What the AI actually knows
A language model has no database in its head. It learned patterns from billions of texts, then generates the most plausible next words. Plausible usually matches true. Not always.
Two limits follow. First, a knowledge cutoff: the model only knows what existed in its training data. A DAC released last month may simply not exist for it. Second, confidence and accuracy are separate things. The model writes fluent, certain-sounding sentences even when it’s wrong. There’s a word for this: hallucination.
Use it accordingly. Models are excellent at explaining concepts, comparing approaches, and structuring decisions. For hard facts like current products, prices, and specs, they need real data behind them. The rest of this academy builds on that distinction.
Go deeper
What "learned patterns" means in practice
Training a language model means showing it an enormous amount of text and rewarding it for predicting the next word. Nothing gets stored as a fact with a label on it. What the model keeps is a compressed sense of which words tend to follow which, across everything it read. Ask it what a phono stage does and the answer comes out right because that explanation appeared thousands of times in similar wording. Ask it the input impedance of one specific phono stage and it will produce a number in the same confident tone, because a number is what belongs in that sentence. Whether it is the right number is a separate question the model cannot ask itself.
This is why models are so good at the shape of hi-fi knowledge and so unreliable on the specifics. Shape repeats. Specifics are rare, scattered, and often contradicted across sources.
Practical takeaway: trust a raw model on the explanation of how something works. Treat every specific figure it gives you as a guess until you have seen it somewhere real.
The cutoff, and why it moves
Every model was trained on text collected up to a certain date. After that date it learned nothing. The cutoff is usually months behind the model’s release, and by the time you use it, a year or more may have passed. In hi-fi that gap matters. Products get revised, prices move, a manufacturer changes hands, a streaming service drops a codec. The model will describe the world as it stood at the cutoff and will not tell you it is doing so.
New model versions land often and push the cutoff forward, which is why the Assistant carries new frontier models the day they are released. A newer model knows more recent history. It still has a cutoff, and it still cannot see last week.
Practical takeaway: for anything with a date attached, whether a product exists, what it costs, whether the firmware fixed the issue, assume the model is out of date and check.
Confident and wrong at the same time
The fluency of a model is a property of its writing, and it is exactly the same whether the content is true or invented. There is no hesitation in the tone when the model is guessing. This is what people mean by hallucination: a plausible, well-formed answer with no source behind it.
Some patterns show up repeatedly in hi-fi. Specs from one product attached to its sibling in the same range. A discontinued model presented as current. A review quoted that was never written. A measurement stated for a speaker that was never measured. Each one reads perfectly.
The model’s confidence tells you about its training data. It says nothing about the world. When you see a spec you did not expect, that is the moment to slow down.
Practical takeaway: ask the model where a claim comes from. If it cannot name a source, or the source turns out not to say that, drop the claim.
Grounding: the fix for all of the above
Grounding means giving the model real data to read before it answers, so it explains what is in front of it instead of reconstructing from memory. When the Assistant answers a question about a specific amplifier, it looks the amplifier up in the Pure Neo database first and reasons from the recorded specs. The knowledge cutoff stops mattering for that product, because the data is current. The hallucination problem shrinks, because there is a number on the page to read instead of one to invent.
Grounding changes what a model works from and leaves its abilities as they were. The explanation is still the model’s; the facts come from outside. That is the division of labor the rest of this academy keeps returning to, and the last unit lays it out in full.
Practical takeaway: when a question depends on a spec, ask it where the model can look the spec up. That is what the Assistant is for.