The Travel Time Collapsed

Something changed in the last two years, and most of the commentary has aimed at the wrong variable.

The claim you hear is that AI makes building faster. Maybe. The measured evidence is thinner than the marketing. But faster building is not the interesting consequence even if it's true. The interesting consequence is that the cost of going somewhere else fell.

Those are different quantities, and only one of them changes what you should do next.

A model from behavioral ecology makes the distinction precise. Eric Charnov published it in 1976, in a six-page paper about animals eating berries.¹ It is the sharpest available tool for reasoning about when to quit a project, and almost nobody in software has picked it up.

The relevant question was never whether AI makes you faster inside a project. It's what AI did to the distance between projects.

The rule

An animal feeds in a patch. Food is plentiful at first, then gradually becomes harder to find. Each additional minute produces less food than the minute before.

Eventually, the animal should leave, but leaving has a cost. The next patch is somewhere else, and reaching it takes time.

Let G(t) be total value obtained after spending time t in a patch. Let τ be the travel time to the next one.

The Marginal Value Theorem (MVT) tells us where the tipping point lies. The optimal time to leave, t*, occurs when:

In Charnov's own framing: leave when the rate at which the current patch is still paying you falls to the average rate you could earn by leaving and starting again elsewhere.²

Note what the rule does not say. It doesn't say leave when the patch stops producing.

Important

It says leave when the next unit of value from staying is worth less than your opportunity cost of remaining — where the opportunity cost already has the price of leaving baked into it.

Why τ is the variable that moved

The comparative statics are the whole argument. Hold everything else constant and increase τ: optimal residence time rises. Foragers stay longer when patches are farther apart.³ Decrease τ: they should leave sooner.

Now apply it.

An MVP that took six months to stand up implied a large τ. That large τ was itself the justification for grinding on a mediocre project. Not loyalty. Not sunk cost. Arithmetic. When the trip to the next opportunity costs half a year of runway, you tolerate a lot of thinning returns before you make it.

Agentic tooling attacks that number directly. Scaffolding, integration, boilerplate, the first working version — the parts of a new project that were pure travel cost — compress.

Be careful about how much you claim here, because the evidence is genuinely mixed. METR's 2025 randomized trial put sixteen experienced open-source developers on 246 tasks in repositories they'd worked in for about five years. Allowing AI made them 19% slower, while they estimated it had made them 20% faster.⁴ The follow-up in early 2026 found a speedup for returning participants and a smaller one for new recruits, with confidence intervals wide enough to drive a truck through.⁵

A large τ was never a reason to be loyal. It was a reason to be patient. Those look identical from the outside and behave differently when the number changes.

Read those results carefully and they don't contradict the argument — they locate it. The 2025 trial measured experts doing incremental work in mature codebases they knew intimately. That is G′(t) deep inside a well-exploited patch. It is the worst case for AI assistance and the least relevant to the claim. The claim is about τ: the cost of standing up something that does not exist yet, in a domain where nobody has five years of context to leverage.

Nobody has run that randomized controlled trial (RCT). Until someone does, the honest position is that τ has fallen for greenfield work and that the size of the fall is unmeasured.

The assumption nobody states

MVT requires depleting patches. G′(t) must decline. Charnov's animals are eating the berries; there are fewer berries.

Plenty of things you might build do not behave that way.

Network effects, accumulating domain expertise, brand, a distribution channel, a codebase whose next feature is cheaper than the last because the abstractions finally settled — these have marginal returns that rise for years before they turn. Run MVT on an appreciating asset and the model tells you to abandon it, confidently, on a false premise.

This is where the foraging metaphor most often gets abused in business writing. The founder who looks like they're irrationally overstaying may be sitting on a patch that isn't depleting. You cannot tell from the outside, and frequently they can't tell from the inside either.

Important

So the model's first demand is diagnostic, not prescriptive: establish that returns are actually thinning before you apply a rule that assumes it. A flat quarter is not depletion. A flat quarter following three flat quarters, in a system with no compounding mechanism you can name, probably is.

Sunk cost, and the thing that isn't sunk cost

Marginal Value Theorem is ruthlessly forward-looking. What you've already spent to reach the patch is not in the equation. Neither is what you've already pulled out of it.

The standard rendering of this is that history is irrelevant. That rendering is half wrong, and the half that's wrong matters.

Past extraction is irrelevant as a cost. It is excellent evidence. Allan Oaten's 1977 extension to Charnov made this formal: when patch quality is uncertain, the yield you've observed so far updates your estimate of what remains, and the optimal policy becomes a stopping rule conditioned on what you've seen.⁶ Green called the resulting foragers Bayesian, which is exactly what they are.⁷

The dollars you spent don't tell you whether to stay. What those dollars bought tells you a great deal.

Two years of weak growth is not a reason to stay because you've invested two years. It is a reason to leave because two years is a large sample.

The failure mode isn't remembering your history. It's mistaking the emotional weight of history for information content.

The edge expires

Here is the part of the argument that should make you uncomfortable if you're planning to act on it.

τ didn't fall for you. It fell for everyone with a terminal.

Fretwell and Lucas worked out the consequence in 1970, six years before Charnov.⁸ When foragers move freely toward the best available patch, they redistribute until the achievable payoff equalizes. Rich patches attract entrants until they aren't rich anymore. The individually rational move, executed by everyone, dissipates the return that motivated it.

Cheap exploration is an advantage while it's asymmetric. Once the tooling is ambient — and it is becoming ambient very quickly — the low-τ world is one where every viable patch gets found faster, entered by more people, and depleted sooner. Lower travel costs don't just make it easier for you to leave.

They make it easier for others to arrive.

Which inverts the strategic conclusion for a specific class of opportunity. If τ is collapsing toward zero everywhere, the durable value shifts to whatever cannot be reached quickly regardless of tooling: regulatory position, proprietary data, distribution, trust, a decade of domain scar tissue. Those are patches with a high τ that nobody's agent can compress.

In an environment where everything cheap gets crowded, the expensive trip becomes the defensible one.

What to actually do with this

Four questions, in order, when you're deciding whether to keep going:

1. Is this patch depleting, or compounding? Name the mechanism either way. If you can't name a compounding mechanism, assume depletion.
2. What is my current τ, honestly measured? Not what it was in 2022. Time yourself standing up the last thing you started from zero.
3. What did the last six months of yield tell me about the next six? Treat it as a sample, not as a debt.
4. Is the patch I'd move to cheap for me to reach because it's cheap for everyone to reach? If yes, price in the crowd before you travel.

Charnov's animals solve this without doing any arithmetic, which is the part that should be humbling. Selection did the optimization; the bird just leaves. Humans, holding the closed-form solution, systematically overstay — the overharvesting bias is one of the most reproducible findings in the experimental literature, in rats and in people, and it survives even when subjects have accurate information about both the patch and the environment.⁹

You will overstay too. The equation won't stop you. What it will do is tell you, precisely, which number you're getting wrong.