Greetings from Goa, India and the second XBE Jugaad!
I’m thrilled to report that the second XBE Jugaad, a feature-oriented hackathon, was a huge success. Most of the product development team was able to attend in-person, which is a rare occasion given that XBE is remote first and distributed around the world.
At the end of day one, we brainstormed two dozen AI-powered features that we thought would add significant value to XBE users. We focused on AI-powered features (leveraging GPT-3) so that we would have more confidence as a group in our ability to leverage AI models (not just GPT-3) throughout the XBE platform.
During the night before day one of the Jugaad, I added a small release note headline suggestion feature as a proof of concept, and to learn some lessons that could be applied by the group. Before choosing a feature, we reviewed that work and a bit about the process required to build it.
The group chose one feature area, safety risk management, to enhance through AI during day two. Our goal was to ship a complete feature based on development built between sunrise and sunset. The business objective of the feature was to integrate safety risk assessment into the job production planning process, and to improve safety risk management communication during construction.
We were able to leverage AI models in three areas related to this feature.
- Automatic generation of headlines for incidents.
- Safety risk suggestions based on job characteristics and recent incidents across the XBE network.
- Automatic safety meeting communication based on the job's key risks and other characteristics.
Ten of us worked all day and shipped a production-ready feature by sunset, in time for an amazing wrap party at HOSA in Siolim. We split the work into three parts.
- Server-side plumbing
- Client-side user experience
- AI prompt design
Leveraging AI made the server and client development work more simple even though the feature is much more capable than what we would have created without AI.
Six lessons that we learned about developing AI-powered features include:
- Prompt design is an art form that requires significant creativity.
- It’s hard to overestimate what GPT-3 is capable of.
- AI shouldn’t be thought of as a product feature, but as a tool used to enhance all product features.
- So much work is pushed to the AI-model that it’s possible to ship a feature early and just refine the prompt over time without anything about the interface changing.
- The unpredictability of responses from the AI models is both magical and a bit unnerving, but the only way out is through.
- It’ll take us many years to full grasp and leverage the possibilities that exist today.
It’s delightful to ship features that integrate structured data and curt user generated content into well written output. This is an underrated implication that users will love. In my experience, many people are very anxious about writing things in public, and when there’s an AI intermediary between their thoughts and the distribution of those thoughts, they’ll be more willing to share.
Given the current tools, one programmer could ship multiple valuable AI-powered enhancements to existing features every single day for quite a while before we’d run out of ideas. The productivity is off the charts, but so is the opportunity. Within 2 years, features that aren’t enhanced with AI will start to feel a bit broken. We’re hustling!
We’re grateful for your support, and we’re always looking for your feedback. Reach out to me or any member of the team if you have ideas or questions.
Founder & CEO, XBE
PS. This newsletter was also written by GPT-3. I wrote the key points as bullets, and it rewrote the bullets into a newsletter form. It also generated the headline! It's AI all the way down.