Tell us about yourself. What are you working on right now?

My name is Wenyu Zhang, and I’m the founder of an early-stage startup. We are building software that uses AI to train employees better and faster on standard operating procedures and also train AI assistants to help with those procedures. Before this, I was the head of product at a startup called Copy.ai.

How do you use AI for work?

I use it almost everyday as a co-pilot for a lot of different tasks. I’m not someone with years of UX-design experience, but as a founder, I have to wear many hats. A lot of times when I’m designing something, I chat with ChatGPT to learn basic principles of UX design but also to dig a little bit deeper. “When does it make sense to apply this principle and when doesn’t it?” Or, “Can you give me examples of products that execute this well and products that don’t?” I use it like a learning assistant.

Sometimes, I use it for mundane things that otherwise would take a lot of time. For example, I needed to find a color that matched a color palette. I had a list of the hex codes and I said, “I need a purple that is between the primary and secondary color.” Using ChatGPT was fast and saved time.

I also use it for researching specific business areas where I’m trying to learn more about distribution models or sales models of different companies. It summarizes a lot of things, which is helpful, as long as you dig deeper yourself.

How did you start using GPT as a learning tool?

It started off replacing a lot of my Google searches. I would be in the process of doing a task and need to think through something. In the past, I might do some Google searches, watch some videos, or talk to my friends who are experts in that area. (I still do. Talking to an expert can be better than ChatGPT because an expert can ask you questions.) Now, I go to ChatGPT instead and see what it says as a starting point.

How do you create a prompt for a new topic that you want to learn about?

I start freeform. Based on what it says, I realize I need to give it different pieces of information. I iterate through it. Usually, my initial prompt is really impulsive. Then, I slowly give GPT more structure and detail about the situation. It’s like a conversation.

How do you revise your prompts?

One of the fastest ways to learn is getting more practice. Someone explaining the principles behind a situation is not as helpful as putting yourself in that situation ten times. I ask ChatGPT for examples, for scenarios, to demonstrate things. Then, I ask it to debate with itself so I can see multiple perspectives and decide for myself. By telling ChatGPT that it has expertise and giving it positive reinforcement, the answers can get better.

Once I asked ChatGPT, “Can you give me a list of companies that have standard operating procedure documents?” It gave me a list of a bunch of types of companies, but its interpretation of “standard operating procedure” was companies where there was a strong regulatory need for compliance. I didn’t want to limit my exploration to that space just yet, but I realized that maybe [its response] was my fault. Maybe “standard operating procedure” is used more commonly with compliance, and I’m [not] using a word that means what I intend. So I coached it: “You have great knowledge of many different industries.” (I wanted to give it positive reinforcement.) Then I was like, “However, the list seems heavily focused on compliance, and we're looking for many other motivations and use cases.” Then, I gave it an example. “In this operations company, where they hire new grads, standard operating procedures is a way to expand their workforce to conduct the task.” Examples help define the range of the pattern I was looking for. I kept iterating and eventually got a long list of different use cases.

How would you describe the value ChatGPT has for your business?

When I was at Copy.ai, there was a user who had the best explanation of GPT at that time. (I don’t think this explanation is as accurate of where the tech is now after the instruct models and heavy use of reinforcement learning, but I think it’s still a useful way to understand the core technology.) This user said, “Imagine your drunk uncle or aunt at Thanksgiving, who happens to be incredibly well read and knowledgeable and worldly, but they’re kind of drunk and ranting, and you have to poke them in different directions.” That is my mental model of it. The more you understand what it’s doing and also its limitations, the more you get out of it.

Do you have a hot take on generative AI?

In the future, the way you teach people and the way you train AI will be quite similar. We’re going to need a way for people to train hybrid AI-people teams, and that’s what I’m working on. I’m excited for someone joining a team, doing new employee training with an AI learning system, then, as they progress in their career, for them to contribute to the knowledge that trained them.

I think vector databases and retrieval augmented generation have already done a lot of work in adding a memory layer to the basic inference, and it will continue to get more advanced. Something that’s not entirely out of the box yet is the ability to guide when it should be creative in inference and when it should stick to the script, or how to work with hybrid structured and unstructured data. We’re going to see products where the architecture is a hybrid of structured prompts combined with sandboxes where the models are encouraged to get a little bit more creative.