Since concluding my travels in Latin America and returning home to Singapore in late May, I’ve been trying to engage in a couple of personal projects. A significant theme underlining these disparate pursuits, unsurprisingly, is my desire to commemorate my travels meaningfully.

In addition, now that I’m back to my chronically online self, I’ve also been paying more attention to AI and thinking about how this technology can be useful in my life. It was in this context that had a brain fart: I could use try to make sense of my travels with AI.

Inspired by the science fiction writer Ken Liu’s take on the potential of AI as an artistic medium that enables artists to “play with captured subjectivities” like no other, I reflected on what exactly AI could let us do that we otherwise could not. What kind of thoughts and gestures would have been impossible or exceedingly difficult if not for AI?

I still don’t have a good answer now, but I suspect that it would have to involve the ability of AI models to perform open-ended simulations—simulations that can be continually orchestrated by natural language over time, unbounded by the strict logic of code.

In other words, I should lean into the simulacra that AI models are adept at generating, and just do more of that. This is the specific affordance of AI I should try to explore and play with, if I want to have a better sense of how this technology can be ultimately transformative and valuable to my life.

In this particular case of trying to interrogate my travel experiences then, I thought of tapping on AI to simulate the act of deciding which places to visit within a particular region. To further lean into AI’s simulation superpowers, I would also direct the AI not to simply recommend me a single travel route, but to imagine a broader set of traveller personas and recommend multiple travel routes tailored for them.

My rough idea here was to use AI to simulate other potential paths my wife and I could have taken while travelling, and in doing so, we could perhaps better reflect on the choices we made, as well as better appreciate the route we actually took.

To keep things manageable in this exercise, I decided to just aim for 10 simulated routes and limit the simulations to Mexico (not the entire Latin America).

First, I asked Google’s Gemini 2.5 Pro (the AI model that I’m predominantly using) to come up with 10 different “personas” for travellers who want to travel across Mexico long-term, giving each of them specific characteristics based on a consistent set of criteria. While I proposed some examples of such criteria to the AI model, I also explicitly gave it the space to propose more criteria.

The generated and consolidated result—with some light edits from me to make them more realistic—is appended in the table below. While some of the personas reflect clear stereotypes, I thought this was really well-done as a first cut.

Next, I attached the table of 10 personas to the same AI model and asked it to recommend a travel route for each persona. To make these routes more comparable, I set a definite start point for the route in Mexico, i.e. Tijuana, to mirror our actual route, and also restricted the travelling duration to a fixed 10 weeks, roughly the amount of time we actually spent in Mexico.

My prompt is laid out in full below:

Taking just a couple of minutes, Gemini 2.5 Pro promptly generated 10 travel routes as I asked for, each one corresponding to the specific personas we generated earlier. I consolidated all the 10 routes together in a table as well and included our actual travel route for comparison.

To provide a sense of the differences in the recommended routes (or the lack thereof), I also mapped out each of the route on Google Earth. Each colour represents a recommended route (many are stacked on top of each other as they go through the same places). Our actual route is marked out in yellow (only the northern parts can be seen).

At first glance, having 10 travel routes might seem like an overkill. A single person would probably just need one or a couple more to have enough reference points. What was the point of generating so many travel routes?

The obvious one to me is that the limitations of the AI model became more apparent with the greater number of simulations.

Firstly, with the outcomes of 10 simultaneous simulations laid out in front of me, I could see the biases of the Gemini 2.5 Pro much more readily. In this case, the AI model was really not fond of recommending northern Mexico. Only three out of the 10 recommended travel routes involved destinations in northern Mexico, i.e. the routes for personas 1, 6 and 7, which included destinations in the Baja California peninsula and/or the Copper Canyon region.

When I probed why, the reasons it gave were not surprising: most of the established tourist attractions were in central and southern Mexico, and there was a higher risk of organised crime in the north. This was exactly what we encountered in the travel literature and actual travel patterns, e.g. we barely met any foreign tourists in the Copper Canyon region. Hence, we should expect the same biases in the data that the AI model was trained on, and having multiple simulations definitely helped to render them more apparent.

In tandem with the biases, having more simulations also provided greater room for outliers in the data to surface. There were quite a few destinations that only appeared once in the 10 travel routes. Furthermore, a couple of them were destinations that I did not know of before, e.g. Tepoztlán (persona 6) and Mazunte (persona 5). Again, these destinations likely have less mentions in the travel literature, and they might not have been surfaced if not for the greater number of simulations done.

Secondly, having more simulations made it easier to spot idiosyncrasies in how the AI model interpreted the prompt (including the specific contexts of the various personas) and decided on its recommendations.

For example, for persona 8 modelled after a content creator, a segment in its recommended travel route was quite impractical in real life. As you can see from the image below, the “Valladolid → Mérida → Izamal → Holbox → Tulum” segment (light green line) would entail quite a bit of backtracking.

When I queried the AI model on this, it shared that it sequenced the destinations based on creating the best possible series of videos, starting with a bang with the adventure-themed highlights in Valladolid, then moving on to a cultural focus in Mérida and Izamal, before concluding with some of Mexico’s most iconic beachfront attractions in Holbox and Tulum. In its words: “while going east (Valladolid), then west (Mérida), then north (for Holbox) seems inefficient on a map, it creates a much better story for a vlogger than a simple linear path.”

I didn’t quite buy this argument since a content creator can decide later on how they want to publish their content. However, this case was still insightful in so far that it showed that the AI model could take the context too literally, at the expense of other considerations we might find more important.

In sum, by forcing the AI model to engage in multiple simultaneous simulations, it becomes clearer to see the imperfections latent in these simulations. They reflect biases in the model’s training data: great at parroting the consensus, not so great at recommending what is not. In addition, they also reveal idiosyncratic limitations in the interpretative and reasoning abilities of the AI model, especially in such scenarios where the choices are highly subjective.

No doubt, with more training, these AI models will likely become better, their shortcomings less stark and less salient. Regardless, I believe that we ought to maintain a critical stance towards AI, especially as this technology is fundamentally about capturing, portraying and giving us the ability to manipulate subjectivities. Simple gestures like what I just did—prompting the AI model to engage in multiple simultaneous simulations—should help us remain on our toes and avoid being overly seduced by it.

More broadly, this exercise is a reminder that context is indeed king—critical for not only human reasoning but machine simulation too.

Not surprisingly, the AI world seems to be moving from “prompt engineering” to “context engineering” as well. People are starting to recognise that what we get out of AI will be determined more and more by our ability to parse, articulate, and simulate different contexts.

Specific to this little exercise, although the differentiated contexts—the 10 personas—were rudimentary, forcing Gemini 2.5 Pro to engage with them has made its outputs significantly more interesting. The manifold possibilities inherent in the act of deciding where to travel to became more apparent with the greater number of simulated contexts. Our choices could be shaped by an endless array of variables, many of which might even be invisible to us. Ultimately, there are just so many ways to travel, and simulations through AI may just be what is needed to render the sheer breadth and depth of these possibilities more accessible to us.

Furthermore, by forcing AI to simulate more contexts, we are also forcing ourselves to confront more possible outcomes. This was certainly possible to do pre-AI, but the difference now is that AI can do so quickly and at scale. The time that I would have taken to conceptualise 10 difference travel routes across Mexico would have been an order of magnitude longer than what Gemini 2.5 Pro had taken. Conceptually then, it seems that AI can help to sharpen our instincts for a more expansive way of thinking, empowering us to perceive the world from 10, 100 or even 1,000 different perspectives at a time.

In this sense, just as how speech rendered our thoughts relational (by allowing us to communicate complex concepts) and writing enabled analytical and abstract thinking (by allowing our words to stand on their own, outside of the immediate context of their author), AI may very well be another transformative cognitive tool. It can enable us to think and learn via simulating multiple contexts at a scale impossible before. Where we once had to devote significant cognitive resources to even put ourselves in another person’s shoes, we now can do so for multiple perspectives in a single prompt. The implications for our cognition here will no doubt be profound.

Early research on ChatGPT use among students appears support this, suggesting that while overreliance on such AI tools may impede learning skills, a more strategic use of such tools after foundational skills are built up may still enhance learning. So instead of trying to use AI to “one shot” answers and completed work, it may be more useful to harness AI as a “multi-shot”—a tool not to narrow down our thought processes, but rather to open us up to new cognitive trajectories.

I probably wrote way too much about an interaction with an AI model that took less than 10 minutes. But I couldn’t help but feel something potent in this simple gesture: to use AI to simulate contexts far beyond what I can practically imagine and gain an expanded way of seeing.

Bringing this back to the subject of travel, which kicked off this exercise in the first place, I’ve always enjoyed how visiting another country can provide a form of “cognitive unlock”. Travel encourages us to see the world through the eyes of another culture, and in the process, we learn that there are just so many different ways to be human.

Travelling with AI shouldn’t be different, except that we can now do so on even more dimensions. Multiple travel routes, multiple backstories, multiple states-of-mind—how all of this will pan out is as yet uncertain, but I’m certainly all for it!

Since concluding my travels in Latin America and returning home to Singapore in late May, I’ve been trying to engage in a couple of personal projects. A significant theme underlining these disparate pursuits, unsurprisingly, is my desire to commemorate my travels meaningfully.

In addition, now that I’m back to my chronically online self, I’ve also been paying more attention to AI and thinking about how this technology can be useful in my life. It was in this context that had a brain fart: I could use try to make sense of my travels with AI.

Inspired by the science fiction writer Ken Liu’s take on the potential of AI as an artistic medium that enables artists to “play with captured subjectivities” like no other, I reflected on what exactly AI could let us do that we otherwise could not. What kind of thoughts and gestures would have been impossible or exceedingly difficult if not for AI?

I still don’t have a good answer now, but I suspect that it would have to involve the ability of AI models to perform open-ended simulations—simulations that can be continually orchestrated by natural language over time, unbounded by the strict logic of code.

In other words, I should lean into the simulacra that AI models are adept at generating, and just do more of that. This is the specific affordance of AI I should try to explore and play with, if I want to have a better sense of how this technology can be ultimately transformative and valuable to my life.

In this particular case of trying to interrogate my travel experiences then, I thought of tapping on AI to simulate the act of deciding which places to visit within a particular region. To further lean into AI’s simulation superpowers, I would also direct the AI not to simply recommend me a single travel route, but to imagine a broader set of traveller personas and recommend multiple travel routes tailored for them.

My rough idea here was to use AI to simulate other potential paths my wife and I could have taken while travelling, and in doing so, we could perhaps better reflect on the choices we made, as well as better appreciate the route we actually took.

To keep things manageable in this exercise, I decided to just aim for 10 simulated routes and limit the simulations to Mexico (not the entire Latin America).

First, I asked Google’s Gemini 2.5 Pro (the AI model that I’m predominantly using) to come up with 10 different “personas” for travellers who want to travel across Mexico long-term, giving each of them specific characteristics based on a consistent set of criteria. While I proposed some examples of such criteria to the AI model, I also explicitly gave it the space to propose more criteria.

The generated and consolidated result—with some light edits from me to make them more realistic—is appended in the table below. While some of the personas reflect clear stereotypes, I thought this was really well-done as a first cut.

Next, I attached the table of 10 personas to the same AI model and asked it to recommend a travel route for each persona. To make these routes more comparable, I set a definite start point for the route in Mexico, i.e. Tijuana, to mirror our actual route, and also restricted the travelling duration to a fixed 10 weeks, roughly the amount of time we actually spent in Mexico.

My prompt is laid out in full below:

Taking just a couple of minutes, Gemini 2.5 Pro promptly generated 10 travel routes as I asked for, each one corresponding to the specific personas we generated earlier. I consolidated all the 10 routes together in a table as well and included our actual travel route for comparison.

To provide a sense of the differences in the recommended routes (or the lack thereof), I also mapped out each of the route on Google Earth. Each colour represents a recommended route (many are stacked on top of each other as they go through the same places). Our actual route is marked out in yellow (only the northern parts can be seen).

At first glance, having 10 travel routes might seem like an overkill. A single person would probably just need one or a couple more to have enough reference points. What was the point of generating so many travel routes?

The obvious one to me is that the limitations of the AI model became more apparent with the greater number of simulations.

Firstly, with the outcomes of 10 simultaneous simulations laid out in front of me, I could see the biases of the Gemini 2.5 Pro much more readily. In this case, the AI model was really not fond of recommending northern Mexico. Only three out of the 10 recommended travel routes involved destinations in northern Mexico, i.e. the routes for personas 1, 6 and 7, which included destinations in the Baja California peninsula and/or the Copper Canyon region.

When I probed why, the reasons it gave were not surprising: most of the established tourist attractions were in central and southern Mexico, and there was a higher risk of organised crime in the north. This was exactly what we encountered in the travel literature and actual travel patterns, e.g. we barely met any foreign tourists in the Copper Canyon region. Hence, we should expect the same biases in the data that the AI model was trained on, and having multiple simulations definitely helped to render them more apparent.

In tandem with the biases, having more simulations also provided greater room for outliers in the data to surface. There were quite a few destinations that only appeared once in the 10 travel routes. Furthermore, a couple of them were destinations that I did not know of before, e.g. Tepoztlán (persona 6) and Mazunte (persona 5). Again, these destinations likely have less mentions in the travel literature, and they might not have been surfaced if not for the greater number of simulations done.

Secondly, having more simulations made it easier to spot idiosyncrasies in how the AI model interpreted the prompt (including the specific contexts of the various personas) and decided on its recommendations.

For example, for persona 8 modelled after a content creator, a segment in its recommended travel route was quite impractical in real life. As you can see from the image below, the “Valladolid → Mérida → Izamal → Holbox → Tulum” segment (light green line) would entail quite a bit of backtracking.

When I queried the AI model on this, it shared that it sequenced the destinations based on creating the best possible series of videos, starting with a bang with the adventure-themed highlights in Valladolid, then moving on to a cultural focus in Mérida and Izamal, before concluding with some of Mexico’s most iconic beachfront attractions in Holbox and Tulum. In its words: “while going east (Valladolid), then west (Mérida), then north (for Holbox) seems inefficient on a map, it creates a much better story for a vlogger than a simple linear path.”

I didn’t quite buy this argument since a content creator can decide later on how they want to publish their content. However, this case was still insightful in so far that it showed that the AI model could take the context too literally, at the expense of other considerations we might find more important.

In sum, by forcing the AI model to engage in multiple simultaneous simulations, it becomes clearer to see the imperfections latent in these simulations. They reflect biases in the model’s training data: great at parroting the consensus, not so great at recommending what is not. In addition, they also reveal idiosyncratic limitations in the interpretative and reasoning abilities of the AI model, especially in such scenarios where the choices are highly subjective.

No doubt, with more training, these AI models will likely become better, their shortcomings less stark and less salient. Regardless, I believe that we ought to maintain a critical stance towards AI, especially as this technology is fundamentally about capturing, portraying and giving us the ability to manipulate subjectivities. Simple gestures like what I just did—prompting the AI model to engage in multiple simultaneous simulations—should help us remain on our toes and avoid being overly seduced by it.

More broadly, this exercise is a reminder that context is indeed king—critical for not only human reasoning but machine simulation too.

Not surprisingly, the AI world seems to be moving from “prompt engineering” to “context engineering” as well. People are starting to recognise that what we get out of AI will be determined more and more by our ability to parse, articulate, and simulate different contexts.

Specific to this little exercise, although the differentiated contexts—the 10 personas—were rudimentary, forcing Gemini 2.5 Pro to engage with them has made its outputs significantly more interesting. The manifold possibilities inherent in the act of deciding where to travel to became more apparent with the greater number of simulated contexts. Our choices could be shaped by an endless array of variables, many of which might even be invisible to us. Ultimately, there are just so many ways to travel, and simulations through AI may just be what is needed to render the sheer breadth and depth of these possibilities more accessible to us.

Furthermore, by forcing AI to simulate more contexts, we are also forcing ourselves to confront more possible outcomes. This was certainly possible to do pre-AI, but the difference now is that AI can do so quickly and at scale. The time that I would have taken to conceptualise 10 difference travel routes across Mexico would have been an order of magnitude longer than what Gemini 2.5 Pro had taken. Conceptually then, it seems that AI can help to sharpen our instincts for a more expansive way of thinking, empowering us to perceive the world from 10, 100 or even 1,000 different perspectives at a time.

In this sense, just as how speech rendered our thoughts relational (by allowing us to communicate complex concepts) and writing enabled analytical and abstract thinking (by allowing our words to stand on their own, outside of the immediate context of their author), AI may very well be another transformative cognitive tool. It can enable us to think and learn via simulating multiple contexts at a scale impossible before. Where we once had to devote significant cognitive resources to even put ourselves in another person’s shoes, we now can do so for multiple perspectives in a single prompt. The implications for our cognition here will no doubt be profound.

Early research on ChatGPT use among students appears support this, suggesting that while overreliance on such AI tools may impede learning skills, a more strategic use of such tools after foundational skills are built up may still enhance learning. So instead of trying to use AI to “one shot” answers and completed work, it may be more useful to harness AI as a “multi-shot”—a tool not to narrow down our thought processes, but rather to open us up to new cognitive trajectories.

I probably wrote way too much about an interaction with an AI model that took less than 10 minutes. But I couldn’t help but feel something potent in this simple gesture: to use AI to simulate contexts far beyond what I can practically imagine and gain an expanded way of seeing.

Bringing this back to the subject of travel, which kicked off this exercise in the first place, I’ve always enjoyed how visiting another country can provide a form of “cognitive unlock”. Travel encourages us to see the world through the eyes of another culture, and in the process, we learn that there are just so many different ways to be human.

Travelling with AI shouldn’t be different, except that we can now do so on even more dimensions. Multiple travel routes, multiple backstories, multiple states-of-mind—how all of this will pan out is as yet uncertain, but I’m certainly all for it!