User comments on a Hacker News thread argued that data mining does not mean models are updated instantly. The original poster noted that instant retraining would be prohibitively expensive at scale and that it is far easier for providers to batch user data with other datasets for later use. They added that even after being included in training data a model may not reliably reproduce specific facts because models are not one to one with their inputs and training is constrained by size and cost.
Several participants expanded on timing and practice. Commenters observed that pretraining does not happen frequently and varies by provider, with one noting that OpenAI and Claude each average roughly one flagship release a year. Others said weekly retraining is still too expensive at scale, and that data mining may only be reflected in the next generation of a model or via fine-tunes and customizations applied on top. Multiple replies emphasized that smaller models or specialized fine-tunes are common ways companies handle more frequent updates without full retraining.
The thread also discussed failure modes and what ´your data´ means in practice. Some users raised tokenization and counting examples, reporting inconsistent answers across chats for simple letter counts and noting that viral questions can lead to ad hoc fixes rather than general learning. One commenter suggested tokenization issues are fundamental to llm behaviour but that many models can now handle such tasks using reasoning, code, or tools. Another contribution argued the value of collecting questions and followups is not only factual answers but improving how models interpret ambiguous queries, which can reduce token generation and operating cost when deployed.
