Training the Perfect AI Assistant

For AI to get real work done for you, you have to leave general use and get something made custom for your problem. 

It only makes sense: AI is powerful because it can be a second brain, it can do things only people could do before. But if you hire someone, you have to train them. You can’t just drop someone off with a general degree in your business and expect them to get things done. So, let’s look at two situations for training a person.  

Option 1: A long training period

You have someone apprentice. They watch everything you do for a while, just taking in the basics. Then, you have them do small tasks while you supervise. You correct their mistakes and commend their successes. At the end of this long process, you’re left with a very competent, independent worker. Now, in the trades, they might just leave you and take a higher-paying job with their new skills, but that’s stretching the metaphor. These steps are extremely close to those you take in training an AI. In AI, the process I just described is training.

Just like with a real employee, apprenticing can be expensive, time-consuming and overkill. You don’t need to apprentice someone to work at McDonalds. In fact, if you take someone who is generally competent, you could hand them a short employee training pamphlet, put labels on the workstations, and then turn them loose. And that’s where our second option comes in.

Option 2: On the job training

You take a person with general competence, and then you hand them all the information they need to do their specific task.  In this case, you don’t really need them to become excellent at their job. They just need to be able to do the bare minimum. In college, I took a part-time job working at a photo counter. I didn’t need to know how to make all of the things they made, like calendars, photo books and canvases. Everything was labeled clearly, and there were detailed instructions on how to make each thing. They put me behind the photo counter because I was generally competent enough to do reading. And then I would get feedback on the specifics: we’re talking small adjustments, not the lengthy, supervised learning that an apprentice gets.      

Which is Better?

These are the two approaches to getting an AI to do real work for you. You can use fine-tuning, where you take an AI and put it through an in-depth training regimen where you give it a lot of examples, and you supervise its learning, and you give it a lot of time. At the end, you will have an excellent tool for working on *exactly* what you’ve trained it on. It will be mediocre on anything else.   

It’s a really strong approach for things that are rigid, unlikely to change. If you need an AI to generate articles in a certain style and on a certain set of topics, this method works wonders. If you’re building an AI-powered customer service bot, this method is too rigid. You need something that can adapt to updated policies and situations.  

The alternative is an AI that is generally competent but is given the specific information it needs to solve problems. The AI gets feedback on its responses to improve them over time, but it can largely do the job out of the box. It’s great at doing a lot of things: Q & A, knowledge bases, and, well, most other places you’d want a non-general AI. It’s not as strong at things like generating articles in a style, but it’s a lot more flexible than the rigid AI.  

This whole approach is something we’re calling Silk-Tuning, and it will be the cornerstone of Odobo, our powerful AI assistant.

Let’s look at the existing tools for both approaches and how they work.  

Long-Term Training

For training, there are a lot of tools, but it depends on how much work you want to do. For the best results, you want to train an AI from scratch. This means you’d be building something like ChatGPT from the ground up. Well, there’s a reason why they say that perfect is the enemy of good. It took around six months for the latest version of ChatGPT to be trained, and that’s with millions of dollars of hardware working on it.   

For most people, you’re going to do something smaller scale called fine-tuning. You take an existing, general AI (like ChatGPT), and then you give it a lot of training data and time. In the end, you’ll have an AI that’s really good at doing what you trained it to do.   

To give you an example of fine-tuning, a year ago, we trained a weak version of ChatGPT to generate articles based on a headline. It took a little over 100 examples, and under $100 worth of training. It got really good at generating a specific kind of article. It also had a lot of problems, like the fact that you have to pay a premium to actually talk to a fine-tuned model if you go through OpenAI. It was also rather inflexible, and it needed you to come up with the titles. Still, it fit the application very well.

Silk-Tuning

Now, for Silk-Tuning, there are some existing tools, but most are missing pieces. You can have an AI look at documents you have and reference those when you talk to it. ChatGPT has that now, even. Most tools have one of two problems (and ChatGPT has both): 

• You can only use a limited number of documents.
• You’re putting all of your documents into a cloud service.

The second one is really limiting, because this tech is perfect for doing things like helping you with your taxes and reviewing medical documents. Exactly the sorts of things you don’t want to be uploading into the cloud. And the limited number of documents really gets in the way, too.  
There are solutions out there with unlimited documents that run entirely on your local machine. They are very slow, and very technical. I’ve run some experiments on them, but it’s not the sort of thing a regular person is going to be able to pick up and use. That’s why we’re building Odobo. It’s an AI assistant built around Silk-Tuning that knows all about you. And it runs on your machine (it doesn’t need the internet), you don’t have to worry about giving it your information. If that sounds interesting to you, you can join our wait list. 

---