A couple of weeks ago Microsoft announced fine-tuning OpenAI models was available in the Azure OpenAI platform. Like everyone else, I got excited — I had been waiting for this for quite some time. And then I looked at the how much it costs to fine-tune a model 😱.
My knee-jerk reaction was something along the lines of “how dare they charge us that much”. But then I ran the numbers and compared the Azure OpenAI offering with that of OpenAI, and ended with “huh, once you go over a couple of hundred million tokens, it’s actually quite reasonable”.
Long story short, for smaller projects with fewer security concerns OpenAI is more cost-effective; however, when token usage exceeds a few hundred million, Azure OpenAI becomes the better choice for most models.
If you want to see see how I've reached that conclusion (and maybe check my assumptions), the in-depth analysis is available on my blog.
Math is wonderful.
Had a lot of fun at Codecamp Iasi last week, talking about how large language models work and how to connect GPT-4, the fairest of them all, to pretty much the Internet. Using Google, of course 😉.
If you want to get a sense of how that went and maybe try and experiment with building a simple chatbot yourself, I’ve posted the slides and notebook here.
Fuyu-8B multimodal was released a couple of days ago, and it's quite impressive for such a small model -- it's both accurate and fast, its only downside being the license (non-commercial).
See below how it handles a series of questions about Dancing Darth Vader 🥳.
Read more about it on the Adept.ai blog and try it out here. Still waiting for transformers to include it in their next release
Cheers,
Vlad
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.