the intent problem and thoughts on solutions
August 2, 2026
proper interpretation of human queries by ai is still a largely unsolved problem today. no matter how much structure you give ai, there’s still gaps it needs to fill, and oftentimes the gaps are wrong.
it misinterprets what you mean by the user experience, or what the actual bug you pointed out is.
even if ai reaches human levels of interpretation, it still won’t be perfect. two humans have different interpretations of someone’s intent all the time!
i think this will be solved in a two ways:
- better context. by giving ai more contex of our lives, and the work we are doing, it have more context necessary for sense making.
- better intuition. this quality is ephemeral, but ai must develop better intuition for determining intent. when i interpret another humans intent, so much of it is rooted in a feeling, not what they said, or what i know about them.
today you can already manually give ai additional context with quizzes, or do a brain dump. but, it’s not perfect. obviously if ai also had context over your whole life, no doubt it could fill the gaps in better.
but, solution #2 is clearly the more difficult one. how can we give ai the ability to intuit the correct gaps?
i don’t really know, and im not an ml researcher, but here’s some rough thoughts:
- intent reasoning. steps in inference process that forces ai to think about feeling, and what the query really means.
- may already be happening? idk.
- training on latent goals instead of explicit ones.
- model learns to infer intent through examples
- determining when something is uncertain, and flagging another question instead of going forward with an assumption
- fable is def better at this, but still not rlly there.
- different training process altogether. autoregressive models don’t seem to be right for this. perhaps something similar to world models would allow for greater sensemaking?
- continual learning for intent
- as we see models actually get good at learning on the job, one improvement will be properly inferring intent.
- parallel exploration
- have multiple instances interpret your intent,
if someone reading this has other thoughts, i’d love to hear them! reach out to youfoundaaron@gmail.com
