Egregoros

@WandererUber the haters are %100 wrong on this one. deepseek/glm/kimi inference is already profitable at low prices

Wanderer atop the sea of clouds or whatever

as they often are.

>profitable
for whom?

lain

@WandererUber for people who have inference hardware

Wanderer atop the sea of clouds or whatever

that's not the issue. If the model is not profitable for the lab making it, they must raise prices (and stop/delay weight releases)

lain

@WandererUber these are all open models, anyone can run them. if a third party can run them at profit right now, they'll be able to do that in the future. glm and deepseek and kimi are worse than claude and codex, but not so much worse that people wouldn't swithc to them if claude would cost $1000 a month

Wanderer atop the sea of clouds or whatever

>if a third party can run them at profit right now, they'll be able to do that in the future.
I don't like that I have to spell everything out WHILE playing devil's advocate. This does not follow at all. They have training costs to recoup and if they don't keep pace they will fall behind. If they fall behind, Anthrobbicc would be able raise prices to e.g. $40 for pro, maybe $100

barrier to entry for writing code and not publishing it under your own name was virtually zero (compared to having a different identity anyway of course) and now is not.

lain

@WandererUber but the open models are trained already, there's no more cost. i'm not talking about anthropic and openai and who ever trains. the open models are profitable right now. if the trainers go bankrupt, well, that's sad, but inference still stays profitable.

Wanderer atop the sea of clouds or whatever

We had this discussion before and you are making an argument that I
-already agree with
-is only tangentially related to what I am saying here

>People can just stay on GLM5.1
they can also hand-code. If SOTA devving costs money and you have to pay again if you want it under a second identity, that's a detriment.
If GLM 5.1 is not profitable for the model trainer, there will not be GLM N.1 in future. Open Weights models *could* lag behind which would make SOTA be OAI / DarioCorp -dependent. I believe quite a few people would be deterred and/or chance it under their real account and catch a C&D from Nintendo for it

lain

@WandererUber yeah, that's true. if current level training turns out to be not profitable, we won't get better models soon. but the current ones still are profitable. so i'm not sure what we are disagreeing on

newper

@lain @WandererUber Diminishing returns that way would honestly be for the best, a lot of hate is due to aggressive hype fear fomo mongering and having it turn into a normal industry would make folk relax. and treat it as a normal technology

Wanderer atop the sea of clouds or whatever

I genuinely think it has too high an "intelligence-to-get-it floor", for lack of a better term.
Sometimes I'm not even sure if the LABS get it (Anthropic publishing a post about a "mind reader" for Claude). A lot of previously smart-enough people are falling for the convincingness of the output and are incapable of checking for *correctness*

all that to say, it could be a long while until it's "treated as a normal technology"

newper

@WandererUber @lain

probably coming off as a bit of the anti llm for the sake of anti llm side but-

idk how that happens tbh ik web ui models are kinda shit especially without harnesses (agentic stuff inmean) and really haven't ever vibe coded.

but I've used it as a reseaech assistant (with all the deep research bells and whistles google afforded me) and while certainly helpful (still doesn't replace normal searching for me, probably as I prefer searching first) l. Its obviously not a smart thing and makes stupid as hell errors or honestly fails to get the point (which is kinda obvious since they are fundamentally word generators)

One example is how It starts phrasing everything with basis on a fact which should have been treated irrelevant.

I tried to use it to write a slop paper and report (uni obligation) it was so bad that I basically didnit manually. Ig it is me actually having standards but it felr quite bad and focised on entirelynthe wrongnthings and had such terrible flow that babying itnto get something worthwhile was harder than actually writing it myself

code (surprisingly due to the NLP roots of the architecture) is probably a best case scenario in hindsight which is why they actually end up being surprisingly useful for thay

lain

@meeper @WandererUber i know you're engaging with it, but somehow you're not getting the results. i don't know why, but it can do a lot more than you found out.

newper

@lain @WandererUber probably the parricukar tooic for the slop research paper.

for example it wqs surprisingly good at hindi poetry (which is actually quite complex to get a good meter in but soinds really good) and analyzed stuff based on the rules I didnt even know at the time.

and recent models (late 2025) suddently became useful for my etymology interest which would otherwise require me to dive into obscure books (which I do want to I digress)

Wanderer atop the sea of clouds or whatever

>based on the rules I didnt even know at the time.
it always seems better if you have no clue
Gell-Mann amnesia but for AI

Replying to @WandererUber@poa.st

newper