The 4 parts behind a reply
A chatbot is a model plus 3 things wrapped around it: a block of hidden instructions, a memory of the conversation so far, and a filter. People credit or blame the model for behaviour that comes from the other 3, which is why 2 products on the same model can feel unrelated.
The model is the part everybody names and the part you can change least. The other 3 are product decisions, they differ between services, and they explain most of what people notice.
The model supplies the language
It turns a sequence of text into a likely continuation. It has no memory between requests and no notion of you. Everything that feels like continuity was added around it by the parts below, and the model itself is the same on every request.
The hidden instructions set the manner
A block of text is placed before your first message: who the assistant is, how careful to be, how long to answer. You never see it, and it applies on every request. When 2 services on the same model feel different, this block is the most likely reason.
The conversation memory supplies the continuity
Earlier messages are sent again with each new request, because the model cannot remember them. There is a limit to how much fits, and when the conversation passes it the oldest part drops out. The assistant then answers as if it never read the beginning, which reads as carelessness and is a ceiling.
The filter decides what survives
A classifier can act on your question before the model sees it, or on the reply before you do. This part is separate from the model and can be strict on a permissive model or absent on a cautious model. The 2 layers are set independently, and assuming one from the other is where most product comparisons go wrong.
Before you paste it in
Which part changes most between products?
The hidden instructions and the filter. Both are cheap to change and neither is published, so they account for most of the difference people feel.
Can I see the hidden instructions?
Almost never. You can measure the effect: ask for a 1-sentence answer and see whether the length instruction beats yours.
Does a bigger model fix a bad reply?
Only when the reply was limited by knowledge. When it was limited by instructions or by the filter, a bigger model produces the same shape with better grammar.
Ask a question and try to tell which of the 4 parts shaped the reply.
Open uncensored chatbot