A 2011 study showed that we prefer slower websites to faster ones if the slow sites appear to be “working harder”. Think about this airfare-search website that takes a minute to show you the result. Of course, it must have found better deals than that site that showed you results in a split second. Or maybe not?
Are we guilty of the same bias when working with LLMs? The one that takes longer, with “extended reasoning” on, must be better, right? Turns out it’s not always the case!
Two recent papers suggest that up to 80% of the “thinking” your AI does on routine tasks may be decorative. The model has already decided. It just keeps spinning. In the episode, I share what happened when I ran the same analysis on my personal CRM with cheap models versus premium ones with full extended thinking. I also get into why bragging about token consumption is like bragging about fuel burn, and why this might be the best news local AI models have had in months.
Your challenge: run one of your regular tasks with reasoning off. Compare. You might find your own spin cycle.
Stay curious!