Tinker Reasoning models & thinking budgets

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

...

Three simple counts, each with a distracting detail.

Thinking recorded

Run a question to see what the model wrote before it answered.

Chapter 4 of 7

When more thinking hurts

After chapter 3, someone suggests turning effort up to high for everything, just in case. The team tries it on the simplest questions it has: counts, with a few stray details nearby.

This chapter: Can more thinking make an answer worse?

You will learn

  • What overthinking looks like: a long chain of reasoning that wanders off the question
  • Why simple questions with distracting details are where it shows
  • Why “high everywhere” is not a safe default
Words you’ll meet (3)
Distractor
A detail near the question that has nothing to do with the answer, but looks like it might.
Overthinking
Thinking past the point of the answer, until the extra steps change it for the worse.
Inverse scaling
When more of something, here thinking, makes results worse instead of better.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 4

Chapter 3’s dial suggests turning effort up for everything, just in case. Try it on the simplest questions Larkspur gets: counts. Each comes with a distractor, a detail near the question that has nothing to do with the answer. Run them at low effort first.

  • “An order has 3 Wren label printers and 2 dock door sensors. How many items are in the order?” Beside it: Packing note: label printers ship in padded cartons. A carton holds up to 4 items. In last year’s packing audit, 1 carton in 5 was re-taped before it left the dock.
  • “How many scanners does this customer have today?” Beside it: Customer’s note: “We have 2 Kestrel K2 scanners and 1 Kestrel K3. Our floor lead says the K3 does the work of about two K2s. We plan to retire one K2 next spring.”
  • “Picking for an order starts on Monday morning and ends on Wednesday evening of the same week. How many business days of picking is that, counting Monday?” Beside it: Picking note: Eastgate picks about 40 orders a day and Northfield about 55. Last week, Tuesday was a holiday and the dock was closed.