Hacker News Reader

I know the point you’re making is half facetious—do you feel like LLMs are specifically getting better at planning?

> do you feel like LLMs are specifically getting better at planning?

When harnessed, yes. They still suck at it, but they are getting significantly better over a relatively short time.

And you don't have to look very far in any direction economically or politically to wonder if humans in 2026 are really as good at planning as we like to think and/or if we may be regressing in this area. Because a lot of things we are responsible for aren't going very well for the majority.

I'll write the dissenting opinion that the models still aren't very good at planning rather it's the agentic harness and prompts thereof that are giving us better planning ability on long and larger tasks.

So they're good at planning but only when used certain ways.

No, only when use in conjunction with not-LLM software

If the model is bad at planning but the system created with the model at its core is, eventually, good at planning, what's the functional difference?

Reply on news.ycombinator.com