AI - Bicycle or Train?

The metaphor works on many many levels

Image generated by Grok (in sloppy train mode)

In my previous post about AI, I discussed the metaphor of using AI like riding a bicycle or riding a train.

The more I think about it, the more I like the metaphor because it works in many ways.

First is the obvious one that was first thought of:

AI as bicycle enhances your own efforts, but you have to provide the effort in the first place. AI as train just takes you to your destination for the price of a ticket and no effort is required.

That’s a good insight. Especially since the big AI companies all seem to want to build the trains and charge us to use them.

Free-riders

It can be extended slightly by observing that the AI trains have a lot of free-riders on them. People like me who have signed up for the free accounts at various AI services and have no intention of converting those accounts into paid ones¹. It is, as I noted repeatedly in the previous article, remarkably hard to get figures out of Big AI that tell you what proportion of users, let alone what proportion of usage, is ticket buying customers vs fare dodgers and that is clearly deliberate because the numbers are worse that the proportion of fare dodgers on any US public train or tram network. However the invaluable Ed Zitron did some reading of a recent article at the Information and figures the following:

OpenAI reportedly had 55m paying ChatGPT customers at the end of Q1 — up from 47m people at the end of the year.

Assuming a userbase of 905m users, this means that OpenAI has a conversion rate of roughly 6%. It’s likely worse, as monthly active users should, at least in theory, be a higher number, as it captures every weekly user in addition to less-active users over the course of a month.

Nevertheless, while this represents an improvement over the 2.583% rate in February of last year, it’s likely improved as a result of cheaper ad-supported ChatGPT “Go” subscribers at $5 or $8 a month, depending on geography. OpenAI also gave away a free annual ChatGPT Go subscription to literally every Indian subscriber in late October 2025, though I cannot confirm if they’re counted in the total.

It’s like they’ve built the lines, bought the rolling stock and are now running the trains without bothering to ask the riders on the trains to actually buy a ticket. True, you get a premium service if you do buy a ticket - almost every serious coder I know of says that the difference between free and “Professional Max” grade service is enormous and that if you are complaining about the errors generated by your free account then you should really try the same with a paid account and see the difference². But even though the output is much lower quality, it isn’t clear to me that the cost of providing the free level is that much lower and there are at least 9 (and probably more like 19) freeloaders for every paying user.

Airlines, who also provide a premium service at a premium price (first/business class vs economy), do not expect the first class flyers to pay for the entire flight. Nor does JR, SNCF or any other high speed rail company. In fact the only transportation services that allow for such enormous levels of free-loading are some poorly run urban public transport services (and I note that even publicly funded services try to make people pay).

The Problem With Trains

Stop, I want to get off here!

The big problem with trains is their inflexibility. You can’t get off between stations. You can’t go to a place that doesn’t have a station.

With a bicycle you can go anywhere³.

That’s an excellent metaphor for the difference between successful AI users and AI slop makers. The slop sorts, and the vibe coders, and so on want to just recycle other people’s original work to solve their current need. They want it now and are mostly uninterested in trying to enhance anything of themselves other than their bank balance. They like the fact that the AI train is fast and they don’t really care that they are following in the steps of thousands or millions of others before them. They typically don’t exhibit the curiosity that leads the bike users to want to get off the train half way and climb up that hill there. They want to go from Tokyo to Kyoto and don’t want to stop in Shizuoka to admire Mt Fuji. But to extend the metaphor, the tourist who wants to really appreciate Japan does want to get off in Shizuoka and the same applies to the user of AI. AI slop is Tokyo to Kyoto and millions of people do that. AI being used productively to generate new insights is the stop in Shizuoka, and that’s, say, one in a million tourists.

Perhaps the biggest problem with trains though, is that they rarely get you to precisely the destination you want. They get you close (for varying values of close) but except for the few times the Tokyo Metro station is in the bottom of building where you have your meeting it is rare that you exit the station and are right where you want to be. Most of the time the train gets you, say, 95% of the way there but you have to then walk that last 5%. That’s very similar to sloppy AI usage.

Consider the picture used to illustrate this post. The prompt I gave grok was:

Draw a cartoon that has a man on a racing bicycle riding on a road beside a railway. The man and bicycle should look like they are a member of a Grok racing team. On the railway there is a train which is advertising Claude and ChatGPT.

As you can see the result is kind of there, but there are issues. Such as the fact that Team Grok’s “road” seems to be another rail line only with the middle filled in. It even has the wires overhead - and in fact there are way too many wires overhead in general and nothing supporting them. If I were trying to do this to get paid I’d have to stuff this into GIMP and do some serious clean up⁴. Fortunately for me having this kind of shoddy image helps make a point in this article.

The reason why the bicycle user is better is that that user can use the same enhancement to his effectiveness both on the flat roads that are by the train tracks and on the steep winding track from the last station up to the pass and then look down the other side to a whole new vista of possibilities that the train user can’t even see, let alone get to. AI as bicycle allows you to go further and do/see new things that potentially no one else has seen before or at least that they have never cycled to before.

Trains Make You Lazy and Weak

Riding a bike forces you to use your muscles. Riding a bike a lot makes you stronger. By contrast riding a train does not (unless it’s during rush hour and you can’t get a seat). In fact if you spend your days riding trains your muscles will atrophy. That is something that has been observed in users of AI. Instead of doing the work themselves people just ask AI to do it and never learn how to do it themselves. In fact they even forget how to do things they used to be able to do.

AI isn’t the only technical product to do this - people are forgetting how to read maps, forgetting how to navigate without a map, let alone a smartphone, and indeed forgetting how to write by hand. Plus of course there’s the whole calculator thing killing the knowledge of addition and subtraction and thus the relationships between numbers that allow you to detect whether there’s been a numerical mistake.

AI makes all this worse and it is particularly bad if you use the train approach for AI. With the bicycle approach a lot of this is mitigated because you are still checking the AI’s output and giving the AI detailed specs that will create what you want. You don’t forget how to write code or draw pictures when you use AI to do some of the fiddly bits.

It is extremely bad for you to become dependent on the train to get you from A to B because you are in big trouble when the train stops running and in moderate trouble if the train operator puts the price up. The Big AI companies are indeed putting the price up and people who are stuck on the train model are now facing a choice of paying a lot more for the AI output or quitting and going back to using much less AI or even none at all. If you built your business model on cheap AI and fired half your staff that’s a problem.

The bicycle user on the other hand can still adapt if it turns out the price of cycling on the road becomes too expensive. The muscles are still there if walking is required and perhaps he can find his own path that isn’t charged for.

Trains Cannot Be Repaired

That leads to another related problem. What happens when the equipment breaks down. If anything in the train infrastructure fails, from the train itself to the tracks, the signalling and the electricity powering it then the train is stuck. If you are on the train when it is stuck you can’t get off and you as the passenger have no ability to help fix the problem. If you are not on it then you are stuck at the station waiting for it along with (probably) dozens of others. In neither case can you get to your destination, nor can you do anything to improve the situation. Its true that trains do not break down very often, but when they do you are in big trouble.

On the other hand if your bicycle gets a puncture or the chain falls off you can fix it. Sure there are accidents that can’t be fixed - I’ve had one where I broke the front wheel - but those are rare and even that one time I could hobble the rest of the way down the hill carrying the bike. The passengers cannot carry a train.

The AI version of this works the same. If you vibe code something and your AI agent decides to delete your entire database while debugging then you are in a world of trouble. This has happened to at least two people/organizations. There are any number of other horror stories that include unmaintainable code with serious bugs that cannot be found or fixed and of course all those embarrassing legal document hallucinations.

The bicycle version is not so disastrous because the bicycle AI user never lets an agent have access to production code and never lets the output of the AI be the entire product. As a result if the AI does something bad, it’s more like the got a puncture problem where you repair the tire and carry on.

Now can you cause serious damage using your AI in bicycle mode? sure. Nothing is perfect. Just as with a real bicycle where you can fall off the side of the road and break your neck. But as with the fall off the side of the road thing, this is rare. Trains being delayed or breaking down are commonplace as are AI horror stories where people just trusted AI to do the right thing and it didn’t.

There’s one other way that AI as bicycle shines compared to the train model, that is in ROI

Does AI Produce Positive ROI

There have been a lot of posts and questions about whether the use of AI is producing a positive ROI or not and whether the massive investment in data centers for AI will ever pay off. That’s a good question and Ed Zitron, as always, makes the case that it won’t.

In this case I think he’s partly wrong, and looking at the profitability of trains is a useful way to see why some AI usage is generating positive ROI and other usage isn’t.

First there’s the question of who is seeing a positive ROI. The AI companies themselves certainly aren’t. The burning of investor dollars by Anthropic and OpenAI is well documented. Now we see google making a secondary offering to pay for more AI, which is not, IMHO, a positive sign. Google are not alone in this; Meta, Softbank, Oracle have also made financial engineering decisions that seem to indicate issues with paying for AI Data Centers as well. I think Amazon and Microsoft have too, but I don’t recall the details.

On the other hand some users certainly are. I’ve mentioned one man bands like ES Raymond before. This substack linked below talks about a number of successes and failures including the author’s own ~70 person company that is finding plenty of success with its AI usage.

Nobody Won the Token Race

Away from the coding space, British comedian Dominic Frisby has found AI extremely useful so I think he counts as getting a positive ROI

How AI Became My Production Company

I am quite sure there are plenty of other examples. However the key thing that Denis notes in his article is how they use AI. And it is absolutely the bicycle model:

We never made using AI the point, and we use it constantly. PMs run tickets through it against our templates and acceptance criteria. QA runs bug reports through it so they’re clear enough for anyone to act on. Marketing crawls the web with it for angles. Product builds per-client RAG out of meeting notes, docs, and history. Engineering, obviously. Every department vibe-codes its own internal tooling, and then the CTO rewrites the worst of it like a human being. The usage is enormous. It just isn’t stupid. The bill stays flat not because we use AI less, but because every run has a task attached, and a task has an edge. We never tried to take the human out of the loop. The AI is a tool the person reaches for, not a replacement we’re proving out, so nobody is burning tokens to hit a number or make a headcount go away. The work still belongs to a person. The bill is just what the tool cost them.

He also points to Uber and the description of that company’s AI use definitely sounds like train model. In fact it sounds like they decided to splurge on first class round trip fares with the goal being the journey rather than the destination. Unsurprisingly ROI on that is non-existent.

But to go back to trains as infrastructure. Very few railways are currently profitable. When you include the capital costs of building them, AIUI, the only high speed rail line to be profitable and to have paid back all its construction debt is the first Japanese Shinkansen between Tokyo and Osaka. Many of the lines built in the 19th century went bust. Many 20th century urban rail projects have never made any money. They have often provided benefits like reducing traffic that make them a positive for society but in cold hard cash terms the lines themselves, even if they are cash flow positive at an operational level can never repay the capital cost to build them. Even the various private railways in Tokyo and Osaka made their money on land sales, the department stores over the stations and so on.

It is hard to see where the ancillary benefits of AI come from and while it is possible to see how AI can raise prices to cover opex (just), if everything gets built out that is promised we are looking at close to $1T⁵ in capex that needs to make a return. If your business processes are based, train like, on specific “big AI” capabilities, you are either going to have to pay a lot for them or handle what happens when you AI provider goes bust (or possibly both).

A bicycle AI user on the other hand has options. We aren’t quite there yet, but it looks like capable private AI, running either locally in your own office / data center or on hardware in Azure or AWS, is going to be a standard thing fairly soon. xAI says it will be releasing it’s current model shortly after it migrates to a new one. Deep Seek and others have fairly good 70B models apparently that are capable in limited use cases. Running these models locally is quite possible. It is very clear from reading posts like this (don’t bother if not technical) that a bit of research into optimization could allow such models to run acceptably on hardware that seems to be well below the expected spec.

If there’s a bust in AI data centers then private AI becomes extremely affordable as many of the servers and chips become widely available at fire sale prices. Ed Zitron and others have pointed out that an individual company may not be able to use a set of latest Nvidia Vera Rubin systems, but a private equity firm, say, that buys a data center that has gone bust for pennies on the dollar (or takes the assets as a creditor in bankruptcy proceedings) could absolutely rent out racks to individual (medium to large) organizations. We don’t get the new improved models in this scenario, but my guess is that existing models are good enough for most users to get a positive ROI on as long as they use AI like a bicycle. Moreover there is considerable evidence that many of the recent improvements in AI results come from tweaking the inputs and improving the tools that use them rather than the AI itself. The question of whether you get better results from Opus 4.7 vs 4.5 or whatever seems distinctly unclear and in fact many people report better results with slightly older models. Or if not better then good enough and cheaper because the older models use fewer tokens per result and thus, as we move to a pay by token business model, the difference in cost becomes important.

The Train AI model gets you AI slop. Copycat apps, derivative images, copyright infringing music and irritatingly written documents that are eminently forgettable and which only generate minimal revenue. It doesn’t solve big problems or massively help with much of anything, in fact it contributes to to the general enshittification of the internet. People like Lucy Pepper write great posts about how AI is bad for artists based on this, and she (and her fellow anti-AI people) are right. It is bad and harmful for everyone. But I think they may be partly missing the point. AI can produce beautifully illustrated Dad jokes that no one would ever spend 5 minutes actually creating if they had to do them by hand (though there are absolutely AI slop).

AI, when used like a bicycle, as a tool to enhance human intellectual musclepower, on the other hand allows people like Dominic Frisby to entertain us all at a price/on a budget that would be impossible without AI. Let AI handle the tedious boring bits and let humans handle the challenging parts - something that IKEA has figured out (via the Denis Stetskov link above). That leads to greater job satisfaction and is therefore a win for IKEA’s customers, its employees and its shareholders.

I think a lot of the problem comes from mis-defining the problem. To take a quote from Rory Sutherland’s Alchemy (via Brad DeLong)

First you define a hotel doorman’s role as ‘opening the door’, then you replace his role with an automatic door-opening mechanism. The problem arises because opening the door is only the notional role of a doorman; his other, less definable sources of value lie in a multiplicity of other functions, in addition to door-opening: taxi-hailing, security, vagrant discouragement, customer recognition, as well as in signalling the status of the hotel. The doorman may actually increase what you can charge for a night’s stay…

As Brad says AI is the automated door opener but it can not replace the hotel doorman. Sometimes the door opener is sufficient. Often, however, you need the extra human part.

1. In the spirit of full disclosure $dayjob does have a paid subscription to more than one service and I have used them for some $dayjob work.

2. e.g https://x.com/esrtweet/status/2060159070674244091

3. OK maybe a mountain bike and even then not absolutely everywhere.

4. well you’d expect someone to do that, but given what we see of AI slop on the internet apparently most people don’t even bother to fix it up at all.

5. Possibly up to 10x more if you do the sums and believe Jensen Huang’s costing for gigwatt data centers ($80-100B/GW)