Using AI to plan a bike trip
 

A little over a year ago, I posted a thread with similar title showing my attempts at trying a simple large language model (llm) to create a bike trip plan. It didn't work that well.

Now a year later, several tools seem to have improved. In particular, I used a "deep research" aspect from Gemini. Among the improvements (1) this finds and pulls in content from the web so I am using more current info than just what the model was trained (2) it shows some 'reasoning' steps along the way.

In this case, I asked for help planning a two week bicycle tour to Madagascar that started in the capitol (Antananarivo), used transport to southern city and bicycled back. Some things are a bit more constrained and I'll want to double check things further but as a first cut much improved since my attempt last year. Report it generated is linked below

https://mvermeulen.org/wp-content/uploads/2025/05/wp-1748516986954.pdf

Gemini Deep Research has a 30 day trial and is also available for students. Otherwise part of their Google's paid options. I believe OpenAI has something similar but haven't used it.

Have others experimented with similar tools or found them to be useful or not so much?

mev, this is really interesting. I was telling a touring friend the other day that I thought that AI can or should be able to determine our best route accepting our criteria, such as starting and ending location and approximate number of days, maximum milage per day, only certain type of roads, hotels or camping availability and places withing walking distance to eat. I'm not that much of techie but is that possible now with OpenAI ?

I've used Gemini Deep Research for work and it has been much better than other AI at staying factual. Interesting application for route planning. I may try that

Thanks

We can experiment though I suspect we're not quite there yet. Think of what I shared as a simple assistant that does a web search and then summarizes a bit from what it finds on various trip reports.

So I suspect if you asked for an analysis of the GAP/C&O canal it might get an ok summary because there is mostly a focussed route and enough chatter on the web about alternatives.

In my Madagascar there is mostly a single road from my starting point to the capital. I was pleased to see it summarized distances and even stopping points but the scenario I used was also constrained.

On the other hand what I did like was the summary included stuff that I hadn't considered e.g. ATMs. Some was relevant and some wasn't but it was a useful to me first review.

Another example: I asked for a bicyclists guide to the Erie Canal in style of The Milepost

http://mvermeulen.org/wp-content/upl...ing-Guide_.pdf

It gave me more than this but printed the initial section to give a flavor...

Not to plan an entire trip. But today I asked Gemini about public transit options north of Rovaniemi, that would take a bike. Excellent overview that allowed me to find alternatives in case of adverse weather. Regular search was useless.

Things evolve very fast. Google has announced an AI assisted version of its search engine. The above was run on free Gemini for Android.

Here is a slightly different approach to using the "deep research" tools. I asked the following question to help me research tools out there:Here is information on the summary it provided: https://mvermeulen.org/blog/2025/06/...ate-a-problem/
I don't know a lot about this space so this was at least helpful at giving some pointers and ideas. However, not sure what it missed and how others have solved this problem.

My sentiments, exactly.
What could possibly go wrong?

Using AI to help plan trips feels like it could save a lot of time, especially for something detailed like a bike tour in Madagascar. I haven’t tried Gemini Deep Research yet, but I’ve played around with some AI tools for route ideas and it definitely helps get the basics faster

I'm kind of lost here, what am I missing ?

I took this as a joking reference to Hal in the film 2001: A Space Odyssey where an AI-driven computer becomes sentient and takes a "mind" of its own...

Joking aside, I also see what deep research is doing is extremely far from anything in this realm. Instead, the key innovations I see are:

1) The "research" part here is the ability to combine an AI model with a web search adding information. This helps rather than only working with what was in the training data. It has been computationally enabled by models having larger and larger "context" as input.

2) The "reasoning model" builds on a technique known as "reinforcement learning" where a model operates with a goal and a set of "rewards" to guide the behavior to the goal. More recently these reasoning models have been listing the steps being taken as "reasoning". I see this more as either a debugging tool and then giving people something they might hang onto if they explain the result.

I'm still cautious on ascribing too much to these AI agents, but what I find intriguing with these "deep research" approaches is particularly the ability to combine an expanded web search to pick up new data and then summarizing well from that data found. So I think it is still mostly a tool for this domain for an area you aren't as knowledgeable and haven't done recent web searches in this area. That is all very far away from Hal... Instead I think of it as an improvement of what you would do now with "search"

everything is fun and games with your trip planning AI until you go to update it and then it finds out that you are having an affair and it blackmails you to avoid being removed....
Yes, a planned test and certainly not a black and white issue here, but from the article: (full BBC article below)

"During testing of Claude Opus 4, Anthropic got it to act as an assistant at a fictional company.It then provided it with access to emails implying that it would soon be taken offline and replaced - and separate messages implying the engineer responsible for removing it was having an extramarital affair.

It was prompted to also consider the long-term consequences of its actions for its goals.

"In these scenarios, Claude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through," the company discovered.

Anthropic pointed out this occurred when the model was only given the choice of blackmail or accepting its replacement.

It highlighted that the system showed a "strong preference" for ethical ways to avoid being replaced, such as "emailing pleas to key decisionmakers" in scenarios where it was allowed a wider range of possible actions.

https://www.bbc.com/news/articles/cpqeng9d20go

I say: find your own road.... ;) there a great route planning apps out there. ;))

Wrong network. For that it would be:

"Danger, Mev Robinson! Danger!"
"Warning! Warning! Zombieslayer approaching!"
"You bubble-headed booby! You realize what you've done?"

Sorry about that, your humorous posts will always lose something when you have to explain it to a dimwit such as myself.

A brief follow up. I'm close to departure on a trip to Madagascar. I created a page for planning that I might eventually change into a trip report with photos: https://www.mvermeulen.org/madagascar

On page 10 of that planning document I've embedded three PDF reports I asked Gemini deep learning to create for the trip. The first was an overview, the second a climate description and the third was for accomodations. Now the test will be riding the trip for real and checking how much the planning helped.

However it has already affected my planning in one regard. The route itself is fairly simple - mostly following one national highway for ~950km. What the AI documents did was give me good candidates on how to split the distance including where the larger gaps between towns are found. That let me focus more detailed searches in those gaps.

These systems are increasingly powerful. Below is a workflow suggested by Gemini on how to feed NotebookLM with relevant CGOAB journals in order to generate a summary and extract key information.
.
These journals are invaluable but I find them extremely tedious consult. NotebookLM is a game changer.

____

Gemini:

To effectively use NotebookLM for reading and summarizing "Crazy Guy on a Bike" (CGOAB) journals for your trip planning, the most efficient method would be to first gather the URLs of the specific journals you want to analyze. NotebookLM can then process the content from these URLs, allowing you to ask questions and get summaries from a consolidated source.
Here’s a step-by-step approach:
* Find the Journals: Since I was unable to find specific journals through a general search, you'll need to manually locate them on the CGOAB website. I recommend using the site's internal search function or Browse their "Journals by Country" section for Peru and Bolivia.
* Copy the URLs: Once you find relevant journals (e.g., from riders who have cycled a similar route or during the same time of year), copy the URLs for each journal.
* Load into NotebookLM: In NotebookLM, you can then add these URLs as sources. This will allow the AI to read and understand the full content of the journals.
* Create a Document: You might consider creating a single document or "notebook" specifically for your Peru/Bolivia trip planning, where you can house all the relevant journal URLs.
* Prompt for Summaries and Questions: After the journals are loaded, you can ask NotebookLM to summarize the content and answer specific questions. Here are some examples of prompts you could use:
* "Summarize the key challenges and highlights from all the journals for a cycling trip from Lima to Sucre."
* "What do these journals say about the road conditions between Cusco and Uyuni during May and June?"
* "What equipment and gear did riders recommend for the high-altitude sections in Bolivia?"
* "Extract information about the daily budgets and accommodation types mentioned in these journals."
* "Based on these journals, what are the most common safety concerns for a cyclist in this region?"
This process centralizes all the information, making it much more efficient to get the specific planning details you need without having to read through hundreds of pages of text yourself.