AI assistant spent 54 seconds thinking and returned no answer to a simple question
On 6 July 2026, a user asked the Merlin assistant a one-line question about what a structured intelligence assistant is. The model used its entire 2,048-token thinking budget without closing its reasoning block, leaving the answer section empty. After 54 seconds, the user received a typed error instead of a response. The developer responded by disabling thinking on the direct-answer path and adding a fallback that salvages answer-like text from unfinished reasoning, tracked through a metric. The developer argues that systems should explicitly define what happens when a reasoning budget runs out mid-thought.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in