How to pick the right Foundation Model for your AI project

What to think about when choosing an AI model

Picking the right AI model is kind of like selecting the right tool for a job—it’s got to be just the right fit. You've got to think about things like what the rules are for using it (that's the licensing part), whether you can tweak it to suit your needs, how it handles private data, and what its technical chops are. We’re going to dive into these things so you can get a sense of what to look out for when choosing a model that’s not going to give you a headache down the line.

At the time of this note, the website Hugging Face boasts a staggering inventory of 371,380 models ripe for the picking, all nestled within an open-source ecosystem bustling with shared AI innovations.

When I asked Bard and ChatGPT, they could only get as far as 180k models (including all the variants). So in a space of just 2 years (the cutoff date for ChatGPT), the number of models have doubled! Of this impressive array, more than 30,000 models specialize in the art of crafting text by the way.

Yet, this number doesn't even touch on the heavyweights of the text generation league, such as the GPT series from OpenAI or Google's Bard. With such a wide field of options, how does one pinpoint the ideal model tailored for their specific needs?

Finding the perfect pre-trained model

Imagine you're in a giant library filled to the brim with AI models—think of Hugging Face and its massive collection. With so many options, finding the one that fits your project like a glove means knowing what to look for.

All about licensing

Licenses are pretty much the rulebook for how you can use a model. Open-source licenses are usually pretty chill—they let you change things up and even use the model for making money as long as you follow a few basic rules and give a shout-out to the original creators.

Proprietary licenses are the strict parents of the licensing world. They lay down the law on what you can and can't do, which can be anything from no making changes to paying up if you’re going to use it for business.

You really need to read the fine print. Some licenses are cool with you using the model for personal stuff or non-profit work but draw the line when you start making money, especially when your user base grows. So, getting to grips with the terms is key to playing by the rules.

Open-Source vs proprietary models

Open-source models are like an open book—you can check out what’s inside, play around with it, and fine-tune it to your heart’s content.

On the other hand, proprietary models from big AI shops like OpenAI or Google are more like black boxes. They don’t let you see or change how they work, but they usually perform really well because they’ve had a lot of money and smarts thrown at them. These companies protect their investment by setting up paywalls that control who get to use their models and how.

The privacy thing

How a model handles your data is super important. With open-source models, you can keep all your data in-house, which is great for privacy.

But with proprietary models, you often have to send your data over to the company's servers, which could be risky if you’re dealing with sensitive stuff.

Tech Specs: the nitty-gritty details

When you’re evaluating a model, you've got to think about what it was trained on and who made it. Models trained on iffy data can be biased or raise ethical flags. The model's origins, like where it was made and who made it, can also influence your choice.

Make sure the model fits your project, too. An AI that’s great at understanding language might be useless for recognizing or creating images.

Performance and the planet

You want to check out how fast and accurate the model is and whether it can handle scaling up. How much power it uses also matters because of its environmental footprint—something we’re all starting to pay more attention to.

Can it handle the real world?

Robustness is about whether a model can keep up its performance no matter what you throw at it. Generalizability is about how well it can take what it’s learned and apply it to new stuff. These are crucial if you need something reliable.

Going further: fan you make it your own?

If you’ve got a really specific project, you might need to train or fine-tune the model on your own data. Check if that’s something you can do.

Wrapping up

As AI tech, like ChatGPT gets more common, there’s a lot of talk about how to use it responsibly. Laws and guidelines are trying to catch up with the tech, shaping how we should use these tools properly. It’s important to stay on top of these changes to make sure you’re using AI in a way that’s not just smart, but also right.