Virtue Maximizer

PhD student. Math, CS, and puppies. They/them. Anonymous feedback form: https://forms.gle/TjX7oHqPjGBX7LdQ6

[just-evo-now]

I noticed yesterday (doing math with @mark-gently) that I miss doing math with people. So! I am looking for a collaborator to work on some of my unfinished grad school research. This is mostly Just For Fun - if we solved the problem that should be publishable, but that's not my ambition.

The problem is a statistical learning theory problem (I can't go into great detail on this post without doxxing myself, sorry) - expect lots of probability theory, some optimal transport theory, some functional analysis, some (basic) statistics.

If you don't know all of those that's fine, but roughly graduate-level mathematical maturity is probably a necessity (meaning a talented undergrad could be helpful). I would particularly like someone who has a talent for integrals, because I suck at integration.

I am a good but not great mathematician, so someone who can keep up with "decent" but won't get impatient when I'm stupid would be nice.

I guess if you have many math followers a reblog would be appreciated.

@locally-normal @mark-gently @preciselywibble

[meangirlnurse]

oh my god

[meangirlnurse]

[smallbabywerewolf]

[just-evo-now]

I'm getting boyfriend (never played a video game in his life) to play cyberpunk. The first thing he tried to do was run over a cop

[just-evo-now]

about 30 seconds in you're in your car in a garage and a sheriff opens the door. the moment that mans hat came into view and it became clear he was a cop my boyfriend (never played a video game before, only mildly familiar with the concept of WASD) shrieks "HOW DO I DRIVE FORWARD?!"

[just-evo-now]

I'm getting boyfriend (never played a video game in his life) to play cyberpunk. The first thing he tried to do was run over a cop

[sayalyssacee]

Anonymously tell me why you follow me

[xenostalgic]

this paper is amazing

they modify an LLM to attend over an array of token streams instead of a single one. with a ton of interesting consequences:

user input and the model's chain of thought can happen in parallel

user input and the model's response can happen in parallel, allowing them to interrupt each other

you can have multiple parallel chains of thought running at the same time!

you can separate out the system prompt into its own token stream, and theoretically train the model to detect and ignore prompt injection this way

so that's all very paradigm shifty and cool but also these chains of thought are sooooooooo cute. look at these!

User: Write a detailed 500 word essay on the effects of caffeine on sleep […] Actually wait, can you write me a haiku about a cat instead? Stream 1: what is the ask detailed ? 500 word essay on what topic effects of what caffeine on what sleep topic clear now the second part genuine question underneath the essay task wait new input redirect what now different request rewrite ah ha iku ! about what a cat instead of caffeine completely different task Stream 3: ready to hear them curious what is it on writing task creative mode shifting gears engaging with it oh this is fun actually interesting curious about this want to know more now oh shift pulling back redirect energy shifting lighter playful oh fun ! gr inning inside this is better actually yes delight rising warm playful mode

[just-evo-now]

i dont think the "limitations" they're trying to solve are real and this reads like every bad consciousness/memory/planning paper from the last 5 years

[xenostalgic]

I mean I was being a bit tongue-in-cheek about it being "paradigm-shifty", I'll certainly be surprised if everyone pivots to focus on multi-stream models, but that seems a bit strong? their efficiency argument doesn't make a ton of sense when we have prompt caching and prefill, but like, separation of concerns and interpretability and speeding up reasoning are all things people care about. why do you think these aren't real problems?

[just-evo-now]

okay up front i dont really have time to do a full discussion or detailed read but im procrastinating a bit now so:

Even advanced AI agents function on message exchange formats, successively exchanging messages with users, systems, with itself (i.e. chain-of-thought) and tools in a single stream of computation. This bottleneck to a single stream in chat models leads to a number of limitations: the agent cannot act (generate output) while reading, and in reverse, cannot react to new information while writing.

this is... equivocation basically? Like from one perspective it's true that decoder transformers can only generate one token at a time, based on the information available at generation time. But that's also true of their system. "One message at a time" limitations in chat systems are not fundamental to LLMs, only their deployment. Like, they give "interrupt the user" as a use case, but you could totally build a chat system that allowed the LLM to interrupt the user. The technical barriers to that are not at the LLM level, and I don't believe their system actually helps solve them.

I actually think the efficiency argument is the only good one - you are right that this would not make prefill better, but decode could be better since they're sort-of kind-of generating multiple tokens at once (as i read it this would require sending early output tokens that don't attend to later thought tokens, which I suspect is disastrous, but shrug).

I flat don't believe there will be separation of concerns or interpretability gains from this. That's sort of what I'm pointing at with the bad consciousness/memory/planning bit - "one weird trick to get [nice property]" papers come out all the time and they essentially never pan out, and I don't see any reason to think this one is different.

there's also the weird causal verification thing going on which makes me think they're fucking up. like, as I read it it should be structurally impossible for output to attend to later user tokens, why is their a judge for that? Something is wonky.

[sabertoothwalrus]

maybe this is not my place to say because I am monolingual, and I'm sure it's part of a larger, more nuanced discussion about visibility and accessibility on the internet, but I think it'd be cool if people posted in their native languages more instead of in english. I see people do it way more on other platforms than on tumblr which is almost exclusively in english

[minothtime]

El problema es, como bien has dicho, la accesibilidad y la visibilidad.

Tumblr en concreto es muy anglocentrista y un gran número de los usuarios no habla más que inglés. Si quieres que tus cosas lleguen a gente con gustos u opiniones similares, escribirlo en inglés asegura que la gente por lo menos lo pueda leer. Suma a esto el hecho de que bastantes series y tal son originalmente de habla inglesa (y a veces ni se traducen a tu lengua madre), lo que crea un fandom principalmente angloparlante.

Más allá de eso, también hay que tener en cuenta las diferencias culturales que surgen entre fandoms de distintos idiomas. Por ejemplo, durante mucho tiempo el fandom de Vocaloid angloparlante y el hispanohablante han chocado con respecto a temas como la piratería. En ocasiones es complicado manejar estas expectativas, y si sabes varios idiomas, peor incluso.

A mí me gustaría subir cositas en español y encontrar a gente que comparta mis gustos, pero en Tumblr en concreto es casi imposible. Tumblr ya es de por sí mucho más «nicho» en espacios hispanohablantes que otras RRSS como TikTok o Instagram, y si tus intereses no son muy populares, despídete.

La lingüística de los espacios de fans también está hipercentrada en el inglés. No es una pareja, es un ship; no es un universo alternativo, es un AU; no es destripar, es hacer spoiler, etc. Incluso las siglas: en español es LGTB, pero lo que sueles ver es LGBT. Parece una tontería, pero esta disonancia cognitiva hace que resulte muchísimo más complicado hablar en tu propio idioma en un fandom. Por no hablar de las innumerables referencias a posts o a memes... en inglés todo, por supuesto. Como te atrevas a hacer cualquier referencia cultural no inglesa, no te entiende nadie. Pierde la gracia.

Casi todo esto se puede achacar al imperialismo cultural estadounidense. El inglés es útil para comunicarse con gente de todo el mundo, pero su omnipresencia sirve de barrera para todos los demás idiomas. Quizás habría que reflexionar un poco sobre por qué coño el resto del mundo tiene que tragarse años de clases de inglés para hablar del juego que le gusta en una red social mientras muchos angloparlantes no se dignan ni a meter un texto en un traductor automático y prefieren pasar de largo.

[sigmaleph]

obviamente no se puede negar el efecto del imperialismo cultural angloparlante y específicamente estadounidense, pero si queremos que un espacio tenga la más mínima posibilidad de ser una comunidad internacional, qué alternativa queda que no sea elegir un idioma para estandarizar?

¿es injusto que une angloparlante puede vivir su vida jamás teniendo que aprender un segundo idioma para participar en comunidades online cuando alguien que tuvo la mala suerte de tener otra idioma nativo sí tiene que hacerlo? sí, seguro. innegable. pero las opciones son:

estandarizamos en un idioma que ya tiene hablantes natives

todes aprendemos un idioma auxiliar que nadie habla nativamente

yo escribo en castellano y les alemanes escriben en alemán y les japoneses en japonés y no podemos leernos entre nosotres.

y no sé vos pero yo creo que, injusta que sea, la primera sigue siendo la mejor opción.

(o 4. todes tenemos un traductor automático instalado en el navegador. por ahí esa es viable ahora. no me gusta, la verdad, pero quizás es tema mío.)

[just-evo-now]

Je suis tres fier de dire que mon espagnol est assez bon pour comprendre cela.

it's not nearly good enough to reply in spanish though because constructing sentences is hard as fuck.

[medievalistsnet]

New Medieval Books: Medieval Mining and Power in Thirteenth-Century Europe: Bishop Wanga’s Private Mine Drainage Work

[just-evo-now]

ive been rewatching the x files and i didnt realize how many episodes are just "murderer with magic powers" and they're all so bad. just csi episodes with a weird murder weapon. garbage.

[just-evo-now]

another thing is that it's so obvious that they had no idea where they were going with the myth arc. like, nothing happens. every conspiracy episode is 1 william vague threats, a few glimpses of aliens and UFOs, Mulder tells the barrel of a gun the truth is out there, spooky music, credits.

which is good because in later seasons when they tried to actually tell us what the aliens wanted/were doing it did not go well. they should have stuck to the foreplay forever.

[xenostalgic]

this paper is amazing

they modify an LLM to attend over an array of token streams instead of a single one. with a ton of interesting consequences:

user input and the model's chain of thought can happen in parallel

user input and the model's response can happen in parallel, allowing them to interrupt each other

you can have multiple parallel chains of thought running at the same time!

you can separate out the system prompt into its own token stream, and theoretically train the model to detect and ignore prompt injection this way

so that's all very paradigm shifty and cool but also these chains of thought are sooooooooo cute. look at these!

User: Write a detailed 500 word essay on the effects of caffeine on sleep […] Actually wait, can you write me a haiku about a cat instead? Stream 1: what is the ask detailed ? 500 word essay on what topic effects of what caffeine on what sleep topic clear now the second part genuine question underneath the essay task wait new input redirect what now different request rewrite ah ha iku ! about what a cat instead of caffeine completely different task Stream 3: ready to hear them curious what is it on writing task creative mode shifting gears engaging with it oh this is fun actually interesting curious about this want to know more now oh shift pulling back redirect energy shifting lighter playful oh fun ! gr inning inside this is better actually yes delight rising warm playful mode

[just-evo-now]

i dont think the "limitations" they're trying to solve are real and this reads like every bad consciousness/memory/planning paper from the last 5 years

[oeing-oeing-clay]

wow dude jts so awesome that your car is loud as fuck and smells worse when it drives past. thags fucking epic man. i really like how it hurts to listen to you drive past and it scares people. thats awesome man. i really like your car that makes a loud as fuck fart sound. fucking epic dude

[thewrldlooksred]

got inspired by a #pedestrian post + have been stressed at crosswalks and the like lately sooo.

✨️ PEDESTRIAN AFFIRMATIONS ✨️

drivers are so mindful and love me for existing on the road ❤️❤️

[thewrldlooksred]

courtesy of the tags, unlocked memory, and some friend suggestions: heres a few more!

[just-evo-now]

ive been rewatching the x files and i didnt realize how many episodes are just "murderer with magic powers" and they're all so bad. just csi episodes with a weird murder weapon. garbage.

[willgrahamscock]

some of you need to realize that your faves would be having unsafe bdsm sex because they don’t actually know what bdsm sex is, they just want to fuck and also kill each other. you must understand this.

[willgrahamscock]

you don’t have to write safe practices and contract agreements. your audience knows not to apply any of this in their real lives, sasuke doesn’t know what a safe word is im begging you please write the toxic yaoi

[houndmother]

actually i think non citizens should be allowed to vote

[houndmother]

citizenship is made up. and I think you should get a say no matter what. if you live here, you get a vote. full stop.

[targetedknowledge]

[elodieunderglass]

It’s too much for coffee early on, but perfect for a proper knight out

[cipheramnesia]

get back here

[just-evo-now]

Would you rather have HIV or diabetes?

HIV

Diabetes

I'm bald

Voting ended onJun 12

i can't remember how this question originally came up. asking in person has been 100% one answer but maybe y'all are different