If LLMs are cars, do we want a “car-centric” internet?

A car allows you to go further, faster than you would on your own two feet, but: - it requires some knowledge to be driven properly - it can make you lazy if you use it all the time for any very short trip - it encourages you to rush, packing more into your day because you can - it has a ecological and financial cost, compared to walking - it has changed the landscape

This post is my random musings on why 1) AI feels very similar at the moment, and 2) maybe we want to go beyond a car-centric internet.

Why LLMs are cars

Driving requires a licence and skills For a car, you don’t just “get in and go.” You need a license, and some skills to navigate. Knowing how to use AI requires you to actually understand risks and limitations, then keep watch and make sure you’re going in the correct direction.

If you don’t steer the LLM properly, it will drive you into a ditch with perfect confidence.

Note: We’re treating AI more like a bike than a car on that front at the moment - just learn as you go, if you crash, you’ll mostly hurt yourself… but the risk to others is increasing with agentic AI.

Going short distances
Just as a car atrophies your legs if you use it to go down two streets instead of walking, AI atrophies your memory/ tolerance for friction/thinking capacity.

We outsource the small cognitive tasks until we forget we ever knew how to do them.

Going long distances
A car open up distance places, at the cost of some carbon consumption and driving fatigue. However, it’s a very different to rush through a highway or to hike the same distance. Both are fine, but sometimes it’s more interesting (though less convenient) to walk for a very long time, get lost, backtrack, take wrong turns, and discover something unexpected…

Sometimes you want to arrive at answers, sometimes you need to journey through the question.

Moving too often
The car can make distance feel trivial, so we pack more trips, errands, meetings into each day. More more more more more. But every item adds cumulative fatigue that doesn’t register until the end of the day.

AI, similarly, changes our relationship to time. Problems that once demanded an hour of quiet contemplation now demand 30 seconds of prompting. So we prompt more, we output more, and the fatigue of double checking accumulates all the same.

Carbon cost
The AI car runs on a fuel that isn’t gasoline, but data and computation. It produces a massive carbon footprint. Every “trip” you take—generating an image, summarizing a document, brainstorming an idea—burns a small amount of fossil fuel in a server farm somewhere. The aggregate cost is staggering.

LLMs/Cars have been changing the city

Cars didn’t just fill roads; we rebuilt the cities around them (especially in countries like the US). That created suburbs, shopping malls, and the death of the corner store.

Similarly, we’re reshaping the digital landscape for AI and agents.

Search engines are becoming answer engines. Social media feeds are becoming synthetic. We are building a “car-centric” cognitive city, where walking (slow, organic research) is not only inconvenient but genuinely harder to do. Finding new creators, tiny websites, exploring the web via linkrolls like we did in early 2000 is now complex.

It’s becoming increasingly hard to get off the highways of information, but that means missing out on hidden gems and getting bombarded with actually irrelevant content instead (ads, SEO-optimized filler, AI-generated noise, among others).

An alternative to cars: bike-centrism and Denmark

In Denmark, it seems like biking is your main “moving around” mode: nine out of ten people own a bike, kids learn how to bike at age 2 or 3. Individuals cycle 1.4 km a day on average, and it’s often much faster than driving.

This is no magic accident (policy/infrastructure/some public organisation). In the 1950s, like everywhere else, Danish urban planners believed the future belonged to cars (wider highways,yadda yadda). Then the 1970s oil crisis hit, protests demanded a car-free city.

Nowadays, Denmark combines heavy taxes on oil and automobiles and cities designed around bikes. Copenhagen alone has around 400 kilometres of cycle paths—clearly separated from car lanes and sidewalks. Even better, there are “cycle superhighways” where commuter’s needs have the highest priority (for example, with traffic lights timed to average cycling speed), usually super well connected to bus and train stations.

What would a bike-centric internet look like?

Right now, the web is becoming a fast highway system - you enter through a search engine like google, through platforms like facebook/tiktok/instagram, and don’t really move around - it’s even worse with LLMs, where you don’t even go out of the chat window. I’ve had discussions with people who want the web to become “agentic-first”, making websites compatible with AI more than focusing on human exploration, thinking that now the web will be accessed “through AI only”.

Agent APIs are needed for some cases, sure. But highways are terrible places to stroll, and you don’t always need to use a car. They’re designed to get you somewhere quickly, not to let you wander - and can make you lazy, rushed and tired.

What if we encouraged the use of different tools depending on different needs instead of reinforcing the highway? What if our designs rewarded slowness, curiosity, and detours (the human experience, in sum), in parallel with the efficient highways for our busy corporate overlords’ demands?

Some ideas that already exist:

• Efficient highway sign posts: Where you actually want to use LLMs, they should be able to access information fast and easily. This means clear, unambiguous, and well documented APIs for sites which want agent integration.
• Natural reserves: Digital “slow web” spaces, where AI is prohibited (no crawlers, no summaries, …). Human-made content that you get directly, as is, without intermediation and distractions, without ads, without popups. A good example might be what people are doing with the Gemini protocol, an alternative to HTML.
• Trail markings: Organic resharing of information - two tools were very good for this, but have kinda disappeared: RSS (replaced by mail newsletters and feeds) and blogrolls (since now most things are platform based) - it could be cool to reintroduce these (I suspect that, if we want to have indiviudal agents tailored to our needs, they would also benefit from RSS and personal curation)

Some random ideas that don’t exist, as far as I know, but it might be fun if they could:

• Stop areas: Just places that encourage you to go offline and think
• “Speed bumps” for information: Introducing deliberate friction to load flagged posts (fake news, strong divisive emotions trigger posts, …), to help make anti-viral content. They could load slowly, or require a captcha to open?
• Viewpoints: Pages automatically aggregating a variety of high quality sources on a topic with a focus on different cultural or historical perspectives, and pointing back to the original authors.

Cities for cars were a choice. The AI internet city can be too. We can have highways for speed and efficiency, bike lanes for daily things, and walking trails for exploration and discovery… but only if we consciously design for all of them.

I hope we build our city with room to wa/onder.