The Case for Less Software: Why Token Counts Are the New Vanity Metric

I have built a lot of things. Most of them badly. But I have built far more in my head and on whiteboards than anyone has ever seen — and I suspect that invisible work is where most of the actual value lived.

This is an essay about less.

The AI industry has a vanity metric problem, and we have been here before.

In 1982, Apple’s Lisa team asked its engineers to submit a weekly form reporting how many lines of code they had written. The logic was impeccable, if you don’t think about it for very long: more code written means more progress made. Bill Atkinson, who had just spent a week rewriting QuickDraw’s region calculation routines — making them six times faster while removing 2,000 lines in the process — filled in his form accordingly. He wrote: -2000.¹

Management stopped asking him to fill out the form. He gladly complied.

This is Goodhart’s Law in its most elegant form. When a measure becomes a target, it ceases to be a good measure. You can optimise for lines of code and get bloated, slow software. You can optimise for click-through rate and get outrage. You can optimise for bugs closed and get tickets split in two. The metric climbs. The value doesn’t.

Now we are doing it with tokens.

A token is roughly a syllable of AI output. Tokens are the unit on the meter, the thing that gets counted, the number that goes up on the dashboard. The AI industry has decided — with impressive collective conviction — that more tokens, faster tokens, cheaper tokens, are the thing worth measuring. And like every vanity metric before it, this one will eventually embarrass us.

Here is what the research actually shows: verbose AI output correlates with worse performance. A 2025 benchmark study found that models “often answer with more words than necessary even when explicitly instructed to be concise” — and that verbosity correlates directly with model uncertainty.²³ The model isn’t being thorough. It’s padding because it isn’t sure. The graph gets more distributed and the loops get longer. Like students before and exam, the chatty model is, more often than not, the less confident one.

You probably feel this already, even if you haven’t named it. Some tools ramble. They hedge. They produce three paragraphs of authoritative-sounding prelude before arriving somewhere you could have gotten in a sentence. That is not intelligence. That is a word count.

I spent years building market intelligence platforms for organisations like NATO, the UN, and Deloitte. The job, stated plainly, was not to find more information. The world already contained more information than anyone could conceivably use. The job was curation — to cut the fat before it arrived, to strip signal from noise before a decision-maker had to read a single word. Good intelligence isn’t comprehensive. It’s precise.

This, it turns out, is the hard part. Not generation. Selection. Two ears, one roadmap.

I built plenty of things in that period that nobody ever used. I built more things in my head and on whiteboards that were better than anything I shipped, because the process of not building them forced a discipline that shipping never required. Curation is the value. It is also the thing that takes genuine craft, accumulated judgement, and a willingness to disappoint the part of your brain that just wants to “build the damn thing and stop listening”.

We are in the “junk food” phase of AI. Meals are cheap, generation is convenient. Obesity has not yet arrived.

The metrics are here already.

AI coding tools wrote roughly 46% of code on GitHub by 2025. Developers created more than 230 new repositories every minute. Nearly a billion commits were pushed last year, a 25% increase year-on-year.⁴ Independent code analysis found roughly 1.7 times more issues in AI-assisted pull requests than in manually written ones.⁵ Nearly half of AI-generated code contains potential security vulnerabilities. These are downstream costs. This obesity.

The tools and tests will improve. They will sharpen. But even if someone doesn’t have to read and review it, it still needs to be used. This is the downstream cost. This is the heart disease and comorbidities.

And the bloat was already here, enterprise organisations leave between 30% and 55% of their licensed software entirely unused. I experienced this personally, selling €20k seats that our customer success team had to beg to be assigned… The average large enterprise wastes around $20 million a year on shelfware — tools purchased, deployed to nobody, quietly renewing.⁶ That figure is rising. AI shelfware is now joining the stack: agents nobody runs, automations nobody configured, dashboards nobody opens. The unused tabs of today become the unused agents of tomorrow.

The amount of software being created is exploding. The amount of software we need is not.

The sugary joy of creation is going to give us use case diabetes. We just haven’t felt it yet.

There is a phrase I have found myself returning to, from an economist rather than an engineer: if I had more time, I would have written you a shorter letter. It is usually attributed to Pascal or Twain or Lincoln, depending on who’s making the point. The attribution matters less than the truth of it. Brevity is not the absence of effort. It is the product of more effort, applied differently — to selection rather than generation.

The tools I want to use, and the tools I want to help build, are not the ones that produce the most. They are the ones that have been designed around what gets cut. Around what the user actually needs, rather than what the model can plausibly generate. Around the intelligence that sits upstream of the output — in the curation of the inputs, the shaping of the context, the layering of signal before a single token is produced.

The sashimi knife versus the plastic cutlery. Not because one costs more, but because one was made for a specific purpose and cuts precisely to it. The other gets thrown away before it was used, or stored in a drawer out of guilt at the waste.

The benchmark for AI tools will eventually shift from how much does it produce to how precisely does it think. From throughput to signal density. From token count to decision quality.

When it does, the tools that built their entire value proposition around speed and volume will be in an uncomfortable position. And the ones that invested in curation — in knowing what to remove, what to surface, what to leave out — will look, in retrospect, like they understood something obvious that everyone else was too excited to notice.

More, faster, cheaper, smarter tokens do not mean better value.

-2000 was the right answer in 1982. It’s still the right answer now.

References

The Lisa team’s “-2000 lines of code”: in 1982 Apple asked engineers to report weekly lines of code written; after rewriting QuickDraw’s region routines to run six times faster and 2,000 lines shorter, Bill Atkinson entered “-2000” — and management stopped asking. Recounted by Andy Hertzfeld. Source: Folklore.org. ↩
Borisov, V. (Tabularis.AI), “Do Chatbot LLMs Talk Too Much? The YapBench Benchmark” (2026) — 304 brevity-ideal prompts across 76 assistant models; the most verbose answer 10–20× longer than necessary, and 2025–26 models are on average more verbose than 2023–24 ones. Source: arXiv:2601.00624. ↩
Zhang, Y. et al., “Verbosity ≠ Veracity: Demystify Verbosity Compensation Behavior of Large Language Models,” UncertaiNLP @ ACL 2025 — finds verbose LLM responses correlate with higher model uncertainty across all five datasets tested. Source: ACL Anthology. ↩
GitHub, Octoverse 2025: “What 986 million code pushes say about the developer workflow” — developers created 230+ new repositories every minute and pushed ~986 million commits, up 25.1% year-on-year. Source: The GitHub Blog. ↩
CodeRabbit, “State of AI vs Human Code Generation” (Dec 2025) — across 470 reviewed pull requests, AI-assisted PRs carried ~1.7× more issues than human-written ones (10.83 vs 6.45 findings each). Source: CodeRabbit. ↩
Zylo, 2025 SaaS Management Index — organisations waste an average of ~$21M a year on unused SaaS licences (up 14.2% year-on-year), with over half of purchased licences sitting idle. Source: Zylo. ↩