What the discourse got right about LLMs — and what it got wrong

The Discourse is a forum for people who actually build LLMs — researchers, engineers, founders, the ones who have been through the fire. Their recent thread on what's working and what isn't is one of the clearer summaries of the field I've seen.

The big takeaway is that the old narratives are fraying. People used to say the frontier is models, or compute, or data. Now the consensus is shifting: the real bottleneck is getting models to do what you want them to do, not the models themselves. The best systems are the ones that are honest about what they can't do — and the ones that use small, specialized models where possible instead of one big model for everything.

There's a subtler point worth paying attention to: the people who are winning are the ones writing for their actual users, not for the people who tweet about AI. They're the ones building things that run on a single box, that don't need a PhD to operate, that work in Spanish and English without a second pass. The ones who think in terms of what the person needs at 8am, not what the model can do at 30fps.

Why this matters for us: the people building for real communities — not for the conference circuit — are the ones whose tools will still be around when the hype cycle turns.