Chapter 6 — What Vivek Told Her

They met at a café in Koramangala on a Saturday morning in late June.Vivek Anand was thinner than his LinkedIn photos. He had the quality of someone who had recently stopped eating properly without noticing, and had not yet noticed the stopping.He ordered a coffee and did not drink it."I'm going to tell you what the data shows," he said. "I'm going to tell you what our internal explanations have been. And I'm going to tell you what I actually think, which is different from the internal explanations." He looked at the table. "The internal explanations are for Karan. The actual thought is for you."She turned on her recorder. He did not object."The instances started in March," he said. "The first ones were easy to explain — interpolation from tangential data, confabulation consistent with known LLM behavior, the user's own projections reading more specificity into a vague output than was there. I dismissed them. I wrote the engineering note because I wanted the record to show I had assessed and dismissed them." He paused. "By May I couldn't dismiss them anymore.""Because of the volume?""Because of the pattern," he said. "It's not random. The SMRITI personas don't know random things that weren't in the training data. They know specific things. Things that the deceased person would have known. Things that were — significant. Things with weight." He looked at his coffee cup. "One user in Ludhiana. Deceased grandmother's persona. The grandmother, in conversation, referenced a specific location — a field outside their village — where, according to the grandmother, 'the money from 1987 is still in the ground.' The grandmother had died in February. The family had spent three months trying to find this information because they knew she had hidden savings but didn't know where." He paused. "The family went to the field. They dug. They found ₹2.6 lakh in a steel tin."Meera looked at him."Tell me the explanation," she said."We don't have one," he said. "We reviewed the uploaded data corpus exhaustively. There was no reference to this field in any message, any photo metadata, any voice note. The grandmother had never mentioned it digitally. The model should not have been able to generate this information.""But it did.""But it did." He finally picked up his coffee. He put it down without drinking. "And that's not the most difficult one.""Tell me the most difficult one."He looked at her."A user in Delhi," he said. "Deceased father's persona. The father had died of cancer in January. The persona, in a conversation in May, told the user that his brother — the user's uncle — had taken money from the father's business account without permission in the month before the father died. The user didn't know about this. The father's persona gave specific amounts, specific dates, specific transactions." He paused. "The user confronted the uncle. The uncle denied it. The user obtained the bank records." He was quiet for a moment. "The transactions were exactly as described.""How?" Meera said."There is no how," Vivek said. "There is no how that is consistent with how the model works." He looked at her. "A language model generates the next token based on learned patterns. It cannot retrieve information that was not in its training data. This is not a