The use of AI and trip planning . Startling results.

[John “Jay” Menna]

I’ve been experimenting with AI/large language models for work, mostly to see how well they handle language-based tasks. They’ve proven to be excellent at things like translation and cleaning up speech—way better than I expected. But I never imagined they could assist with something like route planning and provisioning for a trip.

I’m planning a trekking trip to Greenland this summer, exploring a couple of different areas. While communicating with local outfitters and guides, I realized that working in English was creating some challenges. I thought it’d be smart to translate all my communication to make things clearer on both sides.

So, I wrote up my detailed itinerary and headed over to Anthropics’ Claude model. In my excitement, I forgot to give it the explicit command to “translate this” and instead just pasted my entire letter to the outfitter/guide. The AI ended up processing my whole itinerary instead.

The results were both impressive and a little unsettling. Here’s what I got back:

https://claude.site/artifacts/1ca52510-de1b-443f-b96c-18a3f4c1c3fb

I’m now systematically going through the response out of curiosity, and so far, it seems surprisingly accurate. That said, it rounded the GPS coordinates to just one decimal place, which has moved some waypoints and campsites into the middle of bodies of water. Not ideal.

When I followed up to ask it to refine the GPS coordinates, things took a strange turn—it started hallucinating and generated a three-week itinerary for the John Muir Trail! No matter how I prompted it, I couldn’t get it back on track to refine the original itinerary.

I figured this might be interesting for this group—both as an example of the strengths of these tools and as a cautionary tale about the occasional weirdness they produce.

For those that are an interested, here is my original reqest that ended up being the prompt for the LLM:

I will be in Greenland during the second half of August 2025, arriving in Narsarsuaq from Nuuk on <Date time>.

My wife and I are avid solo hikers, so while we don’t require a guide on the trail, we would appreciate support with the logistics and planning. If your company offers services to assist with these aspects, we’d love to discuss hiring them..

Our initial plan is a five-day hike from Itilleq to Qaqortoq, with one night in Qaqortoq, where we’d like lodging that includes a bath and electricity (nothing fancy). We’re told the view from the top of Mt. Killavaat is spectacular, so we’re considering adding that to our route.

I would like to rent or buy bear spray, flares, and buy gas. Afterward, we’re interested in a resupply at Qaqortoq, followed by a journey to Tasermiut Fjord. From there, we’d like to hike to the col above Tasiusaq. We are still finalizing the details of this section, but Tasermiut is a priority for us. Here’s a summary of the assistance we’d need:

1. Boat transfer from Narsarsuaq to Itilleq on arrival.
2. Transporting a single bag from Itilleq to Qaqortoq for collection upon our arrival there.
3. Arranging transport to and from our hike in Tasermiut Fjord.
4. Return transportation to Narsarsuaq by <date Time>
If there are any alternative routes or scenic spots around Tasermiut that you’d recommend, please feel free to share.

We’re open to adjusting the itinerary for memorable experiences. Some have suggested changing igaliku in favor of Tasiusaq. I welcome your suggestions.

[Terran Terran]

I’ve found it easy to throw off AI, though it probably depends on the source. It may be cheaper to leave the wife at home and bring an AI girlfriend. It’s came a long ways, but it’s not there yet. At least in the versions we get.

[Paul Hatfield]

It seems like you used machine translation, which at this point in time (2024), works quite well.

But at this point in time, “artificial intelligence” is not more intelligent than your family pet.

[David D]

I gave ChatGPT a shot at planning a 3 day fall trip in the Adirondacks with some peak bagging along with some criteria, hoping it would give an interesting starting point to tweak around.  It’s response was still well off from prime time.

Once these models finally get dialed in, they may prove useful as a marker for routes to actually avoid in heavily trafficked areas because I imagine they’ll be giving similar recommendations to everyone.

[John “Jay” Menna]

What’s most interesting is that I’ve been planning this trip for ages, yet the key info wasn’t online… in English.

After grilling Claude/Anthropic for a while, I found it was pulling from a primary Danish source—and even extracted details from parsing a GPX file.

Greenland is linguistically tricky—place names were historically Danish but have recently reverted to Greenlandic. Yet the AI navigates between languages and GPX files like it’s all the same thing.

The linguistic trapeze act isn’t surprising, but extracting longitude, latitude, and elevation from a GPX file and weaving the points seamlessly into the document? That’s pretty cool.

[Zero Flight]

What happens if you feed the same query but with added clauses for GPS coordinates to include up to at least 5 decimal places?

[Dan]

This is interesting. I just asked ChatGPT to plan a loop backpacking trip near a popular destination (geological formation) that I’m familiar with. It gave me a 3-5 day itinerary with an introduction describing the destination and the wilderness area. The itinerary looked very reasonable at first glance, with the name of the trailhead, trail names, trail numbers, suggested campsites, etc. However, looking more closely, it is total nonsense. Half of the trails don’t exist, the trailheads don’t exist, the trails that do exist don’t connect, the trail numbers don’t correspond to the trail names, the entire route makes no sense.

Apparently, the OP has discovered a major gap in the abilities of generative AI. It simply can’t connect up information in the way required to plan a backpacking route. And the fact that it presents the route with total confidence is definitely problematic.

[Paul Hatfield]

This is like asking a Magic 8 Ball a question and expecting an accurate response.

[John “Jay” Menna]

What happens if you feed the same query but with added clauses for GPS coordinates to include up to at least 5 decimal places?

It went from  N60°11.800′ W44°49.200   to    N60°11.80000′ W44°49.20000

Technically correct but worthless.

 

[John “Jay” Menna]

It simply can’t connect up information in the way required to plan a backpacking route. And the fact that it presents the route with total confidence is definitely problematic.

That seems to be the case—it’s untrained AI, after all. I bet it could perform much better with proper training.  The real challenge is deciding whether to spend time training it, only for a new model to drop and render all the hard work obsolete.

[Dan]

Agree, a LLM is probably not even the right model to use for this.

[Geoff Caplan]

Well – I doubt that anyone is going to spend $$$billions on training an AI on marginal topics such as obscure trekking routes documented in minor languages. There just isn’t enough data out there – and the data it did show you pretty much amounts to copyright theft, if there was only a single source.

Though I wouldn’t be surprised if we eventually get decent information on more mainstream routes like the AT and PCT, given all the blogs that cover them.

More generally, I suspect that generative AI is being overhyped. As with any new technology, they make rapid progress while they pluck the low hanging fruit. But then farther progress gets exponentially harder. A number of the analysts I follow feel that this point is rapidly approaching.

This is compounded by the fact that AI is taking traffic from the very websites that supply them with their training material. Why invest the time and money if your intellectual property is going to be nicked? And the big AI companies have retained expensive copyright lawyers in an attempt to intimidate small content producers who might want to take them on.

So will we ever be able to trust AI on safety-critical issues such as hiking routes? Just yesterday, an AI that claims to be able to solve Olympiad-level math questions screwed up  my simple temperature conversion by an order of magnitude. And any time it generates non-trivial code, I spend almost as long debugging as it would take me to write it from scratch. It’s worth it, because it will sometimes suggest genuine improvements, but half the time it’s outputting garbage. So it’s a aide, not an alternative.(And yes – I write clear and comprehensive prompts!)

TL/DR? Worth using for initial ideas, but we still have to double-check everything it suggests.

[David D]

Content will be king.  I work in product development offering Ai tools but customers guard their data like crown jewels, as they should

Some orgs are getting desperate for data.  LinkedIn was making false appeals in email to “select experts to share your knowledge with other linked in users”:  i.e. appeal to the user base ego to create their training content for them.

Similar to the rip off that led to the sale of the Huffington post in the pre-AI era, at the advent of the blog explosion, where the platform was fashioned as a socially conscious vehicle for change soliciting users upload their free work, only for it to be eventually sold for a huge sum.

[Geoff Caplan]

@David D

Yes – they can only improve their models if they can access more and better data. But they are destroying the very sources they rely on, and creators are becoming much more protective of their intellectual property. Just look at the firestorm Adobe ran into when it was suspected their were using the media in their stock library to generate images and video that would put their creators out of business. It’s difficult to see where they can go from here. I haven’t noticed major improvements on any of the 5 generative AIs I use for many months now.

Plust the training and data centre costs are exorbitant – they are all losing money.

It’s a bubble. There’s going to be a huge shakeout, and short sellers who time it right are going to make a lot of money.

[David D]

Geoff, Have you tried the for-pay models?  I don’t use them but I hear that’s where the big improvements are still happening.  I agree it can’t keep on forever

I work in fiber optics and the build out growth in the data centers is outrageous.  I was there for the dot.com boom (and bust) in 2000 and this is something else again.   I’ve never seen a demand curve for optics like this, driven by training and data center interconnect.

One big difference between this bubble and the dot com one is the FANGs are sitting on mountains of cash to burn through whereas the dotcom bubble was built by new entrants borrowing money to put fiber in the ground, and system providers were happy to lend it at their peril (I’m looking at you Nortel).  So when the bubble popped, so many companies when bankrupt.  Once the demand curve for AI pops (as copyright cases wind through courts etc), I don’t think we’ll see near the same scope of dead companies, though select stocks will be badly punished.  I bought a lot of Nvidia a few years ago as this was an obvious one so not complaining, but timing will be tricky.

When the fiber build out happened in the dot com bubble, no one really knew how it would get used. Very similar to the AI build out today, it was a “build it and they will come” mentality.   It took years for a killer app (Netflix, Youtube) to come and take advantage of it.   It’s hard to predict the future and there’s a good chance in 5 years some killer app will use up all this AI compute as well.

I’m surprised there hasn’t been more gov’t pushback, given their goals to reduce carbon emissions, as the technology is shockingly costly in power consumption.  High levels of hypocrisy but no one wants to be seen as the one to slay the golden goose.

[Jerry Adams]

Will AI be using this thread as training data?

I’ve played with Gemini a little.  An amusing party trick if nothing else.

It often produces amazingly good responses.  Yeah, you have to verify it, it does gibberish sometimes.

I told it to write a c program to use aristosthenes sieve method to find the first 100 prime numbers and it seemed to do so but I didn’t test it

[Jerry Adams]

I asked Gemini when the eclipse would be in Austin a few days before hand but it told me it couldn’t because the eclipse already happened.  I told it no, the eclipse has not already happened and it politely apologized and then repeated the same error

[Geoff Caplan]

Jerry – Gemini is strikingly hopeless. For some reason Google is way behind the curve, and isn’t catching up that I can see. Though as you say, the responses are sometimes fine – it’s just that the percentage of mistakes is so high, even on simple arithmetic.

But they can be useful for MYOG topics. I’ve just had some impressive  chats with Claude on optimising the geometry of tipis, for example.

David – yes, the investment in AI infrastructure is jaw-dropping. Nvidia has reportedly generated $70 billion in sales to AI companies and data centres – becoming the most valuable company on the globe in the process. But analysts calculate that AI companies haven’t generated anything approaching that kind of revenue, so it’s difficult to see how this can be sustained.

As you point out, the ecological implications are significant. A generative AI query will require 5-50 times the energy of a conventional search query. I recently heard an interview with an energy planner, and the industry is giving them nightmares. Given the lead time on grid provision they’re having to estimate the future impact, but the uncertainties are so great that this is virtually impossible. The headaches are compounded by all the uncertainties around the transition to EVs. It’s not clear how all this demand can be met by green energy generation.

[David D]

Google’s plan to buy small modular nuclear reactors to fill this need concerns me.  This company’s business strategy is based on “good enough” which won’t cut it for this.  I see this being tied up in regulatory for a long time, unless the new administration nukes the safeguards (which would be bad).

[Geoff Caplan]

We’re wandering OT here, so I’ll try to keep it brief.

On the nukes all I can say is that I know a guy who was chief engineer on one of the Royal Navy’s Trident nuclear subs. For 60 years now, they have been using these Rolls Royce reactors in that highly confined space, and the safety record is impeccable. Rolls is one of the companies working on civilian SMNRs. Based on his hands-on experience my engineer friend believes fervently that mass-manufactured SMNR reactors have a key role to play in transitioning away from hydrocarbons. They are a safe and proven technology.

I spent a little time as an energy economist, and people seem to have a distorted perception of risk. OurWorldInData.org puts it rather well. Imagine an average sized European town of 150k people.  How would it fare if supplied by different power sources?

Coal: 25 people would die prematurely every year (mostly from pollution);
Oil: 18 people would die every year;
Gas: 3 people would die every year;
Hydro: 1 person would die;
Wind: 1 death every 25 years;
Nuclear: 1 death every 33 years;
Solar: 1 death every 50 years.

So nuclear is over 800 times safer than coal! It’s safer than wind, and almost on a par with solar. The widespread fear of nuclear is irrational – and I’m speaking as someone who used to be an active anti-nuclear campaigner.

The real issue with nuclear has been cost per unit – but new community-scale fission technologies coming on stream promise to be competitive with renewables and the grid-scale storage they require – and they achieve even higher levels of safety There is also at least one community-scale thorium salt technology that is ready for production (the Terrestrial Energy IMSR ). Thorium generates significantly less waste, and there are some radical waste reduction and recycling technologies coming on stream in the next decade or so.

As a species we have got ourselves into a potentially terminal mess, and we should be harnessing every technology we can to turn things around before it’s too late. I do wish that people would be more open-minded about the potential role of modern nuclear technologies in this challenging transition. At the moment we have absurdities like Germany decommissioning perfectly good nuclear plants and replacing them with far more lethal and damaging hydrocarbons – all driven by ill-informed moral panic…

[David D]

My tennis partner is a technical director at our national nuclear safety commission. His opinion mirrors your friend’s. I agree it’s use should be expanded and help alleviate the power crisis looming in data centers so we can use these models to plan our trips (see what I did there?)

I’m not worried about safety of the technology with proper oversight.  But that’s not Google.

[Geoff Caplan]

Well, if you know anything about the safety commission, you should rest easy in your bed.

They are pretty ruthless, and Google will have no say on safety matters.

If anything they are too conservative, and are sometimes accused of stifling innovation. Much of the interesting stuff seems to be happening in Canada, where the regulators seem to be more willing to work with innovative companies.

[David D]

I’ve heard complaints that India is possibly moving ahead of the US in SMNR tech, reflecting the stifling of innovation, but don’t follow it close enough.

Way way off topic now: I’m worried that deregulation could loosen this grip too much, but I hope you’re right.  For example, data center customers care about total cost of ownership and will chew through power if its cheaper.  It’s this bottom line sort of thinking at any cost that worries me if the regulatory grip gets loosened on SMNR.  So while safeguards look good now, no guarantee they’ll stay that way

[jscott]

We plunge headlong into AI, that requires immense energy expenditures–like bitcoin–for no real gain. Or have folks found their lives improved due to AI? I’ve heard from some AI developers that it’s far less intelligent than a cat. Or a hamster. so, in the midst of a global warming crisis, we go mad over a technology that requires billions of megawatts so that we can get inaccurate information about a backpacking trip.

there are these things called maps, ya see, that are far more reliable and don’t require nuclear reactors to produce.

[Jerry Adams]

John – “It went from  N60°11.800′ W44°49.200   to    N60°11.80000′ W44°49.20000”

I think the AI has just passed the turing test

That is what a human would do, just to screw with you

[Author]

Posts