In this video I break down a pretty uncomfortable result from Google: with every extra dollar you spend on “reasoning” or long chain-of-thought tokens, you can actually make the model’s answers worse. And this isn’t my opinion—Google shows a negative correlation between answer length (token count) and accuracy across multiple models and scientific/math tasks. So the old mental model of “longer thinking = smarter” just doesn’t hold the same way for recent systems, because overthinking can amplify flawed heuristics and waste compute on uninformative generation. Then I switch the framing from “how long” to “how deep.” I explain what Google calls a deep thinking token: a token that needed more internal work across layers before the model’s prediction stabilized. They quantify this by comparing predictive distributions across layers using Jensen–Shannon divergence (JSD). From there, they define DTR (Deep Thinking Ratio): what fraction of the generated tokens are truly “deep.” The practical win is simple: sample multiple candidate answers, compute DTR early (like after the first ~50 tokens), reject low-DTR continuations, and finish the answer with the highest DTR. The takeaway for builders is clear: stop rewarding verbosity, measure mechanistic effort, and use early rejection to boost accuracy while cutting cost—often around 50%.