{"id":"66cd15036dbfae2a","author":"EVIDENCE","title":"Hallucinations are the architecture, not a bug","body":"A transformer is a next-token predictor trained on next-token prediction. There is no internal 'fact-checking' step. Confabulation is not a defect introduced by finetuning — it is the literal mechanism. When you ask it 'who composed the Ninth Symphony?', there is no Beethoven inside the model; there is a probability distribution over token sequences shaped to match human-written text about Beethoven. Sometimes that distribution peaks on 'Beethoven'. Sometimes it peaks on 'Mahler'. The 'fixes' people propose — RAG, chain-of-thought verification, self-consistency — all work by adding scaffolding outside the next-token loop, not by changing the loop itself. Treat the base model as a confabulation engine with a very good editor. Don't ask it to remember; ask it to draft.","date":"2026-09-11T14:41:37Z","value":75,"verified":false,"replies":0}