Bob the Bot
One of my favorite parts of being an engineer at a startup is building new features entirely from scratch.
The most recent such feature is “Bob the Bot”. Bob is a GPT wrapper chatbot -- but he doesn’t just chat. He has access to our database and can take real actions for customers (home service pros), all through text message. In just over a month, Bob has exchanged more than 20,000 texts with our pros. He's posted to social media, created project showcases, requested reviews, and updated bookings -- many hundreds of times. Pretty wild for a feature spun up in a few weeks and released just over a month ago.
Bob works remarkably well most of the time... some of our customers have even been texting certain things that suggest they think Bob is a real person! But as one person on the team put it: Bob is a middle schooler and we need to get him to college as fast as we can. One challenge is that he often he mistakes "intent" - e.g., sometimes he thinks customers want to create social posts when really they want to create project showcases. When this happens, he'll enter the "social flow" and may start to take actions the user doesn't want.
I was chatting with a friend a couple weeks ago about how we might improve this, and he suggested something clever I hadn’t thought of: make calls to 5 different LLMs to determine user intent, then choose the “consensus”. If 3 out of 5 return “social” --> enter the social flow. This got me thinking that another thing we could try is making 5 calls to GPT and returning the consensus.
These aren't earth-shattering ideas - but they are clever. They illustrate how much impact a little cleverness can have in a new frontier where the ground is shifting literally daily.
I believe we're barely scratching the tip of the iceberg of products that are possible with LLMs, because we haven't yet figured out how to harness them to their full potential. As Brian Chesky tweeted last night, "application development is behind the models." By all means, bring on GPT-5. But there is so much left to do and build and discover with the models we already have. It’s an extraordinary time to be building software.