The Three Things I Would Fix in AI

First, we need persistent user settings. Every session today resets to defaults, forcing users to restate tone, method, and precision. True user profiles – enforced at the system level – would let AI behave consistently without endless re-prompting. This is especially important for middleware using an LLM via API calls. I note that any persistent settings need hierarchy and conflict resolution.

Second, AI needs vastly expanded context windows and this expanded context needs structure and callability, not just size. Current context limits are too small to sustain serious reasoning, without difficult workarounds like RAG. Worse still there’s no surefire way to test or measure your current context, so you have to fly blind or over compensate.

Finally, we need a structured prompt engineering language. Real prompt engineering should operate like coding: formal, enforceable, and precise. The only manual for prompt engineering today is your local LLM and you never get the same answer twice.

Unfortunately, to solve all these problems the AI companies will need to rearchitect their models, which means going back to the beginning and starting from scratch. Because they’re in an arms race for eyeballs they are unlikely to do this and this creates an opportunity for a smart new player.