top of page

OpenAI’s new o1 model: A Generative AI Powerhouse for the Travel Industry

Writer's picture: George RoukasGeorge Roukas

Updated: Oct 3, 2024

TLDR:


  • On September 12th, OpenAI, the maker of the GPT series of large language models (LLMs) introduced a new generation of models called their "o1 models." The first two o1 models are o1 preview and o1 mini.

  • o1 models are being described as 'reasoning models' because, unlike other LLMs they are not constrained by information they're previously seen in their training data. While they still have to be trained, they also learn the reasoning behind how results are achieved. They can then use that reasoning to solve unique problems they've never seen before. Some are referring to the o1 models as belonging to a new class of large reasoning models, or RLMs.

  • The o1 models are considerably slower and more expensive than traditional LLMs, and while they excel at solving problems where there are answers that can be evaluated for correctness, such as writing code, they are often not as good as some traditional models on problems like writing literature or copy. For these reasons, o1 models are likely to be used in more of a supporting role rather than as a 'daily driver' workhorse.

  • o1 will have significant impacts across all industries, including travel. For example, itinerary planning for longer trips with multiple destinations and multiple people with lots of constraints (activity preferences, dietary requirements, multi-modal transfers, etc.) will take a fraction of the time to generate and maintain compared to what humans or LLMs can do.

  • There are other companies working on LRMs such as Anthropic and DeepMind. We expect to see more appear over time.


Generative AI, or GAI, entered the public consciousness in 2022 with the release of Midjourney’s image model and OpenAI’s ChatGPT chatbot. Since then we’ve seen rapid, but mostly continuous, improvements as new competitors for the large models advanced their capabilities. 


We saw one quantum leap in Generative AI (GAI) functionality with the launch of agents that can: (1) plan a series of tasks to accomplish a goal and then (2) execute those tasks in order and (3) combine their results to complete the goal. Agents first started to surface late in 2023 and we’re now seeing a good number of companies that provide applications to help create custom agents.


Now we have another quantum leap in functionality with the release of OpenAI’s o1 model. (Obviously they’re much better at building models than naming them!) o1 models have a superpower that no other models currently have: the ability to reason through a problem to come up with better answers. Of course, o1 can arrive at much better answers for a range of applications, but certainly not all. In general, where there are well defined answers to a question, like in the areas of math, physics, data analysis, strategic analysis, coding, etc.,o1 can provide better answers than other models. The creation of a well crafted itinerary for a multi-destination trip with multi-generational participants can be complex, but the results can be evaluated for correctness--and that's the kind of problem where 01 shines. Writing an essay on courage is not so easily evaluated. That's the province of a model like Claude 3.5 Sonnet.


The difference in abilities comes from the way the model works. You may have heard that traditional LLMs are essentially prediction engines—upon receiving a prompt they predict the best first word (or token) to begin with and then they predict the next, and the next, until they’ve output their entire response for the prompt. Some call them ‘fancy autocomplete’ engines, which isn’t too far from the truth. These models work in serial fashion, moving from one token to the next. Crucially, they can't go back if they detect an error--they just keep going forward, one token after the next, continuing on a serial path to provide the result. Responses from traditional LLMs (e.g., GPT-4o, Claude, Gemini, LLaMA) are based on similar concepts they have seen in their training data.


The o1 models, in contrast, consider a problem more holistically. They generate multiple starting points and follow multiple paths through the model in parallel. When the results are ready, o1 then reviews each of the final outputs, ranks them from best to worst by looking at how well each path's results satisfy the initial request, and chooses the best result for its response to the prompt. It's less like following a recipe and more like having a team of expert chefs collaborating on the perfect dish.


Just to be clear: the o1 models are not like any other models, and they can’t be ‘emulated’ with other models by using ‘chain of thought’ prompting or any of the other techniques to drive better results from a traditional model. This.is.unprecedented.


A good example of how these two types of models differ is a good, old fashioned crossword puzzle. In one test, Ben Thompson (of Stratechery fame) asked a number of frontier models (including GPT-4o, Claude 3.5 Sonnet, etc.) to solve a NY Times 7x7 crossword. Only the o1 Preview model was able to do so. Looking into why, he surmised that when the traditional models begin working on the puzzle they solve it one word at a time. If they make an error along the way, there is no way to go back and correct it, and everything after the first error has a higher likelihood of being wrong as well. The o1 model, in contrast, can generate multiple guesses for each word, keep track of all the trial solutions, and then go back to choose the best one. Further, o1 isn’t just trained on facts; it’s explicitly trained on the logic used to arrive at those facts. It can use that logic on situations it hasn’t seen in its training data to generalize solutions to various problems


The o1 models (Preview and Mini) are pretty amazing in what they can do compared to traditional models, but don’t rush to change the defaults in your OpenAI subscription—these models do have some downsides.


  • As we said before, they’re not the best everywhere. They are superior at problems where they can generate a lot of possibilities and go back to evaluate the various results and choose a winner. That’s fine for things like code generation where you can test your results, but not so good when you want to write a letter or essay. Hang on to your Anthropic (Claude) subscription for that!


  • The process of generating multiple paths through a model for each prompt means the model is doing substantially more work than with a traditional model that only follows a single path. The result is that o1 models are considerably slower and more expensive than traditional models. While we’re used to getting a response from traditional models in a few seconds, a response from an o1 model may take several minutes to arrive.


So, if it’s slow and expensive and can’t beat Claude at writing essays, what good is it? Generally speaking here are some of the things it’s astonishingly good at:


  • Scientific research and analysis, including data analysis. 


  • Advanced coding and software development. In proficiency tests, it scores better than 90-99% of human engineers, depending on the specific benchmark being used.


  • Strategic planning and analysis. You might have guessed that o1 models are pretty good at brainstorming across many strategic options quickly, and you’d be right. 


  • Processing and analyzing complex documents


That covers some generalized cases, but how about in the travel domain?


  • Complex itinerary planning. For example, planning longer trips across multiple destinations with larger numbers of travelers that consider Individual preferences and requirements, budget constraints, multiple transportation options, and preferred activity schedules. o1 need not be used only for complex itineraries though--even moderately complex plans could benefit from its abilities--but the really complicated ones are likely to see the biggest differences in recommendations.


  • Risk assessment and travel insurance. By processing vast amounts of data on geopolitical events, weather patterns, health risks, and traveler profiles, o1 models can provide nuanced, near real-time risk analyses. (It’s not just for the high priesthood of machine learning experts anymore.)


  • Crisis management and travel disruption handling. This could be in the form of pre-planning in advance of situations you want to be prepared for, or in (sort of) real time when a crisis occurs. The models could rapidly assess the impact of a crisis (e.g., a natural disaster or airline strike,) generate multiple re-accommodation scenarios for affected travelers, and prioritize solutions based on traveler status, urgency, and available resources.


  • Advanced customer segmentation and marketing. By considering a wider range of factors - from past travel history to social media activity - o1 models can create highly nuanced customer profiles.


  • More effective personalization in trip recommendations, interleaving supplier options with actual availability. Personalization through all phases of the travel journey will be a key theme in the next few years, and those who get started early will reap outsized benefits.


And let’s not forget that the uses for strategic planning and analysis, advanced coding projects, and evaluating complex documents above apply to travel as much as any other industry. These are high leverage projects where having a better answer can yield far greater returns to the company.


Call on the o1 models when you have those really complex, high-leverage problems to work through and you need the higher level cognitive horsepower to get the absolute best results. If the saying is true that traditional LLMs are like having a really smart intern to work with, then the o1 models might be like reaching out to a world renowned expert—where you pay by the call but get exceptional advice. 


Forward-thinking travel companies that strategically adopt and integrate these advanced AI capabilities will be well-positioned to lead in an increasingly competitive and technology-driven market. The future of travel is AI-enhanced, and with o1, that future is closer than ever. Are you ready to book your ticket?





George is a senior executive with in-depth experience in product management, technology, and competitive strategy. George was a co-founder of Hudson Crossing, LLC, a management consulting company dedicated to the travel industry, in 2007. Prior to that, he was Group Vice President of Product Management for Travelport, where he led the strategy, development and management for all products facing Galileo’s North American corporate and leisure agency partners. He also participated in the leadership of several early e-commerce companies including Biztravel.com, Room12.com, and Clickradio.com.

 

At the end of 2022, convinced that new generative AI models coming out were going to be unprecedented in their impact on nearly everything we do, George left Hudson Crossing to study AI, specifically in its applicability to business, full time. He now advises companies how to best adopt generative AI through his new company, GAIPAN.

9 views0 comments

Comentários


Os comentários foram desativados.
bottom of page