Just imagine your trip and Airial it! Simply describe your trip to Airial, or drop in your favourite 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 Travel
Thanks for the continued support everyone. To share some learning from my experience I'll post thoughts about two questions which I often get from founders looking to build something new with Gen AI or integrate Gen AI into their existing products. Hopefully it adds some value to everyone here who is building with AI. This comment poses and answers the first question. All civil debate is appreciated, feel free to agree or disagree in the comments.
Will foundational models be able do everything of commercial value in a few years' time?
The generative AI revolution is one of most significant revolutions in all of human history. It is comparable to the advent of electricity or the internet. However, it is quite absurd to think that foundational models will be able to capture all commercial value soon or to think that ChatGPT is the only app you will ever need. Ironically, and with no shade intended, this question nearly always comes from people who don't have much experience working first-hand in AI research or applied AI. Those who have worked in AI have seen first hand just how much there is still to be solved, regardless of whether or not they believe in some notion of AGI emerging in a few years' time.
I'm going to answer this from the perspective of my field of research, Reliable Deep Learning. This is one of the most important fields of AI research today since this is where all deep learning models really struggle. Every commercially engineered system to ever exist consists of several modules each of which has to highly reliable, at least under a set of assumptions. The fundamental challenge of building with Generative AI is that of building with a tool that is in many ways far more powerful than anything before it but is unreliable.
Reliability of Deep Learning models is quite simply unsolved although there has been a lot of incredible research work around mitigating or reducing the unreliability of Neural Networks. There are four fundamental unsolved problems in Reliable Deep Learning research:
Miscalibration: Neural networks are overconfident on their mistakes. As state-of-the-art models have become more accurate, they have actually become quantifiably more overconfident on their mistakes.
Out-of-distribution examples: Neural networks are prone to mistakes, and overconfident ones, when they encounter data which is unlike the distribution of the data from their training dataset.
Adversarial Examples: Neural networks are susceptible to adversarial attacks on carefully crafted examples which are meant to deceive them. Adversarial examples crafted for one model are often transferrable to other models trained on the same dataset, which opens up the possibility of blackbox adversarial attacks.
Lack of explainability: Decisions made by Neural Networks are largely black boxes, unexplainable. This practically makes model trustworthiness even worse, since they are confidently wrong without any justification.
The Generative AI revolution has not addressed any of these. The incredible success of LLMs is about phenomenal engineering achievement, scale, training tricks, some new loss functions and even some smart architectural improvements such as sub-quadratic attention mechanisms, but it hasn't fundamentally changed the transformer architecture which continues to suffer from these unsolved problems.
To conclude, just from the perspective of reliability there is a lot to be solved. All commercial value hinges on reliability.
Airial Travel
Here is the second question, building on the first one from my previous comment.
What will remain to be solved as foundational models improve?
In the context of agentic AI, apart from reliability, there are three other problems that are seminal to building a useful solution.
Contextualization: Getting access to and 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.
In the context of engineering, there is plenty to be solved in terms of optimizing cost, latency and accuracy in each of these steps. No complex task that a human agent does can be achieved by just writing one big prompt with the best model available. Such a solution will be infeasible by just inspecting the cost of the attention mechanism, even ignoring the huge accuracy barrier. The RAG approach to reduce this cost has several limitations of its own within its similarity-based lookup mechanism and does not suffice to fix the need for huge prompts for a single-agent system.
Some other engineering challenges to be solved within multi-agent systems are:
Streaming responses through a system of agents
Parallelizing independent or conditionally independent computation when possible
Explaining and clarifying responses with the user to seek clarification for open-ended choices.
Trading off clarification and turnaround time: How much to clarify before kicking off the main agentic computation?
Interfacing with APIs: How can we integrate with a collection of APIs each of which has their own signature and can return unavailable or unexpected responses which give rise to further challenges to solve.
Banva
Awesome product! Has made trip planning very easy for me. Impressed with the robustness of the product compared to other agentic AI products.
Airial Travel
@rpawar7 Glad you found it useful. Most of our efforts have gone towards making the AI reasoning robust. We have built a proprietary reasoning engine that plans the trip like a human and intricately pieces together all the details for your trip
Airial Travel
@rpawar7 Thanks! Creating an Agentic AI which can handle this level of complexity and then taming the beast to make it reliable has been a significant challenge. but we've learnt a lot in the process. There is still some way to go to increase reliability and keep the reliability high as we add more features.
Airial Travel
@rusa4ek Thank you! Let us know feedback once you try it, either directly or via the feedback mechanism in the product.
I’ve been discussing travel plans with my friends these days, and we will try using this product to plan our trip abroad in May. Good luck to your launch!
Airial Travel
@evakk Good to know it resonates with your needs. We're constantly surprised by how universal this pain point of planning travel is. Nearly everyone we've talked to is fed up with the current travel planning and booking process.
That is a cool use case for AI. What is your monetisation model? Are you afraid of your AI bill?
Airial Travel
@leo_le_roy For monetization, we receive a fee when we refer people to our partners for hotels and flights (Similar to Google Flights). Our AI bill has been very manageable because our reasoning based engine does most of the heavy lifting while LLMs do the inspiration part. Which is why we are able to use smaller AI models reducing cost
Definitely a great tool, no denying that. But how effective is it for Indian travellers, considering the majority of activities fall under the unorganised sector?
Airial Travel
@naveen_sharma_25 Thanks Naveen. You're absolutely right. Airial works everywhere in the world and we use Google places data to display the activities you can do. But in India where activities are unorganized, we often dont have a website to give you for that activity. Unfortunately that is a limitation of the infra. But the way the landscape is changing, we're sure this space will be disrupted in India soon!
@archit_karandikar It feels like a fun and stress-free way to organize the journey. Can’t wait to give it a try!
Airial Travel
@archit_karandikar @supa_l We would love to get your feedback. We put together a highly detailed plan thats completely personalized to you and every single aspect of your trip can be changed either via chat or in the trip controls in the UI. I personally prefer the chat :) It has capabilities beyond anything you'll see in any other AI product