Data Machina #195
LLaMAs, Alpacas & more LLM Animals. OpenChatKit in consumer GPUs. Cerebras-GPT: New open LLM models Apache 2.0. Flash attention in PyTorch 2. BloombergGPT. ChatDoctor. Neural Graph DBs.
On LLaMAs, Alpacas, and other friendly LLM Animals. Since Meta AI LLaMA was born, and the model weights escaped the stable, a lot of new LLM animals carrying the LLaMA genetic code were born in the wild. Like the smaller Alpacas…
The LLaMAs and the Alpacas significantly reduce the cost of training, finetuning, and using LLMs without much model performance degradation. I’ve been tracing & tracking these LLMs animals. Here are a few I encountered just this week:
Vicuna: An Open-Source 13B param chatbot model (blog, repo, demo.) “According to a fun and non-scientific evaluation done by GPT-4, Vicuna-13B achieves more than 90% quality of ChatGPT and Google Bard while outperforming other models like LLaMA and Stanford Alpaca in more than 90% of cases.”
LLaMA Adapter (repo): A lightweight method for fine-tuning LLaMA models with Alpaca dataset within 1 Hour and just 1.2M Parameters
Alpaca CoT (repo): Boosting Alpaca reasoning ability with Chain of Thought (CoT.) An Instruction fine-tuning platform with instruction data collection and a unified LLMs interface
Lit-LLaMA (repo): Implementation of the LLaMA based on Karpathy’s nanoGPT. Supports flash attention, quantization, LoRA fine-tuning, pre-training. Apache 2.0 license!
GPT4ALL (repo, data, demo): a chatbot trained on a massive collection of clean assistant data including code, stories and dialogue. Trained with ~800k GPT-3.5-Turbo Generations, based on LLaMa
LION OpenFlamingo: a “reproduction” of DeepMind Flamingo. OpenFlamingo is an open source framework for training vision language models with in context learning. Built on top of frozen LLaMA and CLIP models
Brief notes on AI-Pair Programming. Several Sr. Dev/Engineers I know tell me that they use GPT-4 and Copilot X as their primary AI-Pair coding tools in their day job. They say that -notably- they are performing fewer and fewer searches in Stack Exchange, Google and Github repos. Yet, I know many other Dev/Engineers that are still quite dismissive of LLMs and AI-Pair programming. Let me share my latest notes:
@simonw - a Sr. Dev- wrote about how AI-enhanced development makes him more ambitious, more productive. He says he uses ChatGPT a lot. He describes how he used ChatGPT to build a system to archive ChatGPT messages.
@geoffreylitt reflected on the current state and future of software development, and the pros & cons of using LLMs (and GPT-4 specifically.) This is a great read: Malleable software in the age of LLMs
@bradgessler’s blog post on how GPT-4 can be helpful for both junior and senior devs, and its limitations. Read more here: Pairing with GPT-4.
@chrisdias posted about how Codex models enable the editor in chat and the embracing of the chat view for coding in Visual Studio Code and GitHub Copilot
The ephemeral (dynamic?) UI/ IDE powered by AI is a new paradigm… Checkout: $NAME: An AI-powered IDE: Developers describe what they want to build by writing plain English. Then let AI agents with access to tools do the coding work
@PerfectHQ introduced on demand AI Functions in Marvin: A batteries-included library for building AI-powered software. These AI Functions differ from conventional ones in that they don’t rely on source code, but instead generate their outputs on-demand through AI
@mckaywrigley released AI Code Translator, an easy way to translate code from one language to another
In Developer Tools 2.0 the legendary Sequoia VC mapped out AI in the s/w development cycle. This is a good indication on what the VC money is chasing in the AI landscape.
Having a lazy Sunday? Here is some AI entertainment for you:
Listen to the AI Rap Battle - ChatGPT vs Google Bard
Have a nice week.
10 Link-o-Troned
the ML Pythonista
the ML codeR
Deep & Other Learning Bits
AI/ DL ResearchDocs
El Robótico
data v-i-s-i-o-n-s
MLOps Untangled
AI startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.