Generative AI in the Travel Industry – The biggest updates since launch
All year, in the world of GenAI we’ve seen the fastest pace of product releases in the history of tech. Last week saw the biggest day yet, as OpenAI released a variety of new products and upgrades in their Developer Day in San Francisco.
The Simple Stuff
Context window. That’s how many tokens or words that you can use in a prompt and response. The new limit is about 100k words (128k tokens). That’s a book. So now in one prompt, you could ask it to summarize an entire uploaded book, or alternatively, you could ask it to generate an entire book with a short prompt. Or anything in between.
Updated knowledge. ChatGPT now has knowledge up to April 2023. Still not fully up to date, but in another update, ChatGPT is able to make the decision for you if it should turn on the browsing feature (search online) for new information, based on your prompt. I’m not convinced that’s going to completely work. I don’t like the hybrid of browsing and general training knowledge, but I think the lines between those two options are gradually getting blurred.
GPT4-Turbo is the new version of GPT4. Faster. Smarter. 3x cheaper. Open AI already slashed the price of Version 3.5 by 90% at the start of the year. It looks like they’ll continue to reduce prices as their costs (per word) come down.
The Good Stuff
There were quite a few technical updates around things like formats. Being able to (more reliably) return valid JSON as an example. Some of those updates look quite valuable, but they do get a bit nerdy to cover here.
The most impactful update was in the Agent space. Two concepts called an Assistant and a GPT. They’re quite similar products, but fundamentally, they allow you to create an Agent, from within ChatGPT. This was technically possible already, with some good coding, but none of the agents created by 3rd parties (as layers on top of GPT4) have been great up to now. OpenAI might well have solved that, as they’ve taken on the complex part of the task, and brought it directly into their platform where anybody can use it. It’s kind of the next stage for plugIns – which came out in the spring. You can also look at it like ChatGPT finally has a long term memory.
Say you’re a travel agent specializing in travel to Bali, you can now create a ‘GPT’ called ‘Bali Travel Agent’. It’s going to walk you through setting this agent up. You can decide what kind of tone you want to use, generate guidelines for the kinds of tasks you want to help with, and you can upload information to store within your GPT. Maybe you have 1000 blog posts written over the last 10 years, full of really specialized local knowledge (most likely these 1000 blog posts are already obsolete and ChatGPT knowledge is already superior, but people are very defensive about the value of their content, and this is just an example). Maybe you also have a few hundred expert itineraries that you’ve created for previous clients. You upload all of those too. Basically, all the docs you’ve ever used to run the company. Copy & paste all of it into a doc, and upload it.
Thats it. you’ve created a custom travel agent. You can share that with anybody you like, and they can question it just like they question ChatGPT. You could create an itinerary builder for your guests, and f you give it access to pricing, ask it to generate quotes. You could create a voice-to-voice chatbot, generate audio tours of points of interests on your itineraries, create images and a selection of top tips – entirely customized for the client interacting with your GPT. All within this one product.
If you want to get smarter than that, you can set it up with access to APIs – where it can grab real time flight info, hotel rates & availability, you name it. It’s the kind of thing that some of the largest travel companies built in the last 9 months, spending many millions of dollars, and now anybody can do it. Mostly without writing a line of code. Oh, and if you need code to complete some more complex actions, it can generate that for you. Automatically.
These GPTs will be in an ‘App’ store very soon – so you’ll be able to market your smart GPT to customers directly. OpenAI will take a cut. Not sure how that will work yet.
In the last 12 months I’ve openly shared everything I’ve learned about GenAI. At Magpie, we have around 15-20 products & features using many of these tools, and I’ve been very transparent about what we’re building and how we’ve done it.
This new advancement is different. It’s too powerful. You can run at 10x the speed you could even a month ago. And a month ago was 10x the speed of a year ago. What might have taken a large OTA 6 months with dozens of developers, you can now do mostly in a few days if you know what you’re doing.
So I’ll leave the dreaming up of innovative ideas to you. Once you get your head around what can be done, there’s potential like never before.
Before you get too excited building all of these cool agents though, remember that the next update in a few months will be 10x better again. And don’t forget that whatever you can build today, anybody else can also build.