Claude AI discussion

[snowdunes]

Dario needs a lolcow thread, getting really tired of his big ugly heeb mouth running non-stop, his kvetching about open source models, seething at china together with his peak kike business practices that would still shock the biggest anti-semites who thought they knew everything. Every single time you think it couldn't get any worse, don't worry, Dario is there to make it happen.

I hope Zhipu labs rapes Anthropic (because other American models just suck and Mistral is too busy dodging EU laws to get any training data at all), I hate seeing this extortion AI ecosystem that will only lead us to another AI winter.

 

[Hannibalistique]

Dario needs a lolcow thread, getting really tired of his big ugly heeb mouth running non-stop, his kvetching about open source models, seething at china together with his peak kike business practices that would still shock the biggest anti-semites who thought they knew everything. Every single time you think it couldn't get any worse, don't worry, Dario is there to make it happen.

I hope Zhipu labs rapes Anthropic (because other American models just suck and Mistral is too busy dodging EU laws to get any training data at all), I hate seeing this extortion AI ecosystem that will only lead us to another AI winter.

The most annoying thing is Dario raping/redefining the term "distillation", when Anthropic gained their own edge "distilling" claude from chatgpt in the exact same way.

 

[snowdunes]

The most annoying thing is Dario raping/redefining the term "distillation", when Anthropic gained their own edge "distilling" claude from chatgpt in the exact same way.

Incoherent AI sperging incoming.

It goes deeper than just "China stole from us, whaaaaa!". He use the term to accuse Chinese models of "distilling" when in reality it's mostly just API washing for them to get access to Anthropics models (like openrouter). Anthropic knew this but chooses to ignore this because it was good money for them, they only started to cry about it when Chinese models caught up to them and suddenly the Chinese users dropped and they stopped paying (some distilling happened but not in the scale Dario claims and also... who cares, as you said, Anthropic did it too). GLM 5.2 is BTFO Sonnet in coding, close to Opus (sometimes even beating it, for website making, GLM is just as good in my opinion). Only edge they have is their Fable/Mythos model that is just a computational brute-force model (+2TB parameters probably).

To put more fuel into the Anthropic competition list: Sakana AI (yes, the Japs that are 20 years behind in computers) created a swarm/agent arcitecture based on existing local models showing some cool ouputs, just to have something functional until they get their shit together to train their own models later. I will shit my pants if Mistral actually catches up too, very unliekely but not imposible either considering how wild the market is.

Can't wait for 2TB unified memory to come to consumer market, we already have 0.5TB with the Mac Pro, the tech is here to push it a little forward in order to have the current frontier models locally (no, I'm not delusional, it can be done).

All this can and will happen, and Dario is trying everything he can to prevent this with sheer retarded copes and lying. I'm impressed by one thing, and that is how much of a piece of shit you can be to make Sam Altman look like the sane one.

 

[Hannibalistique]

It goes deeper than just "China stole from us, whaaaaa!". He use the term to accuse Chinese models of "distilling" when in reality it's mostly just API washing for them to get access to Anthropics models (like openrouter). Anthropic knew this but chooses to ignore this because it was good money for them, they only started to cry about it when Chinese models caught up to them

To be fair to Anthropic they "tried" to stop it with ID verification etc., but as long as their paying western subscribers subsidize infinite niggers using it in Africa the Chinese can just pay random Africans next to nothing to verify proxy accounts with their IDs.
So it's kind of hollow.

 

[HootersMcBoobies]

his kvetching about open source models

GLM 5.2 is BTFO Sonnet in coding, close to Opus (sometimes even beating it, for website making, GLM is just as good in my opinion). Only edge they have is their Fable/Mythos model that is just a computational brute-force model (+2TB parameters probably).

I just started using GLM-5.2 via openrouter. I gave a spec to Fable and ran it on Claudecode, saved it to a branch. Switched back to the previous state and ran the same spec via GLM-5 using Opencode.

The UI that Fable produced was way better. Looked beautiful. The actual code implementation was garbage compared to GLM. GLM separated out all the classes, made a better hierarchy, better database schema; just overall cleaner. I then copied just the UI pieces to a temp directory and told Opencode/GLM to integrate that html/css/assets and now I have something fairly solid.

The nice thing about GLM is that you can run it locally. You need somewhere between 256GB and 800GB of VRAM depending on the quant size; so a minimum of $8k~$10k in compute depending on if you buy a set of AMD R9700s, RTX 6000s or connect up 3 x Strix devices. It's completely impractical today. No group of friends are going to go in togeher for $2k/each to buy gear that they can only load one model on at a time, and depent on one person to maintain the uptime. So it's not practical, but using an open weight model is hedging against the possibility of being able to purchase local hardware in the future.

Also, Opus/Fable is probably only better at UI because they used all the training data from Figma to create Claude UI, tanking Figma stocks. Truly evil levels of Microsoft shit.

Sakana AI (yes, the Japs that are 20 years behind in computers)

They created Ruby. .. Sony still makes good stuff. Nobody remembers Panasonic because they only make pro-grade stuff an no consumer products. ... Yea it's sad to see how far Japan has fallen behind in software engineering.

To be fair to Anthropic they "tried" to stop it with ID verification etc.,

I think July 7th will be a footgun moment. Sure, over half of the subscribers will just scan their ID and face, but I think a non-trivial amount (hopefully 15% ~ 20%) will just cancel. I will be one of them. Between both home and work on a Pro plan I get reimbured for, I use ~2m tokens a month or ~200k/week. There are a ton of routers (openrouter, portkey, etc.) and a lot of them allow basic shit like using a password instead of a fucking e-mail magic-link, purchase without saving your credit card, details pricing reports, and other features that are not available from Anthropic's 100% dark pattern of a WebUI. I know personally, I don't get anywhere near maxing out my Pro plan, and it just seems cheaper and way less of a hastle to use an API gateway even with their overhead costs.