Conversation

Notices

1. Embed this notice
   SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 02:49:57 JST SuperDicq

   Looks like Anthropic is currently doing a giant rugpull and they seem to be degrading their models over time while at the same time redacting the "thinking" responses to try and hide it.
   
   This is what you get when you are dependent on proprietary saass products.
   
   There's nothing you can do except complain and beg thatthey will revert the changes. Absolutely pathetic.
   
   https://github.com/anthropics/claude-code/issues/42796

   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 03:48:10 JST SuperDicq

     in reply to

     • theory

     @theorytoe@shin.mugicha.club I believe this specific model is impossible to run on average hardware and without proprietary software such as CUDA, but at least doing it this way you are not dependent on SaaSS.

   • Embed this notice
     theory (theorytoe@shin.mugicha.club)'s status on Tuesday, 07-Apr-2026 03:48:11 JST theory

     in reply to
     glm5.0 my love
   • Embed this notice
     theory (theorytoe@shin.mugicha.club)'s status on Tuesday, 07-Apr-2026 05:46:09 JST theory

     in reply to
     getting a decent coding model running on consumer grade hardware is kind of a pipe dream. Being able to buy compute from anybody I want is enough to make me happy, I'm not so much of a fosstard that I'm worrying about proprietary dependencies or whatever.
   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 05:46:09 JST SuperDicq

     in reply to

     • theory

     @theorytoe@shin.mugicha.club I personally don't think it is a pipe dream. Over long enough time hardware will get better and models will also get increasingly more efficient.
     Being able to buy compute from anybody I want is enough to make me happyYes, but what Anthropic is doing is not renting compute power. They decide how their models and how it behaves, you have no control over it. It's not like renting a VPS where you actually still decide what runs on that server, which would be much better.
     I'm not so much of a fosstard that I'm worrying about proprietary dependencies or whatever.Well I am a freetard, not a fosstard. And I think these thinks are important.

   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 05:47:20 JST SuperDicq

     in reply to

     • protoss

     @nigger@detroitriotcity.com A code leak does not make it "open source" or free software. You're still not allowed to modify or share it legally.
     
     Also the Claude code leak was just frontend client code. None of the important bits got leaked.

   • Embed this notice
     protoss (nigger@detroitriotcity.com)'s status on Tuesday, 07-Apr-2026 05:47:21 JST protoss

     in reply to
     @SuperDicq Claude code is even open source now smh
   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 06:10:26 JST SuperDicq

     in reply to

     • protoss

     @nigger@detroitriotcity.com claude code is 100% LLM produced and therefore public domainThere is unfortunately no legal precedence for who owns the copyright on LLM code yet. But this could be a possibility though...

   • Embed this notice
     protoss (nigger@detroitriotcity.com)'s status on Tuesday, 07-Apr-2026 06:10:27 JST protoss

     in reply to
     @SuperDicq claude code is 100% LLM produced and therefore public domain, which is open source
     and you can still modify and share code leaks if you do not respect the law
   • Embed this notice
     lolitechengineer@loli.church's status on Tuesday, 07-Apr-2026 07:52:38 JST lolitechengineer

     in reply to

     • theory
     @SuperDicq@minidisc.tokyo @theorytoe@shin.mugicha.club I believe this specific model is impossible to run on average hardware and without proprietary software such as CUDANo, that's not the case. Also with groups like unsloth that repack them into gguf you can easily run them on AMD/Intel/Vulkan supporting GPUs or CPU. Like from https://huggingface.co/unsloth/GLM-5-GGUF
     
     You just need 200GB of ideally VRAM. You can use system memory but this was already a really slow model on consumer hardware to begin with.
     
     I suggest 4.7 flash which can be run on as little as 10GB of RAM and will run reasonably well even on CPU https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF
   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 07:52:38 JST SuperDicq

     in reply to

     • lolitechengineer
     • theory

     @lolitechengineer@loli.church @theorytoe@shin.mugicha.club Yeah I have high hopes for Vulkan runtimes for these things. CPU will never be fast enough, but Vulkan is a way to avoid proprietary CUDA or ROCm.

   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 07:58:00 JST SuperDicq

     in reply to

     • lolitechengineer
     • theory

     @lolitechengineer@loli.church @theorytoe@shin.mugicha.club Also with these "flash" models are extremely quaternized and with a very low amount of RAM (i.e. 10GB) you can not do agentic coding. As you will probably be limited to a very small context size.
     
     Still works perfectly fine for a regular chat model though.

   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 08:08:47 JST SuperDicq

     in reply to

     • lolitechengineer
     • theory

     @lolitechengineer@loli.church @theorytoe@shin.mugicha.club I do wanna say I'm not an expert on AI crap at all outside of the fucking around I did with ollama on my RX 6700 XT in my gaming PC and some RAG stuff I implemented at my dayjob.

   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 08:08:48 JST SuperDicq

     in reply to

     • lolitechengineer
     • theory

     @lolitechengineer@loli.church @theorytoe@shin.mugicha.club I know Claude Code has a 200.000 token context size or something. More contex means more ram. With 10 GB of VRAM you can do like max 20.000 context size I think?

   • Embed this notice
     lolitechengineer@loli.church's status on Tuesday, 07-Apr-2026 08:08:49 JST lolitechengineer

     in reply to

     • theory
     @SuperDicq@minidisc.tokyo @theorytoe@shin.mugicha.club yeah, I wouldn't really know.
     The most "agentic" thing I have mine do is when I'm researching something I'll have it do a web search, download a few pages it thinks satisfy my requirements, and summarizes them. Helps filter through the shit.
   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 08:09:53 JST SuperDicq

     in reply to

     • lolitechengineer
     • theory

     @lolitechengineer@loli.church @theorytoe@shin.mugicha.club That really depends on the model. If you're using one of those small 1B flash models or something you can do more than that.

   • Embed this notice
     lolitechengineer@loli.church's status on Tuesday, 07-Apr-2026 08:09:54 JST lolitechengineer

     in reply to

     • theory
     @SuperDicq@minidisc.tokyo @theorytoe@shin.mugicha.club probably more like 8192
   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 08:27:01 JST SuperDicq

     in reply to

     • Johnny Peligro
     • theory

     @mischievoustomato@tsundere.love @theorytoe@shin.mugicha.club I believe you can work around consuming documentation in the context by implementing a RAG system.

   • Embed this notice
     Johnny Peligro (mischievoustomato@tsundere.love)'s status on Tuesday, 07-Apr-2026 08:27:02 JST Johnny Peligro

     in reply to

     • theory
     @theorytoe @SuperDicq I'd like to have a good enough model that can consume documentation of programming langs and knows of SO/blog output to help me code the stuff I do, but while explaining me the things, not really doing my work for me.
   • Embed this notice
     theory (theorytoe@shin.mugicha.club)'s status on Tuesday, 07-Apr-2026 08:40:07 JST theory

     in reply to

     • lolitechengineer
     it's not just context, it's how smart the model is. Try the best Qwen 3.5 you can run locally, then the best one cloud providers offer. it's night and day
   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 08:40:07 JST SuperDicq

     in reply to

     • theory

     @theorytoe@shin.mugicha.club You mean the companies who have scraped the entire internet that are currently leading losses on 10 gorillion GB of VRAM perform better than my gamer GPU? No way!

   • Embed this notice
     SuperDicq (superdicq@minidisc.tokyo)'s status on Tuesday, 07-Apr-2026 08:45:19 JST SuperDicq

     in reply to

     • lolitechengineer
     • theory

     @lolitechengineer@loli.church @theorytoe@shin.mugicha.club Because that's literally how you calculate the VRAM requirement.
     
     Amount of parameters * amount of tokens in the contex = total VRAM usage