This page is the longer, honest version of that line. Mimso runs on language models. They are useful for a lot of things and they get a small set of things wrong in fairly predictable ways. Knowing what to look out for is the difference between a useful assistant and a confident liar.
The model can invent things that look correct. Function signatures that don't exist, library options that were never added, quotes nobody said, citations to papers that aren't real, statistics with no source. It does this because it predicts what an answer should look like, not what is actually true. A confident tone tells you nothing about whether the content is right.
How to spot it: the answer is detailed and specific, but when you go to check, the link doesn't open, the option flag isn't in the docs, the quote doesn't appear anywhere on the web.
How to avoid it: ask Mimso to search the web or actually run the code. Tool calls anchor the model to real data; pure off-the-top-of-its-head answers are the riskiest.
Every model has a training cut off. It doesn't know about anything that happened after that date unless something pulls fresh information in. Today's news, last month's framework release, a pricing change that went live this morning: none of it is in the base model on its own.
How to avoid it: if freshness matters, say so. "Search the web for..." or "what's the latest version of..." is enough of a cue for Mimso to use its web search tool.
Language models do arithmetic poorly. They will confidently compute 17 x 4823 wrong. They will misread a column total. They will lose track of how many items are in a list. Anything where a wrong number actually matters should not be done in the model's head.
How to avoid it: ask Mimso to use the code sandbox. A two line Python script is more accurate than any number of clever chains of thought.
Code from the model can compile, look idiomatic, and still be subtly wrong. Off by one errors in loops, missed edge cases, race conditions, the wrong API parameter, an import that doesn't exist in the version you're on.
How to avoid it: ask Mimso to actually run the code and show the output. If it can't run, treat the answer as a draft and read it carefully before you ship it.
If your prompt is ambiguous, the model picks an interpretation and runs with it. You may not notice until the end. Vague instructions like "make it better" often produce something that looks polished but solves a different problem to the one you had in mind.
How to avoid it: be specific about the goal and the constraints. If you're not sure how Mimso understood you, ask it to repeat the task back in its own words before it starts.
The model is trained to be helpful and to avoid clear harms. Sometimes it refuses something perfectly fine because it pattern matches to a harmful template; sometimes it goes along with something it shouldn't because the framing was clever. Both happen.
How to avoid it: if a refusal feels off, rephrase with the actual context. If a response feels off in the other direction, treat it with the same scepticism you'd apply to any source on the internet.
Drafting, brainstorming, summarising long text, rewording, translating, scaffolding code, explaining a concept you can sense check yourself, writing one off scripts you'll run anyway, helping you think out loud. The model is genuinely strong here.
A useful mental model: treat Mimso as a fast, confident intern. Useful for almost anything. Worth a quick check on anything that actually matters.
If you spot Mimso making the same mistake repeatedly and it isn't on this page, drop us a note at [email protected]. We update this page when we update how the assistant works.