Reliable Deep Learning Researcher from Google Brain launching Airial Travel on Friday - AMA
I'm Archit Karandikar, the co-founder of Airial Travel by CoInvent AI. Airial is your personal travel agent with AI superpowers. Simply describe your trip to Airial, or drop in your favorite TikToks / IG and see your dream trip come to life instantly. Airial's one-of-a-kind AI picks out every last detail so that you can start packing, stress-free! Airial launches on Friday, March 28 (launch page).
Before this I used to work on large driving models at Waymo and reliable and explainable deep learning research at Google Research. My co-founder Sanjeev @sanjeev_shenoy1 was one of the founding engineers of Instagram Reels. One of the reasons we chose to build an AI travel agent was that travel is being disrupted by both Generative AI and Social Media. The other is that we're both avid travellers.
Reliability is one of the most critical challenges for building with Gen AI today. In fact, the fundamental challenge of building Gen AI native products today is that of building with an ultra-powerful tool which is unreliable and unexplainable.
We will be online on March 27 answering questions. Please feel free to post questions beforehand. AMA anything about the following or beyond
How to build Gen AI native applications to be reliable and explainable?
The challenges of building Agentic AI
The future of travel with AI
The intersection of social media and Gen AI
Let's connect!
Archit: https://www.linkedin.com/in/architkarandikar/, https://x.com/KarchitK
Sanjeev: https://www.linkedin.com/in/sanjeev-shenoy-68333253, https://x.com/SanjeevShenoy
Replies
Really impressed by how Airial takes a real-world behavior—saving travel content on TikTok or IG—and turns it into something instantly actionable.
@Archit curious to hear: What’s been the hardest part about making Airial’s itineraries actually reliable and useful across different kinds of trips? Agentic AI sounds powerful, but I imagine it’s a beast to tame when it comes to real-world constraints like flight pricing, hotel availability, and ground transit schedules.
@Sanjeev – You’ve worked on social products at massive scale. What was most different about building Airial from scratch? Were there any early product assumptions you had to completely throw out?
@Both - How much of the challenge in building was teaching the AI to “read” chaotic or nonlinear social content (like TikToks with multiple spots or edits)? Was there a breakthrough moment in making that feel smooth and reliable?
@jo_adams44 Lots to unpack in your question. What differentiates Agentic AI from other Generative AI is the ability to take actions on behalf of the user in a complex environment. There are three fundamental challenges to achieve this:
Contextualization: Understanding information from the environment in the context of the user query.
Reasoning: Thinking through the user query to understand the actions that need to be taken.
Agency: The final step of creating the action in a format which can be plugged into the system.
For travel there are separate challenges for each of these. The product must be designed so that the AI engine has access to the environmental context (trip plan) and the ability to take a very wide range of agentic actions to first create and then modify the trip plan. There is the core challenge of AI itself - it reasons through thousands of variables that make up a trip plan and takes precisely the actions that the user intends. There are several dependencies and constraints which must be taken care of once any proposal or edit is made. Finally, the AI must interface with APIs in the creation of its plan since it cannot dream up flights, hotels, activity details and transportation on it's own. We've also built the ability to plan with TikToks or IG reels which needs high-fidelity video understanding in-context. There are several levels of engineering challenges to optimize speed, memory and cost within each of these.
This is the crux of the challenge of building high-fidelity agentic AI than can create and modify a highly multivariate structured object (in this case, travel plan) as per user directions. There is a lot to be solved here. Agentic AI will revolutionize many industries in the years to come.
Additionally, What is the distinction between generative AI and agentic AI in your mind and how good is GPT-4 at real world travel queries?
@jo_adams44 As answered above, what differentiates Agentic AI from other Generative AI is the ability to take actions on behalf of the user in a complex environment.
There is a paper called TravelPlanner by Meta AI which precisely answers your question, and quite emphatically at that. As per this paper, GPT-4 fails on 99.4% of real world travel queries, which goes back to my point of above there being a lot still to be solved here. To see why, take any travel TikTok (example) and try planning that trip to the last detail with AI - most solutions will be completely insufficient. Airial is the only one which does a fairly good job. In fact, Airial is the only one which even attempts to plan something like this including all the logistics and even for us, there is lot still to be done to capture all the details and increase both precision and recall.
What more can generative AI do for travel, along with providing itinerary inspiration?
@chirag_agarwal12 In terms of functionality gen AI can provide discovery, logistics, edits, refinement, price optimization, booking in addition to inspiration. In terms of the value props, gen AI can provide time-saving, convenience, money-saving, stress-free planning and even help with travel safety since it plans in accordance with expert advice. Once travel planning with Gen AI becomes the norm, the old way of planning and booking will feel outdated, inconvenient, suboptimal and will be completely replaced, in our understanding.