Cade Brown 쫞

Programming, Music, Math

[Cade Brown 쫞 @dev_ceb]

@karpathy GitHub + https://t.co/KXqecXO0Io + PorkBun. Lets you write markdown/mdx/astro components and compiles to HTML. Best DX by far

[Cade Brown 쫞 @dev_ceb]

@karpathy i’m imagining a Tails-like OS on a USB thumb drive, with weights on it, precompiled binaries of llama cpp. boots to REPL on anyone’s machine

[Cade Brown 쫞 @dev_ceb]

@pickover @genekogan “We live on a placid island of ignorance in the midst of black seas of infinity, and it was not meant that we should voyage far” - lovecraft

[Cade Brown 쫞 @dev_ceb]

@DrJimFan Also, newer implementation: https://t.co/6UJh0gP2k6

[Cade Brown 쫞 @dev_ceb]

@DrJimFan Although this paper offers new theoretical framings and cool diagrams, it’s unfortunate they didn’t compare against existing methods in this space.

NNCP was written by 1 person, 2 years ago. And it gets similar ratio to 70B…

[Georgi Gerganov @ggerganov]

sam.cpp 👀

Inference of Meta's Segment Anything Model on the CPU

Project by @YavorGI - powered by https://t.co/jFknDoasSy

[Cade Brown 쫞 @dev_ceb]

@fchollet Cool feature, but why use tensorflow at all at this point? If using modular components is the goal then PyTorch is already king…

not a TF user but torch.compile should also give better performance for frozen models

[AI at Meta @AIatMeta]

Introducing SeamlessM4T, the first all-in-one, multilingual multimodal translation model.

This single model can perform tasks across speech-to-text, speech-to-speech, text-to-text translation & speech recognition for up to 100 languages depending on the task.

Details ⬇️

[Jim Fan @DrJimFan]

IDEFICS, a 9B and 80B open-source reproduction of DeepMind's Flamingo by the @huggingface team.

Flamingo was an early milestone of multimodal foundation models that support arbitrary interleaving of image and text. IDEFICS is not quite at DeepMind's level yet, but outperforms an earlier community effort (OpenFlamingo).

Another gem is the OBELICS dataset, an open-source collection of image-text web documents: 115B text tokens & 353M images.

Blog: https://t.co/gQF6VM6sKB
Demo: https://t.co/TzFvzdI5T7
OBELICS: https://t.co/1M5MXVPB56

[WizardLM @WizardLM_AI]

🔥🔥🔥
We share you the online demos of all the WizardMath models:

🧙‍♀️WizardMath-70B: http://47.103.63.15:50083/
🧙WizardMath-13B: http://47.103.63.15:50082/
🧙‍♂️WizardMath-7B: http://47.103.63.15:50080/

❤️Welcome everyone to use your professional and difficult instructions to evaluate WizardMath, and show us examples of poor performance and your suggestions with our Discord (https://t.co/jjYGPeAEqq) or Twitter comments.

✅Please use the same systems prompts strictly with us, and we do not guarantee the accuracy of the quantified version.

🚫For the simple math questions, we do NOT recommend to use the CoT prompt.

🏇Please refer to (https://t.co/OwvWSO0XUx) for the latest available demo URLs.

[AK @_akhaliq]

BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis

abs: https://t.co/XQFMthf3oP
project page: https://t.co/4xb5xihmcb

[Cade Brown 쫞 @dev_ceb]

@pikuma @vinc17 Except on CRTs, QD OLED, and basically any display without pixel perfect rendering enabled… still cool!

[Cade Brown 쫞 @dev_ceb]

@ID_AA_Carmack For MMA, consider round off and symbolic symmetry of dot product operator, which can simplify lookup table size and access.

[Cade Brown 쫞 @dev_ceb]

@ID_AA_Carmack Integer rings with lookup table. u8 mapping as 256 unique fp32 values. Common op graphs can be “pre compiled” at exact precision and stored in small lookup tables. So, runtime kernels are simple integer accesses and bit ops

[Yannic Kilcher 🇸🇨 @ykilcher]

Men who use LARGE language models, is it possible that you're compensating for something?

[Cade Brown 쫞 @dev_ceb]

@ezyang Thanks for the link! I'm also wondering if you've ever heard of discrete unique ring-like structure (i.e. not based on IEEE). For example, an int8 representing 256 elements chosen to minimize error for common inputs. Ops/subnetworks based on computed tables with integer arith?

[Cade Brown 쫞 @dev_ceb]

@ezyang We are nearing that point with newer formats…

[Cade Brown 쫞 @dev_ceb]

@ezyang I often wonder about ML math using non-floating point formats, for example an int8 that has a table look up and operators but the values themselves are basically indices

[Cliff Pickover @pickover]

Here is a list of the books I've published, providing an eclectic learning journey and spice for the imagination.

So far, which book did you like best? (Select image to magnify.)