Live event at FabLab Iasi, an awesome generative AI compendium, sample code, and OpenAI assistants in Azure
Hey there friends, long time no see!
I’ve just realised Christmas was like, two months ago, and now I’m depressed 🥶. Luckily, a lot of new & interesting stuff has been happening, so that’ll keep all of us busy for a few more months until summer hits.
Take a look:
First things first, in less than two weeks I’ll be talking about the mechanics of large language models and RAG at Fab Lab Iași, delving into how large language models work, how to use them successfully, and how you can write the world's simplest chatbot with just Python, Jupyter Notebooks, and Azure OpenAI.
The event is here if you want to join.
So many language (and multimodal) models coming out 😮💨! So, I've decided to write down the most interesting ones down on this page right here — https://vladiliescu.net/awesome-gen-ai/
It's very much incomplete.
Everything's still in flux, but this is the inclusion criteria I've settled on for now:
1️⃣ A model needs to be interesting to me personally, not necessarily beat by 0.1% whatever llama finetune was on top of OpenLLM leaderboard at the time of its release.
2️⃣ A model needs to be available for commercial usage. I need to be able to use them to generate code/ideas/whatever that may or may not generate revenue. That excludes models like Orca-2 (but not Phi-2!) which can only be used for “non-commercial, non-revenue generating, research purposes”.
3️⃣ I need to be able to run the model locally, on a Mac M2 Max. This means either smaller models (34B tops), or something that runs in llama.cpp. That excludes Qwen-72B.
Take a look, and let me know if I'm missing something.
In other news, I've just updated my article on using Whisper, GPT-4, DALL*E 3, and Text-to-Speech to also support their Azure counterparts. It’s so cool being able to run this from the Azure infra (but not being censored by Azure’s aggressive content filter, I yearn for the day when they’ll turn that off).
While the switch is relatively straightforward, it's important to note that each model's availability is restricted to specific Azure regions. To make the transition smoother I recommend using dedicated Azure OpenAI clients for each model, to avoid having to keep track of which model is available in which region.
Here are find the sample code plus a summarized chart of the Azure regions where each model is available.
Last but not least, some cool news from Azure OpenAI have dropped recently, and this is my favorite bit -- the Assistants API is now in public preview! 🔥
This means you'll be able to create more dynamic, and most importantly stateful copilot experiences in your applications without the heavy lifting. The API simplifies managing conversation states and supports a variety of formats and tools, such as a Code Interpreter for Python and function calling for app integration.
There's a bunch of other (mostly coming soon) stuff in there as well 🤩.
I'm really excited about the new embeddings models that will be released later in the month (higher capability and performance, all at a lower cost, you know the drill), but the new text-to-speech models sound great (heh) as well.
You can dive into the details over here.