What's this about?
Ethan Mollick argues we've barely scratched "harness engineering" — the scaffolding around a model matters as much as the model — echoing OpenAI's ARC-AGI-3 result the same day. He also showed Flux 3 generating a strikingly coherent cinematic clip from one prompt.
What exactly is "harness engineering"?
See the GPT-5.6 Sol entry: the harness is the prompting, memory, tools and retry logic around a model. Mollick's point is that much of the next round of gains will come from better harnesses, not just bigger models — a cheaper, faster lever most teams haven't pulled yet.
Why it matters
999 + 571 + 278 likes. It reframes "is the model smart enough?" into "have we built the right scaffolding?" — practical guidance for anyone shipping with today's models.