Post by Johannes Van Tongelen

Last week, AI invented a source that looked perfect: one property colliding with another. We are going deeper into the properties AI runs on, one collision at a time. This week, two more collide, and the failure is quieter. AI drops the rules you set, without ever telling you. At the start of a long chat you set two rules. "Free tools only," "Under 200 words" Forty messages later, the answer is excellent. It is also 400 words and recommends a paid tool. Nothing broke. No error. The conversation just quietly drifted from where you started it, and you almost did not notice. When this happens people say the model "forgot" the rules, or "ignored" them. Both words point the finger in the wrong direction, and that is why the drift keeps winning. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝗰𝗼𝗺𝗲𝘀 𝗳𝗿𝗼𝗺 The model did not forget in any human sense, and it did not disobey.  It only ever acts on what is currently inside its context window, and it faithfully follows your most recent instruction.  Your opening rules are still true to you. They have simply scrolled out of what the model can see. 𝗪𝗵𝘆 𝗶𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 Two properties meet at once. Working memory: as the thread grows, your early constraints fade out of the window.  Steerability: the model keeps obeying your newest steer. Put them together and it can contradict something you agreed forty messages ago,  with no flag from either side, because each answer looks right for the thing you just asked. 𝗛𝗼𝘄 𝘁𝗼 𝗽𝗿𝗲𝘃𝗲𝗻𝘁 𝗶𝘁 Here the fix is mostly about how you close a thread, not how you open it. You delegate in bounded chunks. Long, multi-step work gets broken into fresh sessions at natural boundaries, with the key rules restated at the top of each.  A short memory should not be asked to hold an hour-old constraint. When you add an instruction deep in a thread, you re-state the standing ones with it. "Still in force: free tools only, under 200 words. Now do this."  You refresh what has faded so the latest steer and the original intent stay aligned. You judge the output against your original requirements, not just the last prompt. A locally correct answer can be globally out of bounds. And the anchor is the read-through on the way out. Before anything leaves a long thread, you re-read the whole thing against your original must-haves.  That pass is where accumulated drift gets caught, because a thread long enough to lose the thread will not announce that it has. → When you pull a result out of a long AI thread, do you check it against your first instruction or your last one? Next week, the third collision: a clean, confident number that is simply wrong, and why the tidier it looks the less anyone checks it. #AILiteracy #WorkingWithAI #AIstrategy #CIO #DigitalLeadership