🚗 There have been a lot of funny adversarial attacks on car dealership chatbots lately. Things like having the car dealer’s chatbot write a python script for you, or convincing it to do math that results in it offering you the car for $1.
🎓 Ethan Mollick, Associate Professor at Wharton, used these to highlight that LLMs aren’t ready for sales roles or customer success roles. One of his comments on that post stood out to me most: If you want to work with the LLM, then there is not as much of a problem as there is if there is an adversarial role between customer and seller that is very different.
👍 I completely agree - we can abuse and fool LLMs. That does not mean they aren’t powerful tools that can transform how we work when used to support our efforts, if motivations are aligned.
🤝 One example of this is in software testing. Qualiti has been working with LLMs for 3 years and can attest to their ability to transform how we work. Instead of doing the manual work of writing tests ourselves, the LLM can take the bulk of the load and let us focus on determining what is most valuable for our customers.
🎯 The key is working with models where incentives are aligned - Qualiti wants and expects high-quality results just as much as our customers do.
