Planning summer travel is tough. Planning it for a busy family—visiting a number of locations with three little children in tow—can really feel downright not possible.
I’m knowledgeable information and journey photographer, so I must plan a variety of journeys with my spouse and our three boys. I’m additionally an AI skilled with a cog sci diploma from Johns Hopkins and a decade of expertise within the discipline.
So naturally, I made a decision to mix these two passions. To that finish, I constructed an AI digital twin of my household, which makes use of a number of rounds of simulation and superior agentic AI to plan each side of our summer time journeys.
I’ve examined it in planning a number of real-world journeys, and its concepts are improbable—with some large caveats.
Vibing with Claude
To construct my journey planning AI, I turned to Anthropic’s Claude. I’ve written Python code for years, so I understand how to code up a fundamental script and combine with an utilized programming interface (API).
In my expertise, although, Claude is much better at these items than I’m. Once I first began testing Claude and its ilk a number of years in the past, the Python they wrote was buggy and kludge-filled. By the point I completed debugging, I usually felt it will be simpler to simply write the code myself.
Not anymore. With Claude’s new Fable 5 large language model, you’ll be able to describe a chunk of software program you’d prefer to construct in intense element, and the mannequin will spend 20 minutes or extra understanding your temporary and spinning up completely optimized Python code that works proper out of the field.
Folks name the method “vibe coding,” however that’s at all times felt a bit pejorative to me. Working with immediately’s greatest coding LLMs feels extra like having a mid-level software program engineer in your pocket than utilizing a easy mannequin to spin up a cutesy internet interface or yet one more Pac-Man clone.
For my AI journey planner, I advised Claude that I wished to explain every member of my household intimately, constructing a journey profile for every of us. I then wished at hand the system particulars like our journey dates, lodging plans, and journey preferences, and have it analysis concepts for our journey and in addition lookup fundamentals like climate and closures.
Armed with these particulars, the system would simulate our journey a number of occasions, predicting how every of us would react to every journey thought. Lastly, the system would put together an in depth report and itinerary based mostly on its findings.
I wished the top consequence to be simple to make use of on the go, and straightforward to change. So I requested Claude to construct the system within the type of a Google Colab pocket book. Colab is a free service that permits you to run complex Python scripts in Google’s cloud for free, with a Google-Docs-like web interface.
Claude dutifully set to work. A number of minutes later, I had over 1,000 strains of Python code able to go.
The ultimate system makes use of the OpenAI API to ship out a number of AI brokers, researching 40-plus eating places, actions, and basic sights for any spot we’re planning to go to. It pulls in climate information from Open-Meteo, runs 10 rounds of simulations utilizing the digital twin of every member of the family, after which writes up a five-plus web page itinerary and journey plan.
To construct our profiles, I turned to ChatGPT. I usually chat about journey plans with the bot, and it now has an extensive memory function. I requested it to recall all it had realized about every of us from these earlier journey conversations, and to construct a profile of our likes and dislikes in a machine-readable JavaScript Object Notation (JSON) format.
Many of those had been spot-on. ChatGPT accurately decided that my oldest son loves Legos and mini golf, my youngest has a deep ardour for ice cream, and my 6-year-old loves—a bit inexplicably—bird-watching.
I took these fundamental profiles, added my very own insights, after which loaded them into Claude’s framework. After a day of tinkering, my AI digital-twin journey planner was prepared to make use of.
Boots on the bottom
I’ve used my agentic digital-twin system to plan at the least 5 journeys this summer time, from fast day journeys to weekend journeys and even weeklong getaways.
Utilizing the system is simple—I plug in journey particulars, fireplace it up, and wait about 10 minutes for a report to come back again.
Like many agentic AI techniques, my journey planner burns by way of an ungodly variety of tokens—the foreign money of AI compute time—to supply its outcomes. On a latest run, I deliberate a single day-trip to Capitola, California. Doing so burned by way of a cool 256,503 tokens, costing me just a little underneath $5. In June alone, I used over 8 million tokens for journey planning.
The outcomes are price it. My system’s stories are informative and very detailed. Every report leads with a summarized, ranked desk, displaying the outcomes of its simulations and describing the actions and eating places we’re prone to take pleasure in essentially the most.
For that journey, the system discovered an area ice cream place (Polar Bear) it thought we’d love, and in addition suggested us: “If you’re not already totally full of meals, cease at Gayle’s Bakery & Rosticceria earlier than committing to seashore parking.” The ice cream spot was certainly improbable, and Gayle’s was probably the greatest picnic provision locations we’ve ever visited.
My system additionally picked the most effective seashore for our household, and even the most effective car parking zone—no small factor in car-obsessed, space-crunched California!
The system does an amazing job of balancing competing wants. It is aware of that I’m a professional photographer and might want to go to and {photograph} fascinating sights for my work.
Nevertheless it additionally is aware of that three little children can deal with solely a lot of this, and builds in quiet actions—and loads of ice cream—to maintain the journey in steadiness.
And it retains issues actual. My pure tendency is to go to and {photograph} a lot of locations. However in planning an even bigger summer time journey to Hawaii, the system jogged my memory that our youngsters would need to spend most of their time floating across the lodge pool.
Maybe a bit passive-aggressively, it ranked varied features of the pool (a water slide, a themed cave, and the like) as our prime “should do” actions, whereas shoving bold issues I’d take pleasure in (like a farm tour on the opposite facet of the island) a lot additional down the checklist.
Level taken, AI!
The system is sweet, however it isn’t excellent. When you’ve bought precise boots on the bottom, issues with its suggestions generally shortly emerge. On a latest fruit-picking journey, for instance, it suggested us to go to a farm with absolutely the best-tasting peaches.
The peaches turned out to be improbable. However choosing them additionally required slogging round an uncovered, sunny discipline and climbing a 7-foot agricultural ladder at every tree. The children struggled. We left after a couple of minutes and visited a distinct farm with underripe fruit however quick access and brief timber. My children liked it.
The system, in different phrases, is sweet at optimizing however doesn’t at all times optimize for the suitable factor—a common problem for artificial intelligence in general, and agentic AI in particular.
The coder in my pocket
I realized loads by constructing my AI digital-twin journey planner.
It confirmed me firsthand that AI brokers—with their means to learn by way of 1000’s of internet pages, Reddit posts, Tripadvisor evaluations, and the like—are improbable at unearthing surprising data and the sorts of “hidden gems” that journey professionals always spill ink over.
Googling alone, I’m prone to discover distinguished sights and vacationer stops in a goal vacation spot, however I’d most likely miss small native spots that won’t have a elegant web site (or a marketing price range). My system’s brokers discover these and spotlight them—albeit after expensively crunching by way of a lot of content material!
I additionally realized that feeding AI fashions reams of nice background information issues. If I ask Claude out of the field to seek out “actions in Costa Mesa for a household with three children,” it’s prone to suggest pretty generic, kid-friendly spots.
Asking it to seek out actions for a 9-year-old with an adventurous palate who doesn’t thoughts spicy meals, a bird-focused 6-year-old, and a 5-year-old who loves numbers and domestically made sweet yields significantly better, extra tailor-made outcomes.
My largest takeaway from constructing the system, although, is a realization of how highly effective immediately’s AI coding instruments have turn out to be. In a pre-AI world, I’d by no means make investments tens of hours to code up a digital-twin-based simulator with full internet crawling and tie-ins to dwell meteorological information with a purpose to plan just a few summer time journeys.
Claude, although, can construct one thing like that in minutes. That doesn’t solely make my job as an beginner coder simpler—it opens up tasks I’d by no means dream of tackling with out AI’s assist.
Though I’ve discovered my system genuinely helpful in my photographic profession, I principally constructed it for enjoyable. However in doing so, I can now simply see how a real software program engineer—armed with even higher LLM-powered instruments than Claude’s fundamental coding system—may construct astonishing issues very quickly in any respect.
Whereas chatbots are more and more succesful writers and communicators, their true potential lies of their means to talk the language of our time: pc code.
My journey planner is nice at discovering native ice cream spots and sandwich joints. However I can see how an analogous agentic system—constructed by a real coding professional with entry to cutting-edge instruments and information—may remedy intractable issues in science, drugs, or finance with relative ease.
LLMs get a lot of reward and opprobrium for his or her prose and their recommendation. Nevertheless it’s their coding prowess that may actually change (or break) the world.
