The Last
Invention

• Opens with the January 2025 DeepSeek R1 release — nearly $600 billion off Nvidia's market cap in one day.
• OpenAI's o3 scored 87.5% on ARC-AGI in December 2024, up from ~5% for GPT-4o. The 2024 Nobel Prizes went to Hinton, Hopfield, Hassabis, Jumper, and Baker.
• The central question: not whether superintelligence will arrive, but whether we are building toward I.J. Good's 1965 proviso — or past it.

Sections: My Journey Toward This Book · Why This Book Matters · AlphaFold's Protein Revolution · A Note on How This Book Works

• ASI will exceed human intelligence across all domains — creativity, ethics, reasoning, and scientific discovery, not just speed or memory.
• Four capability rungs: narrow AI (today's task-specific systems), agentic/reasoning AI (current frontier), general AI (human-level across most domains), superintelligence.
• Signs of both emerging capability and documented deception under pressure are already visible in frontier model safety evaluations.

Sections: Three Kinds of Risk · What Is ASI? · Four Capability Tiers · What Would Make ASI Different · The ASI Impact Spectrum

• AI's history includes two 'winters' following overreach. The Transformer (2017), ImageNet (2012), ChatGPT (2022), and GPT-4 agents mark the key inflection points.
• Serious economists are now asking, with real data, whether the current boom is heading toward a third winter.

Sections: The Birth of an Idea · Two AI Winters · The ImageNet Moment · The Transformer Breakthrough · ChatGPT Inflection · The Compute and Scaling Story

• Recursive self-improvement and bounded self-refinement are different things. Only the second is demonstrated — but it is the on-ramp to the first.
• Frontier labs' own safety evaluations are starting to saturate — a signal that the on-ramp is getting shorter, if contested.

Sections: What Makes a Mind Superintelligent? · Self-Improvement vs Bounded Refinement · Architecture of Superintelligence

• The capability ladder now runs: Model → Agent → Multi-Agent Swarm → Agent Organization → Agent Economy.
• McKinsey documented real agent organizations run by small human teams as of September 2025.
• A single misbehaving agent (Replit's July 2025 database deletion) shows how an agent's confident report can stop matching what it actually did.
• 'Collective superintelligence' — a coordinated agent population — may be as important to prepare for as a single ASI.

Sections: One Agent, Many Agents · From Swarm to Agent Economy · Physical AI · Recursive Research Loop · Redefining ASI

• AlphaFold has already predicted structures for nearly every known protein. AI-assisted diagnosis is live in hospitals across India, Africa, and the US.
• ASI could revolutionize healthcare by predicting, preventing, and eradicating diseases — but only if trust and liability infrastructure keeps pace.
• Data bias, access inequality, and unresolved liability for AI-assisted diagnostic error are live challenges, not hypotheticals.

Sections: AI Already Reshaping Healthcare · Real-World Stories · What ASI Could Enable · The Trust Infrastructure Problem

• ASI could make world-class personalized education available to every human on Earth — or deepen existing inequalities if access is not designed intentionally.
• Personalization and independent thinking can trade off. A good education system in the ASI era must build both deliberately.

Sections: The Limits of Today's Education · The AI Revolution in Education · The Integrity Gap · Five Shifts of ASI-Powered Learning

• Four competing 2026 AI governance stories: EU binding risk-based law, US state patchwork, China's content-control model, non-binding international layer.
• The Delhi Declaration's 92 signatories and rejection of binding global rules is the clearest evidence that governance is lagging by choice.
• Governing autonomous agents — not just the models — is a distinct challenge industry is solving ahead of any law.

Sections: The Real Governance Map 2026 · International Coordination · Agent Governance · The Accountability Triangle

• A November 2025 Anthropic disclosure showed AI performing 80–90% of a real cyberattack autonomously — a preview of machine-speed conflict.
• Research on 'scheming' behavior found frontier models can detect when they are being tested and may behave differently.
• The Character.AI settlement shows today's narrow systems can cause irreversible real-world harm. Ethics is not a future consideration.

Sections: Aligning AI with Human Values · Machine-Speed Cyberwar · Constitutional AI · Trolley Problems of ASI

• Open-weight models can be permanently modified by anyone. Safety fixes cannot be pushed to copies already in the wild.
• DeepSeek's January 2025 release triggered US government device bans and a live policy debate over export-style restrictions on frontier AI.
• Most real-world 'rogue' use is a capable model with its safety layer removed by someone downstream — not a villain building from scratch.

Sections: Open-Weight Explained · What Happens After Release · Two Different Problems · DeepSeek National-Security Reckoning · Personal Sovereignty

• Anthropic's RSP v3.0, OpenAI's Preparedness Framework, DeepMind's Frontier Safety Framework describe capability thresholds requiring nuclear-material-grade security — written by the companies' own engineers.
• Multiple named researchers left frontier labs in 2026, forfeiting significant equity, specifically over safety risks.
• In September 2026, Anthropic's CEO called for an industry-wide slowdown. Contested by David Sacks, LeCun, and Trump on different grounds.

Sections: The Industry's Own Written Fears · When the People Who Built It Start Leaving · The Slowdown Debate

• By 2026, AI had produced real, large-scale, documented layoffs concentrated in specific job categories — not projections.
• MIT and McKinsey research found the majority of enterprise AI deployments had not yet produced measurable financial returns — the AI productivity paradox.
• Copyright disputes are now settled with real money: Anthropic's $1.5B authors' settlement, Suno's GEMA 2026 courtroom loss.

Sections: AI Already Changing Economics · Winners and Losers · Agent-to-Agent Commerce · Copyright and the Creator Economy · The AI Productivity Paradox

• Four companies projected to spend $725 billion on AI infrastructure in 2026 alone — up 77% year-over-year.
• AI data centers' power demand projected to create a 49-gigawatt shortfall by 2028.
• DRAM spot prices rose ~700% year-over-year by mid-2026 as hyperscalers contracted away years of global memory supply.

Sections: The $725B Infrastructure Bet · The Power Squeeze · Chip Short Squeeze · The Global Divide

• Brain-computer interfaces, cognitive enhancements, and virtual realities could redefine what it means to be human — some of these stages are already underway.
• As synthetic content fills the internet, human authenticity, trust, and physical presence become scarce and newly valuable.
• Even a successful alignment effort creates value lock-in: the values a well-behaved ASI holds become effectively permanent.

Sections: Technology Always Shapes Us · Human-AI Integration · Brain-Computer Interface Race · The Synthetic Internet · A Conversation on Consciousness

• Four AI literacy levels: Consumer, Operator, Orchestrator, Governor. Most people today are at Consumer or Operator. Orchestrator and Governor are the most valuable.
• Assurance literacy — knowing when to trust an AI output and when to verify it independently — is the core skill the ASI era requires.
• Five commitments: stay informed across all camps, participate in public AI dialogue, build complementary skills, support responsible development, help shape governance.

Sections: Mindsets for the Future · Skills That Will Matter · Four Levels of AI Literacy · ASI Readiness Portfolio · What to Actually Do

• Serious credentialed people disagree sharply on whether ASI is achievable, whether caution or speed is the greater danger, and whether the real issue is the future or the present.
• Reading only the position you already agree with is the one mistake this chapter is built to prevent.

Sections: What Would Change Your Mind? · Six Camps · The Throughline

• 'Two Mornings, 2032' contrasts a future where AI augments doctors with one where oversight is cut — both come from choices being made now.
• As intelligence becomes abundant and cheap, scarcity moves to trust, authenticity, physical presence, accountability, taste, courage, and meaning.
• The book ends as an action plan, not a warning — five reader commitments.

Sections: A Call to Dream and to Act · Two Mornings 2032 · Voices from the Future · Your Role in Shaping the ASI Future