LinkedIn·Monday, 24 August 2026·2d ago

A reasoning model doesn't answer you straight away. It writes to itself first, sometimes thousands of words you never see, and only then…

Betacowork

975 followers

A reasoning model doesn't answer you straight away. It writes to itself first, sometimes thousands of words you never see, and only then replies. That hidden writing isn't free. It fills the same context window your actual question lives in, and you pay for it either in money or, if you're running locally, in time. Usually it's worth it. Sometimes it isn't: The model starts going in circles, revisiting the same thought, and never converges. Andrew calls it the loop of doom. The interesting part is why it happens more on compressed models. Shrink the numbers inside a model from 16 bits to 8 and each dimension drops from ~65,000 possible states to 256, and there are hundreds of those dimensions. Less room to move, more chances to end up back where you started. Frontier models do it too. You just don't see it, because there are guardrails on the server side quietly cutting it off. Nobody wants to publish that the answer took two hours and this is the bill. 📅 New: we're running an AI Weekend for entrepreneurs at Betacowork on 26 and 27 September, in Brussels. Two days, hands-on, built for people running a business rather than a research lab. Registration opens shortly. 🔔 Follow so you catch it when it does. The rest of this series continues in the meantime, with hardware next. Coworking in Brussels, one #free day to try it: welcome.betacowork.com/try #coworkingspacebrussels #event #Brussels #Freelancing #Entrepreneurship #AI