We recently released a video demoing 4 applications of GPT models (ChatGPT).
Our first demo shows how ChatGPT can recommend items from a list based on a customer’s intent to purchase. This example, originally done in partnership with a neighborhood market, accepts a list of beers available near the checkout, and items purchased (received from a point-of-sale system). The output is a recommendation for one of the inputted beers.
For smaller players who may not need, or are not able to afford, a robust recommendation engine/loyalty system allows you to make a just-in-time recommendation to a customer. Whether this is integrated at a point-of-sale system in a real store, or near the checkout flow on an e-commerce store the barrier to entry is almost zero now.
While robust recommendation engines may do a better job of identifying and recommending better products, ChatGPT will better be able to articulate to the customer the value proposition of its recommendation.
The use-case for code generation, iteration, and testing is well covered by the plethora of tools (e.g. GitHub Co-pilot). I may be wrong, but I believe that prompt engineering and partnering with AI tools will be key skills in the technical engineering landscape 5-10 years from now (+/- 7). For now, I’ll focus on the prompt engineering:
Through this, and the following, example we show the possibility to model ChatGPTs output. It doesn’t have to just output a text, in our case, it responds with a robust JSON. There are many ways to achieve this, we did a few-shot prompt that gave examples, clear instructions, and some validation logic to ensure it matched (and if not, we would retry with feedback).
Showing off the various areas of expertise, we see that ChatGPT is very capable of understanding spacial queries and designing areas for us. In this case, it acts as an interior decorator recommending items for us to place.
It would have been easy for us to just ask for suggestions and explanations (as we did in the code assistant), so instead, we stretched its ability to give us an embedded list of complex objects that would need to reasonably fit inside a real room and theme.
It did so perfectly, identifying a reasonable theme, furniture items, and formatting the response well. If you’re curious about our prompt engineer, leave a comment below or reach out to me at firstname.lastname@example.org.
Whisper Demo (Car Guru)
Whisper is awesome. The ability to apply AI and the broad knowledge to transcribe speech to text accurately is incredible. To be able to “talk” to a GPT and, with a delay, interact is awesome. We already see this capability being integrated into Copilot X to “talk to your IDE”.
The demo shows a simple case of it acting as a mechanic, but imagine a “robot” able to talk to you with intelligent, accurate responses in a shopping mall or a dealership. The future is cool and coming fast.
As we see these services expand, we will see them compose in different ways. With Whisper, the floodgates are opening for new inputs to these GPT models (API, voice-to-text, CLI, and more coming soon). And with plugins, the ability for 1-to-many outputs will change the game. Not only can we recommend furniture, but hook into retailers to add them to carts. Now that it can perform searches, book tables, schedule meetings on our behalf… Skilled labor is about to get easier and learning to optimize working with these models will be a critical skill.
These are four examples we created quickly to show off different aspects of working with GPT models. If you enjoyed please like the video, clap for this article, and follow-us to hear more. And if you want to have a conversation about AI, ChatGPT, or how we can you and your team please feel free to get in touch at email@example.com.