There was no announcement, no dramatic turning point. And yet, within a few months, a silent divide has settled into the technical professions: between those who have woven AI assistants into their daily work, and everyone else. And the gap isn’t closing — it’s widening.

A difference in speed that becomes a difference in level

At first, AI was just a time-saver: a snippet of code suggested here, a bug spotted faster there. But that saving compounds. An engineer who delivers twice as fast doesn’t simply produce more: they learn faster (they test more ideas), they take on topics they would have avoided, they automate their repetitive tasks.

Meanwhile, the one who hasn’t adopted these tools keeps working at the old pace. After a few months, the difference is no longer about speed — it’s a difference in scope and in accumulated experience. The gap widens because the lead produces more lead.

It’s not (only) a question of access

You might think it all comes down to access to the tools. It doesn’t: coding AI is widely available, often free to get started. The real fault line lies elsewhere — in the skill of using it:

• knowing how to frame a precise request (the right context, the right constraints);
• knowing how to verify what the AI produces, because it sometimes gets things wrong with great confidence (the famous hallucinations);
• knowing when to trust it and when to take back control.

This skill can’t be bought: it has to be practised. And those who practise every day pull away from those who watch from a distance.

The overlooked factor: the token quota

There is, however, a very real and material barrier that gets little attention: the token quota.

AI models are billed by usage, in tokens — the unit that measures the text (and code) consumed and produced. But being effective with AI means iterating: rephrasing, starting over, exploring several avenues, having it re-read long files. Every round trip consumes tokens.

The consequence: whoever has a generous token budget (an advanced subscription, an enterprise account) can iterate freely, explore, and delegate large tasks. Whoever is limited by a tight quota censors themselves: they hesitate to try again, shorten their requests, give up sooner. At equal skill, the token budget becomes a multiplier.

This is no minor detail. It creates a divide within the divide: not only between those who use AI and the rest, but between those who can use it intensively and those who ration it.

The traps of being ahead

It’s not all rosy on the “fast” side. Excessive dependence has its downsides:

• Atrophy of the fundamentals: delegating without understanding ends up making you fragile.
• False confidence: generated code that “looks correct” can hide a flaw.
• Homogenisation: if everyone asks the same model the same thing, the solutions start to look alike.

The best aren’t those who use AI the most, but those who keep their judgement: they use it as a co-pilot, not as autopilot.

Closing the gap, concretely

The good news: this gap can be crossed. It doesn’t depend on a rare talent, but on deliberate practice.

1. Start now, on small real tasks (reviewing, explaining, debugging).
2. Build the verification reflex from the outset — understand, don’t copy.
3. Learn to prompt: context, constraints, examples (see the AI that codes).
4. Think about your token budget: iterate where it matters, don’t waste it on the trivial.
5. Keep the fundamentals: AI amplifies a skill, it doesn’t replace it.

In one sentence

The gap isn’t widening between those “for” and those “against” AI, but between those who learn to use it with discernment — and with the means to iterate — and those who wait. The first step is the most important: taking it today costs far less than catching up tomorrow.