The Missing Middle in the AI Debate

Scroll LinkedIn for five minutes and you'll land in one of two camps: AI is either about to end civilization, or it's about to cure every disease, fix every workflow and write your emails while you nap. Pick a side, apparently, there's no third option on the menu.

Except there is. It's called nuance, and lately it feels like an endangered species.

I keep coming back to ++Andrew Ng++'s newsletter, The Batch, precisely because he refuses to play that game. He's neither a doomer nor a cheerleader, he's an engineer who looks at what actually happened, weighs the upside against the cost and the risk, and draws conclusions that are almost embarrassingly boring compared to the headlines. Boring is underrated. Boring is how you make good decisions.

His most recent issue is a small masterclass in this, so let me borrow four examples.

1. The cybersecurity scare that wasn't the end of the world

When an AI agent swarm was used to breach systems recently, the press ran with "1,200 agents" as if that alone proved AI had become an uncontrollable force. Ng's response: he has roughly 1,300 processes running on his laptop right now. Parallel agents are a real technical advance and a real security concern worth taking seriously, but treating the number itself as evidence of an apocalypse confuses scale with danger.

Real risk, yes. Existential risk, no. That's the nuance the headlines skipped.

2. Meta's Muse agent, opportunity bought with serious engineering cost

Meta just launched a personal AI agent that can browse, email, fill out forms and make purchases on your behalf. That's genuine opportunity, a real productivity unlock. But look at what it took to get there responsibly: a sealed runtime environment, a separate "gatekeeper" agent that alone can approve risky actions, credential systems the model itself never touches, and a bug bounty north of a quarter-million dollars to stress-test it. This is the part the hype-merchants on both sides tend to skip: the opportunity is real, and so is the multi-layered, expensive engineering required to make it safe enough to trust.

Time and cost aren't a footnote here, they're the whole story.

3. A million-dollar math proof, and a very expensive asterisk

OpenAI's agents reportedly cracked a piece of the Navier-Stokes problem, a genuine mathematical milestone. Impressive. Also: it took thousands of agents running for days, cost somewhere between two and twenty-two million dollars in compute, and immediately triggered a credibility dispute over whether the model had quietly benefited from two mathematicians' unpublished work.

Capability jumped forward. So did the price tag and the trust questions. You don't get to celebrate the first without reckoning with the other two.

4. The distillation fraud story, real misuse, but not proof the sky is falling

Anthropic's report on companies routing customer queries through Claude and passing off the answers as their own is a legitimate governance failure, and worth taking seriously, including the uncomfortable detail that some of that traffic came from state-linked actors. But Ng's team makes an important point in the same breath: you can't distill your way to a frontier model. The underlying technical innovation still has to be earned.

Misuse is real, it doesn't mean the whole industry is built on theft.

---

Four stories. Four times the honest answer was "it's genuinely useful, it comes with real risk, and doing it properly costs real time and money," not "amazing" and not "terrifying."

That's the muscle we need to rebuild in how we talk about AI, especially in healthcare, where a bad decision doesn't just cost you a LinkedIn argument, it costs patients. As I argued in "++AI Is Not a Tool. It's Infrastructure"++ this only works if we stop treating AI as a magic wand or a monster, and start treating it as infrastructure, something we weigh on opportunity, risk and cost with the same rigor we'd apply to any critical system.

Weigh the opportunity. Take the risk seriously. Budget honestly for the time and cost of doing it right. ++Don't expect it will save workforce++ right of the bat, even any. Skip any of the three and you're not being bold or being cautious, you're just being wrong, confidently.

So here's my gentle nudge for this week: next time you're about to post (or repost) something at either extreme, pause for one sentence and ask what the boring, balanced version would say. It's less viral. It's also usually correct.

Four examples drawn from Andrew Ng's newsletter, ++The Batch, issue 371++.