Joe Gibbons

[Joe Gibbons @JZGibbons]

@todaymare decent rack

[buun @spiritbuun]

Introducing a new KV format: VBR (Variable Bit Rate)
VBR dynamically quantizes your KV cache layer-by-layer as your session grows, giving you the highest quality possible within your VRAM constraints. This is my dream format.
Pushed to master. Available now. 1/15 🧵

[Darth Powell @VladTheInflator]

Hey @grok are you fucking telling me the guy who discovered the double helix in DNA and won the Nobel Prize found that IQ was determined by genetics based on scientific data and they stripped him of his fucking career to maintain a false narrative of all humans being equal.

[Joe Gibbons @JZGibbons]

@NousResearch Holy carrot!

[Joe Gibbons @JZGibbons]

@elonmusk If you keep the price reasonable, then we'll all be on board here. God bless you on many fronts.

[Joe Gibbons @JZGibbons]

@LechMazur fair point, I'm concentrating on keeping my sw dev job in use for a few more years - not worried about generating SEO slop/blog posts etc.

[Joe Gibbons @JZGibbons]

@ivanfioravanti So true but I'm still yearning for Qwen3.7-27B or 25B-A3B more than than the frontier models

[Joe Gibbons @JZGibbons]

@pupposandro @Takyon @AMD @luceboxai Great, I hope you can do it and keep local inference costs down

[Joe Gibbons @JZGibbons]

@antirez @theo This is fantastic, but things may not stay this way forever - we need a proper way to "crowdtrain" open source weight models, must be millions of consumer GPUs that can chip in.

[Joe Gibbons @JZGibbons]

@sudoingX shitloads of local models using a DSPy harness, too many to list but limited to RTX 3090 24GB etc. and GLM-5.2 to write the DSPy harness, refine it, every bloody day! I'll get there, just need more VRAM

[Joe Gibbons @JZGibbons]

@chefbama @sudoingX very interested in this

[Joe Gibbons @JZGibbons]

@HMICFRS @StaffsPolice lol - this proves "diversity" over ability is so wrong once again.

[Joe Gibbons @JZGibbons]

@aaronburnett or just straight to machine code

[Joe Gibbons @JZGibbons]

@aaronburnett cool, try assembly language for various platforms to eek a few more % over C? - why not, it's just guided prompts to an LLM, right?

[Joe Gibbons @JZGibbons]

@RollingHedge @DanNeidle @garyseconomics I'd be wary of Neidle, he's a major-leftist himself

[Adam Wren @aswren]

Oh man what a satisfying video, brutal stuff

[Joe Gibbons @JZGibbons]

@malikwas1f @Alibaba_Qwen @migtissera https://t.co/5r47UdiAHW

[Migel Tissera @migtissera]

Not sure of the Mac Studio, but since it's a dense model -- Mac would be slower. I'm doing 45 tokens/second on my 4090 though, the 4-bit quantized GGUF. I also just realized that I built the GGUFs without MTP, so it could get faster. I'll upload the MTP enabled ones shortly, keep an eye out here: https://t.co/q4RmgUkJgr

[J.K. Rowling @jk_rowling]

Whistleblowers in the NHS and at gender clinics have been trying to raise the alarm for a decade. There is no excuse for continuing to pretend that extremely vulnerable children haven't been irreparably harmed.

[Joe Gibbons @JZGibbons]

I co-founded Wikipedia, but an anonymous mob runs the show — and now I’m banned https://t.co/uV3Q35gkUc