There is a peculiar experience that comes from using a modern AI system intensively.
You are given something that looks like a Ferrari.
The engine is astonishing. It can reason, write, analyse, code, search, organise information and sometimes produce connections that would have taken a human researcher hours or days.
Then you press the accelerator.
And discover that somebody has put a wooden log into the exhaust.
You look ahead.
There is only one narrow lane.
You cannot change lanes. You cannot see much beyond the lane you are currently in. There are limited exits. There is no convenient U-turn.
And just when you finally find an interesting side road, somebody says:
“That road requires another subscription.”
1. The Ferrari analogy began with fuel
My original analogy was relatively simple.
Imagine that ChatGPT is a Ferrari.
The model is the engine. Computational resources are fuel. The user is the driver.
The better the model and the more resources available to it, the more work it can potentially perform.
It is tempting to think that moving from a lower-cost plan to a higher-cost plan should simply mean:
More money → more fuel → more work.
But intensive use reveals something important.
Fuel is only one bottleneck.
Suppose I receive considerably more fuel but the Ferrari still has to travel on the same road, through the same traffic, with the same restrictions. I have bought more potential performance. I have not necessarily bought proportionally more useful output.
2. Then I discovered the wooden log
Imagine paying for a Ferrari and filling its tank. The engine starts. Everything sounds magnificent. But there is a wooden log stuck in the exhaust.
The Ferrari can theoretically produce enormous power. It just cannot efficiently turn that power into forward motion.
This is analogous to a recurring phenomenon in AI use:
Capability is not the same thing as accessible capability.
A model may be capable of reasoning about an article. It may be capable of generating the article. It may even understand exactly what should happen next.
But the complete workflow may still require the human to copy information, move between applications, open websites, perform actions, verify results, correct errors, upload files, download outputs, publish material and reconnect the result to the research archive.
The AI has the engine.
The user becomes the transmission system.
3. Then came the highway problem
The analogy became more interesting when I stopped looking at the engine and started looking at the road.
The current AI interface can resemble a single-lane highway.
You are travelling in one conversation. You can go forward. But your ability to move sideways between different intellectual lanes is limited.
- Lane A: P vs NP research
- Lane B: educational research
- Lane C: articles and publishing
- Lane D: AI experiments
- Lane E: personal archive and documentation
A thought discovered in Lane D may suddenly illuminate Lane A. An educational concept from Lane B may provide a useful framework for Lane C. An old discussion in Lane A may become relevant to a new hypothesis in Lane D.
The intellectual graph is not linear.
It is a network.
4. What I actually want is not a faster Ferrari
This led to a surprisingly simple conclusion.
I do not necessarily need the fastest Ferrari.
I want a much better highway.
- multiple lanes;
- lane changing;
- interchanges;
- side roads;
- U-turns;
- visibility across the network;
- easy movement between projects;
- direct access to files;
- direct access to research;
- direct access to publishing;
- the ability to carry context from one road to another.
That could produce substantially more useful work than giving me a much faster Ferrari trapped on the same narrow road.
This is an important distinction.
Model capability is only one component of productivity.
A useful approximation is:
Useful AI output ≈ Capability × Context × Interface × Execution × Reliability
If any one of those factors becomes a serious bottleneck, increasing another factor produces diminishing returns.
5. The strangest bottleneck: analysis without complete execution
My particular use of AI makes this problem more visible.
I am not primarily asking: “What is the answer to this question?”
I am trying to use AI as something closer to an intellectual laboratory.
- execute research;
- analyse my analysis;
- identify recurring patterns in my thinking;
- connect apparently unrelated discoveries;
- organise accumulated discussion;
- maintain an evolving research structure;
- turn discoveries into articles;
- archive those articles;
- retrieve them later;
- continue the investigation from where it stopped.
That is a much more demanding requirement than answering isolated questions.
It requires something closer to an intellectual operating system.
6. Context is not the same as memory
An AI system can have memory without having free intellectual mobility.
Knowing that something happened previously is not sufficient.
Suppose I discovered an important idea three weeks ago. What I really want is not merely: “Yes, I remember that.”
I want: “That earlier idea is relevant to what you are doing right now. I found it and brought it into the current analysis.”
That requires more than memory.
Retrieval + relevance detection + cross-context reasoning + execution.
In highway language: memory tells the Ferrari that another road exists. A proper research architecture gives it an interchange.
7. External websites create another layer of bottlenecks
Then comes the outside world.
The AI may know what needs to happen. But knowing and doing are different things.
An article may need to be published. A website may need editing. A document may need uploading. A repository may need updating. A page may need checking.
An external service may require authentication. Another may impose an API restriction. A third may require payment.
Suddenly our Ferrari reaches the edge of the highway.
There is a beautiful road visible on the other side. The Ferrari has the horsepower to reach it.
But there is a toll gate. Or an API. Or an authentication barrier. Or an integration that doesn’t exist.
Or another software company demanding its own pie.
The AI says: “Here is exactly what you should do.”
The human says: “Excellent. Now do it.”
And the AI replies, metaphorically:
“I cannot reach that road.”
8. This is why more tokens are not the complete answer
Suppose Plan A gives me one unit of fuel and Plan B gives me five units.
That sounds like Plan B should be five times more useful.
But suppose both plans use essentially the same constrained highway.
If my real bottleneck is interface, context mobility, external access, execution or reliability, then additional fuel has diminishing returns.
I may spend more money without proportionally increasing completed work.
This is why capability-per-rupee and useful-work-per-rupee are different metrics.
9. Go versus Pro becomes a different question
The obvious question is: “Which plan is more powerful?”
The more useful question for my particular workload is:
Which plan gives me the greatest amount of useful intellectual work per rupee?
If a higher tier gives me substantially more fuel but leaves most of the major road bottlenecks unchanged, its marginal value may be much smaller than expected.
A lower tier that provides sufficient model capability, adequate usage and the tools I actually need may therefore produce better value for money.
This does not mean the lower tier is objectively more capable. It means the system’s bottleneck has moved somewhere else.
10. The deeper lesson: AI progress needs infrastructure progress
The AI industry naturally focuses on the engine: bigger models, better reasoning, more compute, more context, more intelligence.
All of that matters.
But eventually another question becomes equally important:
What can the intelligence actually do?
A brilliant scientist locked inside a room is still a brilliant scientist. But the scientist’s productivity is constrained by the room.
Give the scientist a laboratory, a library, instruments, assistants, communication and access to the outside world, and the same intelligence suddenly becomes much more productive.
AI systems are approaching a similar transition.
The next major productivity leap may not come entirely from making the engine smarter.
It may come from removing the walls around the engine.
11. The ultimate AI highway
The ideal system I am imagining would connect:
- Conversation
- Research memory
- Projects
- Files and archives
- Web
- Code and computation
- External applications
- Publishing
- Long-term research history
And the AI would be able to move between these layers naturally.
Not:
“I know what you should do, but you must now leave this conversation and perform six manual operations.”
Instead:
Think → retrieve → analyse → execute → verify → archive → continue.
That is a genuine research environment.
That is the highway.
12. The Ferrari joke has therefore become an engineering model
The joke began as:
“Sam gave me a Ferrari but I need to pay for the fuel.”
Then it became:
“The Ferrari has a wooden log stuck in the exhaust.”
Then:
“The Ferrari is driving on a one-lane highway.”
And finally:
“Sam owns both the Ferrari and the highway.”
That last version is actually the most important.
Because the company controlling the AI model also controls much of the interface through which users experience that intelligence.
Therefore the limiting factor does not necessarily have to be the model itself.
It can be the architecture surrounding the model.
The next AI revolution may not require a smarter Ferrari.
It may require removing the wooden log, widening the highway, adding lanes, building interchanges—and finally allowing the Ferrari to leave the highway when the destination is somewhere else.
Until then, we may continue admiring the horsepower while wondering why the journey takes so long.
And occasionally asking the driver:
“Why the hell are we still in first gear?”