Quantum Howl

Innovating with less: what the evidence says about constraints

  • 5 min read
  • 918 words

Scarcity does not always sharpen ingenuity. What research says about constraints and innovation, and what it means for AI that runs inside an organisation.

Innovating with less: what the evidence says about constraints

People often say that scarcity sharpens ingenuity. We have said it ourselves: building AI that has to fit on an organisation’s own server forces you to think before you build. We went to see what the research says, and the answer is more nuanced than the slogan.

They help, up to a point

The most thorough review of the subject looked at 145 empirical studies across several disciplines. It concludes that the relationship between constraints and innovation is an inverted U: some constraint helps, too much hurts (Acar, Tarakci and van Knippenberg, 2019).

The authors separate three kinds of constraint: input constraints, such as money or time; process constraints, which fix how the work is done; and output constraints, which fix what the result has to meet. How they are experienced matters too. The same limit can be read as a challenge or as control imposed from outside, and the effect is not the same.

The mechanism has been described in experiments. When scarcity is made salient, people stop using things only for their intended purpose and come up with more novel solutions (Mehta and Zhu, 2016). These are experiments with consumers, not companies, and should not be stretched.

None of this says that having less is better. It says there is a range in which a limit helps, and beyond that range it gets in the way.

In AI, constraints push towards specialisation

Here the evidence is more direct. A study that measured the energy use of dozens of models concludes that general-purpose ones cost orders of magnitude more than task-specific ones. A thousand text classifications use 0.002 kWh on average; a thousand generated texts, 0.047 kWh; a thousand generated images, 2.907 kWh (Luccioni, Jernite and Strubell, 2024).

And small models already go a long way:

  • Fewer parameters for the same result. The smallest model able to score above 60% on a reference benchmark had 540 billion parameters in 2022 and 3.8 billion in 2024: 142 times fewer (Stanford HAI, AI Index 2025).
  • Distillation works. A reduced version of a well-known model keeps 97% of its language understanding while being 40% smaller and 60% faster (Sanh et al., 2019).
  • So does chaining models. Starting with the cheapest and moving up only when needed can match the best model at up to 98% lower cost (Chen, Zaharia and Zou, 2023).

One caveat: the same Stanford report notes that the cost of using a model of a given level fell more than 280-fold in two years. That makes small and local cheaper, and makes the cloud cheaper just the same.

What you pay for running locally

Telling only half the story would be dishonest. Running AI inside the organisation has two measured costs.

  • You run behind. Open models that fit on your own hardware match the most advanced ones with a lag of 6 to 12 months, depending on the benchmark. The authors warn that the real lag may be longer, because small models tend to be tuned for the benchmarks they are measured on (Epoch AI, 2025).
  • You use more energy per query. With the same model, local hardware delivers at least 1.4 times less per watt than a data centre (Saad-Falcon et al., 2025). Cloud at scale is very efficient: Google measures 0.24 Wh for a median text query, 33 times less than a year earlier, according to its own figures (Elsworth et al., 2025).

So the argument that local saves energy does not hold. We found no independent comparison that supports it.

The one that does hold is different. That same Stanford study tested more than twenty local models on one million real queries, and they answered 88.7% of them correctly. Two years earlier they covered 23.2%. For most everyday work, what fits in-house is already enough.

Our experience

This last part is not a finding from the literature. It is what we have lived through, and we present it as such.

When the model has to fit on a specific machine, you cannot solve a problem by adding capacity. You have to choose the task, compare models against a test defined in advance, and measure the answer. We wrote about it when choosing a model for dental imaging, when tuning local response time and when explaining why our dental AI lives in the clinic. Those three articles are in Spanish.

We have no study showing that working this way produces better systems. We have our own, and the observation that almost every decision we were later glad of came from an awkward question: does this fit?

In short

  • Constraint does not replace budget, and with too little you do not innovate either.
  • In AI, running locally costs between six months and a year of lag and somewhat more energy per query.
  • In return, it forces you to specialise and to measure, and the models that fit already handle the large majority of real queries.

Sources

Get every new publication

New chapters and reports, by email. No advertising.