📅 What AI knows and what it doesn’t
The neural network is not omniscient - it only knows what was in its training data up to a certain point (cutoff date). Everything that happened later seems to not exist for her. Move the “date of training” slider and see what the model can tell you about and what it can’t.
For events after the cutoff date, the model will either honestly say “I don’t know” or, worse, come up with a plausible answer. Therefore, for the latest news, prices and facts, neural networks connect an Internet search (this is called RAG). Never rely on a neural network to figure out what happened recently without checking.
How it works: knowledge cut-off date
The model is taught once on a huge cross-section of texts collected before a certain date. After this, her knowledge is “frozen”: she does not read the news in real time and does not remember yesterday. Therefore, each model has a cutoff date (knowledge cutoff), beyond which there is emptiness for it.
The danger is that the model does not always honestly admit ignorance. Ask her about an event after the cutoff, and she can confidently come up with a plausible but false answer (this is called a hallucination). Outwardly, such an answer cannot be distinguished from the real one, so fresh facts, figures and dates should always be double-checked.
To provide the model with fresh data, they connect it to an Internet search or a document database and put what they find directly into the query - this technique is called RAG (retrieval-augmented generation). Hence the practical conclusion: for the actual, ask for a model with access to search, and for the important, check with the original source. Understanding the limits of AI is part of the AI skill and can be practiced in a free lesson.
Start for free