The Coming Wave — a working engineer's review

“The challenge of our era will be containing the technologies we cannot uninvent.”

I read The Coming Wave twice — once for the headlines, once for the engineering. Then we hosted Mustafa at MS Zürich in March 2026 and I came away with a third reading: this is the most useful book on AI for people who actually ship AI in regulated environments, even if you disagree with its predictions.

Why an engineer should care

Most safety writing is either too abstract for production engineers, or too narrow (one technique, one threat model). Suleyman’s framing is operational: containment isn’t a single technique, it’s defense-in-depth.

The book’s “ten steps” map onto things I now ask in customer reviews:

flowchart LR
  A[Safety research] --> B[Audits]
  B --> C[Choke points]
  C --> D[Makers]
  D --> E[Businesses]
  E --> F[Governments]
  F --> G[Alliances]
  G --> H[Culture]
  H --> I[Movements]
  I --> J[Coherence]
  J -.-> A

When I sit with a customer rolling out an agent to 50,000 employees, I don’t argue containment from first principles anymore. I point at the chapter and ask: which of the ten do we have, and which are we faking?

A formal sketch: oversight has to scale faster than capability

Here’s the worry, in one inequality. Let ( C(t) ) be the capability of a deployed AI system at time ( t ), and ( O(t) ) be the effective oversight applied to it (humans + automated evals + auditors). Containment requires

$$ \frac{dO}{dt} \ge \frac{dC}{dt}. $$

Capability scales with compute and data — both roughly exponential. Oversight, by contrast, scales with trained reviewers, audit infrastructure, and red-team capacity. In practice,

$$ C(t) \sim e^{\alpha t}, \qquad O(t) \sim t^{\beta}. $$

For any ( \alpha > 0 ) and finite ( \beta ), the ratio ( O(t)/C(t) \to 0 ). So either:

1. (\alpha) comes down — slower deployment, choke points, safety research that reduces useful capability per FLOP, or
2. (O) goes super-polynomial — oversight automated by AI itself: recursive eval, scalable oversight, debate.

Both end up in The Coming Wave under different chapter names. The book’s value is that it doesn’t pick one; it argues for all of the above, simultaneously, layered.

Where I push back

• The book is light on technical mechanisms for containment. Engineers still need their own reading list (which is what /ai-safety is for).
• “Containment” risks sounding like control. In practice the most effective containment I’ve seen has been transparency + reversibility, not lockdown.
• The asymmetry argument (offense vs defense in AI security) deserves its own book.

What I’d pair it with

• Concrete Problems in AI Safety — Amodei et al., 2016 — the closest thing alignment has to a Rosetta Stone
• The Alignment Problem — Brian Christian — the human-centered view
• Power and Prediction — Agrawal, Gans, Goldfarb — the economics
• A subscription to your own customer’s incident log

My take in one line

Read it once for the policy, twice for the operations. The best gift it gives an engineer is a shared vocabulary with the non-engineers who decide whether your agent ships.

Rating: ★★★★★ — on permanent loan from my desk.