Obi Ohagwu

-

[D.K. @DanKantori]

本谷 利明

[Obi Ohagwu @ObiOhagwu]

@finbarrtimbers Distribution entropic difference is a goated distance metric tbh

[Obi Ohagwu @ObiOhagwu]

@konet Not any public repos I’m aware of. Although tinygrad does, it’s still limited to CPU. It’s being worked on though

[Obi Ohagwu @ObiOhagwu]

4-bit quantized models do seem to be pretty fast on CPU, but is the precision tradeoff against fp16 gpu based models worth it? This seems to only be an issue with smaller models (7B). > 35B models work great with 4-bit quantization though

[Georgi Gerganov @ggerganov]

Here is 4-bit inference of LLaMA-7B using ggml:

https://t.co/NSetwCti03

Pure C/C++, runs on the CPU at 20 tokens/sec (M1 Pro)

Generated text looks coherent, but quickly degrades - not sure if I have a bug or something 🤔

Anyway, LLaMA-65B on M1 coming soon!

[Obi Ohagwu @ObiOhagwu]

@ggerganov Interesting paper on linear scaling methods
https://t.co/vHOnnLPAc4
https://t.co/0G8EDEc2iw

[Obi Ohagwu @ObiOhagwu]

@ggerganov Nystrom self-attention type linear scaling? I wonder how you’d go about that. Will be following. Great work so far though!

[Obi Ohagwu @ObiOhagwu]

How open of you guys :)

[OpenAI @OpenAI]

Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: https://t.co/TwLFssyALF

[Obi Ohagwu @ObiOhagwu]

Nice development

[anton @abacaj]

LLaMA has been fine-tuned by stanford,

"We performed a blind pairwise comparison between text-davinci-003 and Alpaca 7B, and we found that these two models have very similar performance: Alpaca wins 90 versus 89 comparisons against text-davinci-003."

[Obi Ohagwu @ObiOhagwu]

@Kdenkss That’s a pretty optimistic outlook. More likely there’ll be a sharp divergence in what it means to be 1st world and 3rd world. One has all the compute, the other is an open mine for compute resources. Think Elysium

[Obi Ohagwu @ObiOhagwu]

Broke: Immanetize the Eschaton
Woke: Discretize the Eschaton
Bespoke: we like bit manipulation

[Obi Ohagwu @ObiOhagwu]

You call it depression, I call it a sub-optimal minima. Just tweak your optimization hyperparams

[Obi Ohagwu @ObiOhagwu]

Latent diffusion manifolds as Jungian artefacts…

[Davis Blalock @davisblalock]

"What Makes Convolutional Models Great on Long Sequence Modeling?"

CNNs—not transformers—now dominate the hardest sequence modeling benchmark.

Here's how this happened: [1/14]

[Glenn Marshall @GlennIsZen]

Twitter viewable version of my completed film, taking extracts from James Joyce's novel 'A Portrait of the Artist as a Young Man', used word for word as prompts, in the style of artist John Lavery, using #stablediffusion #deforum

[Obi Ohagwu @ObiOhagwu]

@Kdenkss Prompt: “A Yoruba man returning home after many years away:: in the style of Caravaggio, —ar 16:9”

[Obi Ohagwu @ObiOhagwu]

@Kdenkss You should give midjourney AI a try. Their model seems to be a lot better suited for picturesque/cinematic art. Prompt: “An Igbo man looks over the village he grew up in:: in the style of Caravaggio”

[Obi Ohagwu @ObiOhagwu]

@Kdenkss Prompt: “ A young Yoruba man writing his PhD thesis on the Reinmann hypothesis:: in the style of Caravaggio, —ar 16:9”

[Obi Ohagwu @ObiOhagwu]

@Kdenkss Prompt: A busy afternoon in an Enugu market:: in the style of Caravaggio, —ar 16:9”