A few days ago, during one of our regular catch-up calls — where travel industry adviser Ellen Keszler and I discuss all things boards, travel and tech — right as we were running out of time, Ellen asked me a simple question.
She asked, nine months into the GPT era, what have we learned?
It is always the simple, thoughtful, direct questions that hit hardest. As someone who watches and studies this closely, I froze and pondered how I could possibly answer this question in the couple of minutes remaining on our call. And while this discussion deserves more time, three things came to mind.
It’s about the product! The model is not enough
One thing we have certainly learned is that artificial intelligence solutions that are effective and delight users require a lot more than a great AI model. I have used and met with roughly 50 generative AI solutions built for the travel and hospitality space and probably just as many doing things in other industries. One of my key learnings is that the major differentiating factor between the ones that seem to be gaining traction and the ones that died on the vine is the quality of the productization.
When studying those startups that are achieving early success across industries, the thing that stands out is the quality of the product and the team’s focus on best-in-class product thinking and product management. In the early months of the availability of generative AI APIs, we saw hundreds (if not more) of startups launch. Startups that took those APIs and added new rules, heuristics and other enhancements and threw those into a basic chat window. And it seems that despite funding and lots of press attention, none of those has achieved material traction or growth. While on the other hand, the startups and large companies that took a bit longer and utilized the APIs as an enhancement to a strong digital product are the ones seeing early success.
Subscribe to our newsletter below
Rather than picking on any startups we can use a few large companies as examples. Think about the difference between the Expedia Group integration and HubSpot. The former simply took a chat window with a basic OpenAI API and connected it to its databases. The results are interesting but do not seem to be any better than just utilizing OpenAI’s public GPT Chat app in one tab and the traditional Expedia OTA homepage in another. This is quite contrary to the HubSpot integration of those same exact APIs. HubSpot’s integration has taken these incredible AI capabilities and put them directly into the core of its product; which has been designed and iterated on for years. Rather than two disparate experiences, the new HubSpot is a super version of itself. All the same great products and features users loved before with a new superpower to make everything more efficient, effective and enjoyable.
Further, unlike Expedia, which is finally using this technology to play in the “plan your actual trip” arena and by doing so redefining its “jobs to be done,” HubSpot has not redefined itself, its value props, nor launched brand new user interfaces but rather has strengthened and supercharged its core value proposition and app.
The point here, it’s not the AI, they both use the same API, it’s the quality of the product that makes the difference.
Further training the model is a must
The only possible moat in a world where tens of thousands of companies are utilizing a small number of nearly identical large language models is a model further trained on proprietary data. It is that simple.
Successful businesses almost always have a point of view. A brand that goes beyond the color pallet and fonts and really focuses on what the business is, what its point of view is, how it goes about providing value, etc. Without further training your model, it is nearly impossible to tune the quality, style or brand fit of the AI’s outputs to fit your business and customers. Further, it is this additional training (and productization) that are the moats and differentiators that will lead a customer to reliably choose your business over the competition.
Model decay is real
If we did not know this before, operating an AI product is very far away from the “set it and forget it” of our dreams. In just the short time since generative models were released into the wild, we have seen the very real nature of model decay. The fact is that over time AI model output quality seems to deteriorate. And sometimes reinforced learning seems to make the problem worse not better. This is counterintuitive.
Further, we have learned that AI models trained on content created by other AIs seem to have inferior results and lead to more rapid decay. There is probably a lot we can learn from that.
The good news for industries including travel and hospitality, retail, entertainment, sports, real estate and many others is that these industries have everything it takes to fight back against decay and staleness. That is, they have real people engaging other real people, in the real world, in real time, 24/7. It is this advantage, ubiquitous availability of real data and content, that is one of the elements that has me so excited about generative AI in our space.
There is so much more to say, including the fact that high-quality training data has always been critical for AI and that first-party data is the best data of all. And of course, that generative AI is just one tiny subset of deep learning, and that deep learning is just one subfield in machine learning, which of course is a subfield of AI overall. My point: There is so much more to be excited about in the world of AI outside of the generative realm we are all currently captivated by.
