Home › Hard problems › Inside the model

Inside the model: the core of LLMs

Thinking cost: models overthink easy questions

Solo: strongOne person can run a real test of this in about 8 weeks.

"What is 2 + 2?"an easy questionthousands of hidden thinking tokens, all paid for"4"needed about tena littlea lot?how much should it think?nobody has a reliable dial for that
Reasoning models spend thousands of hidden tokens on questions that need ten, and nobody can reliably tell in advance how much thinking a question needs.

Why it is still open

Reasoning models spend thousands of hidden tokens on questions that need ten. Nobody has a reliable way to decide how much thinking a question needs.

An experiment one person can run in 8 weeks

Route 1,000 questions to "think a little" or "think a lot" using a cheap classifier. Measure accuracy kept against tokens saved.

Verdict: Solo: strong

Cheap, measurable, and it saves money for every buyer.

The solo test behind the verdict

  1. Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
  2. Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
  3. Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
  4. Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.

Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.

PreviousReliable multi-step reasoningNextContinual learning: learn after deployment without forgetting

All 79 hard problems · Working on this, or know of a result we missed? basant@cleandataengineering.ai