Learning AI Poorly: Mama llama code my... next project?

(originally posted on LinkedIn)

Normally, on Thursday morning I take a look at the open tabs and various screenshots I took through the week regarding AI or ML and write something. Usually it works out, but this week for some reason all I had was a ton of “is AI conscious” garbage and even if I might have purchased a domain name I don’t have anything useful to contribute.

But, guess what? Meta released Code Llama, a state-of-the-art large language model for coding which is a Large Language Model that will write software for you. neat. Does it work? yes. it does. Is there a paper to read? yes, right there on the announcement. Can you use it right now? errrr. maybe? If you have a mac, you can.

There is a project called llama.cpp that came about in an effort to do inference “at the edge” or sort of efficiently do transformer inference on regular computers or devices. If you want, you can download the code, build it on your mac without too much trouble and then run inference on a variety of models. The beauty of this is that there’s a large community of people who work on this project. Within a few hours, TheBloke converted the model so it could be used on llama.cpp with just a small patch.

How about the code it generates? Well, subjectively it is pretty good at code completion and writing small functions (bubble sort an array of integers in python) but it isn’t going to help you build a complete project. At least not yet.

In other news today, HuggingFace raised a $235M series D round at a $4.5 billion valuation with participation from Nvidia, Intel, Google, Amazon, AMD, Qualcom, IBM, Salesforce and Sound Ventures. OpenAI opened GPT-3.5 Turbo for fine tuning which means customers can tune the model with their own data to improve inference for custom use cases. ideogram.ai launched this week with a $16.5 million seed round. They do generative AI with a twist - They can put legible text within images. And finally, the first person who paid me to write code was awarded a new NIH grant to improve diagnostics of Aortic Stenosis using AI and 4D flow MRI awesome!