AI craze
Like most people there’s an urge to jump on a craze purely so you can talk to people about, but once you’ve been in tech long enough you tend to try and ignore the fads and the new stuff at least until they’ve been around a bit.
Safe to say the AI craze is in full swing and now that the peak has dropped a little I thought I would take a look at see what it’s all about.
I’ve been using chat GPT for a while now and it can be very helpful for some things, particularly when learning a new programming language or framework, it can answer very specific questions very well, and at least give you an idea where to start and at best it will give you usable code. I have also been using Github copilot for work, once you figure out where its useful, its great for completing a line of code and generating tests, writing a comment about what the code should do first seems to help get better results, I have noticed it is getting better at using the context of the code file you are in, I expect it will get significantly better in the future.
There are a ton of apps that transcribe audio clips to text or let you question a pdf or generate cartoon avatars using a picture of yourself, and these are all essentially wrappers around the openai api, the api themselves are simple and easy to use with only a handful of configurable parameters to play with, the magic happens in the prompting, calling it engineering might not be a stretch there’s definitely a lot to learn to get your desired result. I took the prompt engineering course on deeplearning.ai and found it very useful.
OpenAI isn’t the only show in town either, there are lots of good solutions for self hosting a VM or serverless access to a GPU with runpod and replicate to name a few, that make it fairly easy to setup. Cloudflare is currently in beta with its workers AI solution. I have created a prompt to generate articles using openai text generation, and testing with other models llama 2 and mistrel noticed very good results from both open source models.
Fine tuning a model for a specific use is also very simple, certainly with openai anyway, generating the data will be the trickiest part, a few api calls to train the model.
We are also now seeing tools like langchain built on top of LLM’s make it easier to build applications and abstract away more of the LLM specifics.
For a layman developer it seems possible to be able to build something pretty cool and useful without having an intimate knowledge of ML or vector math, what’s the old saying,
Anything that a AI can be used for a developer will build
