The Coming Wave
Mustafa Suleyman · 2023
Editorial rating
- Evidence
- 8/10
- Actionability
- 7/10
- Originality
- 9/10
The thesis
AI and synthetic biology will spread because their benefits are too valuable to reject, yet their power, accessibility, and autonomy make traditional control methods inadequate. The central challenge is not stopping progress but building layered containment before failures scale beyond institutions that can correct them.
Who this is for
Executives adopting frontier technology, policymakers designing AI rules, founders building powerful systems, and risk leaders who need a framework broader than compliance checklists or speculative catastrophe scenarios.
My favorite quote
Just because consequences are difficult to predict doesn’t mean we shouldn’t try.
Why it matters
Uncertainty is often used as an excuse for delay, even though early choices shape which risks become difficult to reverse.
Do this
Name one plausible failure in an AI or automated workflow and assign a person who can stop deployment if it appears.
Start here
Containment is not a single regulation, shutdown switch, or promise from a responsible company. It is a layered system of technical safety, audits, incentives, laws, international coordination, and institutions capable of intervening throughout a technology's life cycle. Use that lens whenever someone presents one safeguard as sufficient.
Critical summary
Mustafa Suleyman helped found DeepMind and later built Inflection AI, giving him an unusually close view of how quickly artificial intelligence moves from research result to widely available capability. Working with Michael Bhaskar, he places AI beside synthetic biology as the core of a coming technological wave that will alter knowledge work, medicine, warfare, politics, and the distribution of power. These technologies are hard to govern because they combine four properties: asymmetric impact, hyper-evolution, omni-use, and increasing autonomy. A small group can cause effects once reserved for states, capabilities improve at digital speed, the same systems support beneficial and destructive uses, and machines can act with less direct supervision. Suleyman calls the resulting control challenge the containment problem.
What it gets right
- Explains why general-purpose AI cannot be governed like a narrow product with one predictable use
- Connects technical risk to incentives, state capacity, corporate competition, and geopolitical pressure rather than treating safety as an engineering problem alone
- Rejects both effortless techno-optimism and blanket prohibition, forcing readers to confront the narrow path between uncontrolled proliferation and oppressive surveillance
What it overstates or misses
- Treats the rapid arrival of highly capable autonomous systems as more certain than the evidence allows
- Gives less attention to ordinary harms already visible in labor markets, discrimination, market concentration, and environmental costs
- The ten containment steps are directionally strong but remain too broad for organizations that need operating rules, budgets, and enforcement mechanisms
The evidence combines technological history, current research, industry experience, and forward-looking scenarios. Suleyman is strongest when describing proliferation incentives and weakest when converting plausible danger into estimates of likelihood. His insider status adds valuable detail but also creates tension: the builders warning about runaway competition are still competing to build the systems. The final containment agenda correctly spans safety research, audits, corporate reform, government capability, treaties, culture, and public participation, yet coordination at that scale is precisely what the book shows to be difficult. The verdict: an essential risk framework for the AI era, even if its timeline confidence outruns its proof.
Key concepts
The Coming Wave
A cluster of fast-moving technologies led by AI and synthetic biology that can create broad prosperity while redistributing power and risk.
Containment Problem
The difficulty of monitoring, limiting, redirecting, or stopping powerful technologies at every stage from research through mass deployment.
Four Features
Asymmetry, hyper-evolution, omni-use, and autonomy explain why small actors can wield rapidly improving systems across countless purposes with declining oversight.
Pessimism Aversion
The tendency to dismiss disturbing forecasts as alarmism, which prevents serious preparation before evidence becomes impossible to ignore.
The Narrow Path
The difficult route between uncontrolled technological catastrophe and a surveillance-heavy state that suppresses innovation and freedom in the name of safety.
Core insights
-
Capability spreads faster than control
Falling costs and reusable digital knowledge allow powerful tools to proliferate before regulators, companies, or norms can adapt.
-
General-purpose systems resist simple rules
A model useful for education, coding, research, and design can also enable fraud or attack, so regulating one application misses the underlying capability.
-
Competition weakens restraint
Companies and states hesitate to slow down because each fears that a rival will capture the economic or military advantage.
-
Safety must be layered
Technical safeguards fail without audits, incentives, accountable leadership, capable government, international agreements, and mechanisms for public scrutiny.
-
Containment starts before deployment
Design choices, access controls, infrastructure concentration, and funding conditions shape risk more effectively than cleanup after proliferation.
Implementation steps
Today
- Choose one high-impact AI use case and document its owner, users, data sources, failure modes, and shutdown authority.
- Ask whether the system is general-purpose, autonomous, easily copied, and capable of asymmetric harm.
This week
- Run a ninety-minute containment review covering technical controls, human oversight, access restrictions, monitoring, incident response, and appeal routes.
- Add a deployment gate requiring evidence that benefits, foreseeable harms, and rollback procedures have been reviewed by named decision-makers.
This month
- Create a tiered governance model that applies stronger testing, logging, approval, and external review as capability and impact increase.
- Conduct a red-team exercise where a small actor misuses the system, the model behaves autonomously, or a vendor suddenly removes access.
Ongoing
- Track capability changes and incidents rather than assuming the risk assessment remains valid after launch.
- Maintain independent challenge from security, legal, domain experts, affected users, and leaders who are not rewarded for rapid deployment.
Suggested 30-day practice plan
An editorial application plan created by Monolithic Vault - an interpretation of the book's ideas, not part of the original book.
- Day 1
Inventory every AI system used in one team and identify which can influence customers, money, employment, safety, or sensitive information.
- Day 3
Score each system for autonomy, generality, ease of proliferation, asymmetric impact, and speed of capability change.
- Day 7
Select the highest-risk system and map technical, operational, legal, and vendor controls across its full life cycle.
- Day 14
Test one realistic misuse or failure scenario and record whether monitoring, escalation, and shutdown procedures work.
- Day 21
Close the most serious control gap and assign named owners for remaining risks, deadlines, and accepted exceptions.
- Day 30
Approve a recurring containment review with risk tiers, deployment gates, incident reporting, external challenge, and executive accountability.
Free PDF summary
Take this analysis with you: a designed two-page field-notes sheet with the thesis, my favorite quote, the key concepts and core insights, and the full 30-day checklist. Print it or keep it - free, no signup.
Go deeper
If this analysis earned your attention, the full book goes further than any summary can. The original is always the primary source.
