You can enjoy a lot of things with just one map. This time, I asked ChatGPT a slightly tricky question.
It is known that SOM can be used for strategic planning of companies. In that case, decide which area to attack next based on the positional relationship between your own area and your competitors’ area on the map. If SOM represents numerical data, the “actions” are defined by examining the node values of its target area on the map.
In the case of SOM for unstructured data, we are testing how far we can go with the help of LLM. We randomly selected a node on the 1000 Business Ideas map and found 11 Nearest Nodes at that node. We presented 10 ideas and their PC scores to ChatGPT, and for the 11th, we presented only the PC score and asked about the content of their business ideas.
Here is ChatGPT’s answer:
Business Idea: AI-Powered General Pet Wellness Service
Description: A comprehensive service leveraging ChatGPT to offer general pet wellness advice, including basic behavioral tips, dietary recommendations, and routine health care strategies. This service aims to be a one-stop solution for pet owners looking for accessible, reliable information to maintain their pets’ health and happiness.
Target Market: General pet owners looking for accessible, reliable pet care advice without the need for highly specialized consultations or training sessions.
Revenue Model: Given the generalist approach and the emphasis on accessibility suggested by the PC scores, a freemium model could be ideal. Basic advice and standard wellness tips could be offered for free, with a subscription model for access to more detailed guides, personalized wellness plans, and interactive features like real-time advice or emergency care tips.
Here is actual idea in the data:
Business Idea: AI-Driven Virtual Pet Health Tracker
Description: A platform that utilizes ChatGPT to monitor and advise on pet health, offering personalized care schedules, dietary recommendations, and activity tracking tailored to each pet’s breed, age, and health condition.
Target Market: Pet owners seeking to optimize pet health
Revenue Model: Freemium model, premium features for detailed health tracking and consultations
ChatGPT has been known to include hallucinations in its attempts to answer any question, but this one seems to have a reasonably good answer. However, after trying it with other nodes, it seems that if there are few examples presented, or if the node you want to predict is in a different cluster from the example node, you can get very different results.
LLM uses embedding vectors to perform logical operations on the meaning of words, so using embedding vectors would likely improve accuracy. However, we found that LLM can also perform calculations using PC scores obtained by reducing the dimensionality of embedding vectors.
What we learned from this series of experiments is that text information in a specific situation uses only a small portion of the semantic space that LLM has, so it can be adequately represented in a space with significantly reduced dimensions.
In this experiment, the information presented to ChatGPT is only the 10 ideas and their PC scores, and ChatGPT is not informed of the entire map. Still, it seems that ChatGPT can make a semantic inference, at least with some basis.