Llama: update history

This commit is contained in:
Akemi Izuko 2024-02-11 21:30:59 -07:00
parent 1ca736dad5
commit 410573e03d
Signed by: akemi
GPG key ID: 8DE0764E1809E9FC
3 changed files with 21 additions and 3 deletions

View file

@ -1,7 +1,7 @@
--- ---
title: 'A Brief History of Local Llamas' title: 'A Brief History of Local Llamas'
description: 'A Brief History of Local Llamas' description: 'A Brief History of Local Llamas'
updateDate: 'Jan 01 2024' updateDate: 'Feb 11 2024'
heroImage: '/images/llama/tiny-llama-logo.avif' heroImage: '/images/llama/tiny-llama-logo.avif'
--- ---
@ -161,4 +161,19 @@ the turning point where open source models finally break ahead of commercial
models. However as of writing, it's very unclear how the community will break models. However as of writing, it's very unclear how the community will break
through GPT4, the llama that remains uncontested in practice. through GPT4, the llama that remains uncontested in practice.
#### Early 2024
This is where we currently are! Hence, things are just dates for now. We'll see
how much impact they have in a retrospective:
- 2024-01-22: Bard with Gemini-Pro defeats all models except GPT4-Turbo in
chatbot arena. This is seen as questionably fair, since bard has internet
access.
- 2024-01-29: miqu gets released. This is a suspected Mistral_Medium leak.
Despite only having a 4bit-quantized version, it's ahead of all current
locallamas.
- 2024-01-30: Yi-34B is the largest local llama for language-vision. LLaVA 1.6
based on top of it sets new records in vision performance.
- 2024-02-08: Google releases Gemini Advanced, a GPT4 competitor with similar
pricing. Public opinion seems to be that it's quite a bit worse that GPT4,
except that it's less censored and much better at creative writing.

View file

@ -1,7 +1,7 @@
--- ---
title: 'Local Llama Quickstart' title: 'Local Llama Quickstart'
description: 'A collection of guides to get started with local llamas' description: 'A collection of guides to get started with local llamas'
updateDate: 'Dec 31 2023' updateDate: 'Jan 28 2024'
heroImage: '/images/llama/llama-cool.avif' heroImage: '/images/llama/llama-cool.avif'
--- ---
@ -49,6 +49,7 @@ anyone looking to get caught up with the field.
- [Karpathy builds and trains a llama](https://www.youtube.com/watch?v=kCc8FmEb1nY) - [Karpathy builds and trains a llama](https://www.youtube.com/watch?v=kCc8FmEb1nY)
- [Build a llama DIY by freecodecamp](https://www.youtube.com/watch?v=UU1WVnMk4E8) - [Build a llama DIY by freecodecamp](https://www.youtube.com/watch?v=UU1WVnMk4E8)
- [Understanding different quantization methods](https://www.youtube.com/watch?v=mNE_d-C82lI) - [Understanding different quantization methods](https://www.youtube.com/watch?v=mNE_d-C82lI)
- [LLM workshop](https://github.com/mlabonne/llm-course)
#### Servers #### Servers
- [Ollama.ai](https://ollama.ai/) - [Ollama.ai](https://ollama.ai/)
@ -58,6 +59,8 @@ anyone looking to get caught up with the field.
#### GPU stuff #### GPU stuff
- [Choosing a GPU for deeplearning](https://timdettmers.com/2023/01/30/which-gpu-for-deep-learning/) - [Choosing a GPU for deeplearning](https://timdettmers.com/2023/01/30/which-gpu-for-deep-learning/)
- [Last gen servers for
AI](https://www.kyleboddy.com/2024/01/28/building-deep-learning-machines-unorthodox-gpus/)
#### Cloud renting #### Cloud renting
- [Kaggle](https://kaggle.com) - 30h/week free, enough VRAM for 7B models - [Kaggle](https://kaggle.com) - 30h/week free, enough VRAM for 7B models

View file

@ -29,7 +29,7 @@ const posts = (await getCollection('unix')).sort(
</h1> </h1>
<p <p
class="text-center mb-4 text-lg"> class="text-center mb-4 text-lg">
Over the past few years, I've extenstively configured at Over the past few years, I've extenstively configured and
learned about using MacOS, Linux, and even some BSD learned about using MacOS, Linux, and even some BSD
systems. Along the way I documented some of the fun aspects systems. Along the way I documented some of the fun aspects
of learning Unix. The code for most guides is part of my of learning Unix. The code for most guides is part of my