ChatGPT's life - saving skills prove more reliable than maps after a five - hour wilderness ordeal in the Canadian jungle.

Imagine that you're lost in a Canadian forest for 5 hours, your phone battery is down to 3%, Google Maps malfunctions, and the signal is weak. But ChatGPT came to the rescue with real - time coordinates. It's a textbook example of AI navigation. Come and take a look!

Recently, a post on the X platform went viral. A group of people were riding all - terrain vehicles (ATVs) in the remote Canadian town of Mabou. They got lost for a full five hours and finally made it back safely with the help of ChatGPT navigation.

Mabou is a small place surrounded by undeveloped forests and small paths.

They originally planned to start from Upper Southwest Mabou Rd and ride 18 kilometers to Whycocomagh for a visit.

The plan was good, but they accidentally deviated from the main road and got into a small path not marked on the map.

As a result, Google Maps and ATV - specific apps were useless because these tools only recognize main roads and don't cover those hidden paths at all, leaving them completely lost.

The phone signal was also poor, and the battery was down to 3%. One of the riders had a brilliant idea and tried to use ChatGPT for navigation, sending GPS coordinates to ChatGPT every 5 - 10 minutes.

Surprisingly, this method really saved them.

ChatGPT's Step - by - Step Rescue

In the first screenshot, ChatGPT gave a clear and practical response, breaking the route into several simple steps.

Send Coordinates for Help

They sent their real - time location (45.9697°N, 61.4119°W) to ChatGPT using their phone's GPS and said they wanted to go to Whycocomagh (45.96435°N, 61.1426°W), asking if there was a passable route.

ChatGPT's Great Response

ChatGPT analyzed the coordinates and terrain and gave extremely clear step - by - step instructions:

Step 1: Head east on Upper Southwest Mabou Rd. It's a dirt road that ATVs can pass through.

Step 2: Merge onto Chestico Trail / Celtic Shores Coastal Trail. This trail starts from the east and is legally available for ATV use. It's about 17 kilometers long, stretching from Fort Hood to the Mabou River.

Step 3: Follow the trail eastward, parallel to Highway 19. Go through the forest and follow the Mabou River.

Step 4: When you're approaching the village of Mabou, turn south or east to connect with Route 252 (the road leading to Whycocomagh), and you're there!

These instructions not only include road names but also tell you the directions (east, south) and the terrain (forest, river). They're very practical and easy for lost people to understand.

Adjust the Route at Any Time

They sent new coordinates every 5 - 10 minutes, and ChatGPT adjusted its suggestions based on the latest location.

The following picture shows that ChatGPT confirmed they were still on the right track, suggested they continue along the Mabou River, and also gave the next - step options.

Outperforming Traditional Navigation

Google Maps and ATV apps failed completely because they lacked data on small paths.

ChatGPT, on the other hand, used GPS coordinates and satellite views to figure out how the small paths were connected and gave instructions in plain language. It was a real lifesaver, especially in the wild.

Some people praised ChatGPT as a "lifesaving tool," but others complained: "With a phone signal and GPS, why not just look at the satellite map?"

The replies to the post explained that although satellite maps can be viewed, they can't provide real - time personalized text instructions like ChatGPT, especially in complex terrains.

Research has proven that large language models (LLMs) like ChatGPT have real potential in outdoor navigation.

For example, the PathGPT framework, which turns historical routes into text and then uses AI to generate personalized routes, works really well.

PathGPT: Asking a Friend for Directions

Imagine opening a navigation app and saying, "Find me a route from the office to home that avoids traffic and allows me to pick up a cup of coffee on the way." The app immediately gives you a precise plan.

Recently, a research team from Shanghai Jiao Tong University launched PathGPT, which has completely refreshed our understanding of navigation with LLMs.

Paper link: https://arxiv.org/abs/2504.05846

Previous navigation algorithms, such as Dijkstra's shortest - path algorithm, are like rigid robots: the shortest distance is the only thing that matters, and they don't care about anything else.

But in reality, driving or walking isn't that simple.

Some people like to take detours to enjoy the scenery, some are in a hurry and want to avoid traffic jams, and others want to pick up their kids on the way. Traditional algorithms can't handle these complex requirements at all.

Later, there were machine - learning models that could learn patterns from historical trajectory data, such as recommending routes based on traffic conditions.

But they also have a major drawback: once a model is trained, it can only work according to fixed rules.

Want to add a requirement like "pass through a certain business district"? You have to retrain a new model, which is time - consuming, labor - intensive, and costly.

The core idea of PathGPT is: since large models (like GPT) can understand natural language, why not let them translate users' requirements?

For example, when a user says "the fastest route" or "a scenic route," the large model can understand and generate a plan by combining historical route data.

But LLMs also have limitations: they may not know the specific road conditions in a certain city and may even have hallucinations.

So PathGPT has an "add - on": Retrieval - Augmented Generation (RAG) technology.

Simply put, first build a knowledge base of historical routes: convert the starting point, ending point, and names of the roads passed through for each route into natural - language descriptions. For example, from People's Square to the Bund, passing through Nanjing Road and Zhongshan East 1st Road.

When a user asks a question, PathGPT first finds historical routes in the knowledge base that are similar to the user's starting and ending points.

Actual Test Results: High Flexibility with Room for Improvement

The research team conducted experiments on taxi trajectory data in cities such as Beijing, Chengdu, and Harbin.

The results showed that when generating "the fastest route," although the precision and recall rates of PathGPT were slightly lower than some traditional machine - learning models (for example, on the Harbin dataset, the precision rate for the fastest route was 48.4%), it has a killer feature - it can handle those strange requirements that haven't been pre - trained.

For example, if a user asks for a route passing through three parks, traditional models may be completely confused, but PathGPT can combine natural language and historical data to give a reasonable plan.

The greatest significance of PathGPT is to turn navigation into a natural conversation. In the future, you can just talk to the navigation like asking a friend for directions.

Of course, it still has room for improvement. For example, it may give wrong directions occasionally (the hallucination problem of LLMs), and its reliability in complex traffic conditions needs to be improved.

But think about it. The large models have only been around for a short time, and they've already achieved this much. In the future, if combined with more accurate real - time traffic data and a more complete knowledge base, they may really become everyone's intelligent travel butler.

ChatGPT can not only chat but also give practical suggestions based on real - time data, which is especially useful when traditional maps fail.

AI is changing the way of outdoor exploration. Navigating in a map - less forest using coordinates and logic is definitely a milestone.

Reference Materials

https://x.com/rohanpaul_ai/status/1937199835318485177

https://arxiv.org/abs/2504.05846

This article is from the WeChat official account "New Intelligence Yuan". Author: Ying Zhi. Republished by 36Kr with permission.