The People Building AI Just Told You to Stop Waiting for a Better Model

If you've been holding off on delegating real work to AI because the model needs to get better first, the case against waiting came from an unusual place this week: the people building it.

On September 12, Anthropic's Dario Amodei published "We Must Pace the Frontier," arguing for deliberately slowing model development. Sam Altman endorsed it within two days: "No amount of American competitive pressure should justify recklessness." Demis Hassabis called it "the direction correct for meeting this critical moment." Three CEOs whose companies compete directly agreeing to slow down is itself the story — but the more useful signal came from the conference floor, not the stage.

At Dreamforce the same week, business leaders told reporters something quieter: they're still learning to deploy AI models that are a generation or two old. Nobody on that floor was asking for a smarter model. They were asking how to use the one they'd already licensed.

Ethan Mollick named the pattern directly in "The Overhang," published September 18: current models can already do far more than most people ask of them. The gap isn't capability — it's four things a person brings that the model doesn't supply on its own: deep knowledge of the specific situation, wide knowledge across domains, taste (knowing what good looks like), and agency (deciding what to hand off and what to keep). Anthropic's own numbers back this up without mentioning a new release: Claude now handles 26% of Anthropic's internal AI R&D work, up from under 1% in February. That jump didn't come from a smarter model. It came from Anthropic getting better at supervising the one it had.

What to do about it Monday morning

Stop treating the next model release as the thing standing between you and delegating more. The upgrade won't supply taste, context, or judgment about what to hand off — those are yours to give it, with the model you already have open right now.

Write down what "good" means before you generate anything. Mollick's "taste" isn't a personality trait — it's a standard you can hand the model as a check. If you can't say what makes one draft better than another, no model version fixes that for you.

A marketing manager at a 40-person SaaS company had been telling herself she needed "a model that actually understands our brand voice" before letting AI draft full blog posts instead of just outlines. Eight months and two model upgrades later, the drafts still needed a full rewrite. The brand voice had never been written down anywhere the model could check against — three adjectives in a slide deck, not a standard. She spent one afternoon pulling ten sentences she'd call genuinely on-brand and ten she wouldn't, and used those as the check before every draft after. The next model upgrade changed nothing that afternoon hadn't already fixed.

An operations lead at a 200-person logistics company kept switching between AI tools chasing "smarter," while never telling any of them the three dispatch constraints a new hire learns in their first week — which routes can't be combined, which customers get priority holds, which changes need a supervisor's sign-off. Every tool she tried made the same scheduling mistakes, because none of them had been told the constraints a smarter model wouldn't have guessed either. The fix wasn't a better model. It was deciding, once, which scheduling decisions the tool could make on its own and which ones needed her first.

Neither of them was missing a capability their next subscription renewal would have supplied. They were missing a standard and a boundary, and both are gettable from the tool already open on their screen.

The one thing

The three people running the labs just told you the tools are good enough. The gap left standing is the one only you can close, and it doesn't wait for the next release.