things ai still sucks at
August 28, 2026
(this is a running list, and if you have any you think im missing or disagree with - reach out to me youfoundaaron@gmail.com)
a few notes on where i think ai is lacking and some ways to patch these gaps in the short term:
- interpreting intent
- ai is quite bad at filling in the gaps between what you've said explicitly and a possible solution. often makes poor assumptions, both about user experience, and the tech. more written [here](the intent problem and thoughts on solutions).
- patch: q+a tool, more context, explain everything, even things you think don't need to be. but to be honest its not perfect, and can still falter.
- reasoning about caches at scale.
- not able to evaluate on its own, what is important enough to be cached? where does it not make sense to cache this entity at all?
- patch: reviewing work, asking it to think about what data is important vs not when designing cache.
- problem solving
- oftentimes, ai gets stuck inside its current context iwndow, and can't effectively "think outside the box" to debug things.
- an example recently for me is that ai was having a hard time determining properly implementing an ultrawide band protocol. i had to prompt it to use a sniffer board, and have my iphone send pulses to catch it.
- patch: explicitly asking it to think outside the box, consider options with no bounds and having clear documentation it can reference for past actions. better context definitely solves some of these problems on its own.
- explaining things
- oftentimes it keeps using abstractions instead of just explaining the thing.
- also just the entire claude slop issue.
- solution: using skills like "humanizer", specifying how many words the response should be.
