automatically generates skeletal data that’s good enough for research use. no expensive hardware & proprietary software, it’s completely open source. > Honestly, this removes one of the biggest barriers in mocap and makes it accessible to almost anyone.

10T params. Wild. Are we still not past the existing scaling regime?

@hanssium … and a few people believed that it lets them work with much larger reasoning trails in lesser characters. Of course this comes with drawbacks like learning the CJK-specific priors for relations that don't necessarily universally semantically exist, like linking broom and wife&c

@hanssium Agreed. This got me wondering what's happening with CJK characters now, and the effects it has on dense representations of things inside LLMs, I remember there was a phase some 2 years ago where some frontier models would switch to Chinese characters for "reasoning" tokens…