A Glimpse of ChatGPT's Limitations
Exploring AI's Potential By Example
Today's news feeds are absolutely brimming with information about Artificial Intelligence changing the world. While there have been many years of changes, it's gotten public attention largely due to the massive adoption of ChatGPT and other AI products. It's a popular view that we're entering the next major technology revolution since the iPhone or maybe even the internet. It feels like everything is going to change and there is a lot of speculation about what it all means.
But it's important to note that this isn't magic, at least not yet. While so many amazing things are now possible that were never possible before, a lot of the previous AI challenges remain. GPT just introduces a much stronger reasoning engine to learn from and understand text.
I've spent the last week or two working through a lot of the problems that need to be solved to work with an AI. It's incredible what I've been able to do, but it still takes quite a lot of technical and data knowledge to put together a robust product. Let me help clarify some of the challenges with some examples, and you'll also see just how limited things are right now.
Introducing HomeGPT
Let's say we're buying HomeGPT, a bot that will help you make your home even better with the power of AI! The product promises to be able to improve the livability and optimize the layout of your home. But when we got the product it isn't quite as impressive as we thought it would be.
The first feature is a space optimization feature. You buy a new vacuum cleaner and ask HomeGPT to find somewhere to put it. HomeGPT learns about things the same way we do. You show it each room, and it rates each room to see which one is best for the vacuum to be kept in. Then, once it knows the best room, it asks you to open up all the storage spaces in that room, and it looks in there. Based on this, it tells you where the best place for the vacuum is.
But say there isn't enough space for the vacuum in that room. In that case, the space optimization feature finds items in the room that could be relocated out of the room. It takes the top 3 items that could be moved and goes back through and looks at all the other rooms one at a time. It then repeats this process a few times until the vacuum finally fits in its right space and all the other items fit in their correct spaces.
This shows the first issue with HomeGPT, it can't remember your entire house to be able to answer complicated questions about your home. It can only remember a bit about each room at a time. This limitation is also called "token length" in terms of how much data GPT can process at once. Until that limit grows, it's unlikely for HomeGPT to make a holistic evaluation of your home, meaning it could settle for a good solution instead of finding a great solution.
Going further, let's say that HomeGPT has a few different attachments to know what's happening with each room. These take photos, measure humidity, check for lighting, and collect other data about your home. The more "senses" that HomeGPT has, the less of the room it can view at one time because it can't hold all this extra data in its head and see the entire room. Without that extra data, your vacuum might get damaged by water in the bathroom pantry or under the sink.
The data HomeGPT can see helps it make better decisions, just like you or me, but that further limits its ability to see the entire picture. Once impressed with HomeGPT we want to push it further and ask more questions. The questions keep getting bigger and broader: "Where do I put the couch to get the most light?" "How do I get more floor space?" "What rug would match my furniture?" "How can I save time in the morning?"
Frustratingly, HomeGPT can show us magic with some of these answers and completely fails at other tasks. It starts to feel a little random what sort of home optimizations it's capable of. Great idea, but it is a little frustrating to work with. Hopefully, they will make it a better product in the future.
What this story tells us about limitations
HomeGPT helps show us some details about where the limitations are with ChatGPT and other AIs reliant on Large Language Models. We noted:
- We can't show it everything
- It only knows about what we show it
- It only sees what it has the capability to see
- The more in-depth the question, the more capabilities it needs
That last one is exactly why ChatGPT introduced Plugins last week. For these models to be empowered to answer a wider range of questions, they need more capabilities to "see" more and leverage existing tooling to give it data it understands. OpenAI is also working to extend GPT-4's token size so that soon it will be able to make decisions based on more data.
But the data still matters. Data Scientists will tell you that one of the hardest aspects of their work is getting good data sources that they can use to make decisions. The information and the prompt that you feed these systems need to be engineered to be understood the same way a human would look to understand them. The structures can be very complex but still require some structure.
While ChatGPT introduced the world to a rationalizing engine, it'll still be a long time before it can do magic and perform miracles. Until then, we'll have to keep organizing our homes and doing our best to keep up.
 | A guest post by
|