In 2015, stick-figure diagrams on a quirky blog jolted the internet into imagining a world of superintelligence. By 2025, new scenario planning documents like AI 2027 were charting not just concepts, but year-by-year roadmaps of how humanity might stumble—or sprint—into an age of machines smarter than us. A decade of breakthroughs has transformed both the technology and the way we talk about it. Here’s how our thinking has evolved—and what it means for the years ahead.

When Tim Urban hit “publish” on his Wait But Why two-parter, AI was not yet dinner-table conversation.

Siri was clumsy, Tesla’s Autopilot had just launched in beta, and DeepMind’s greatest public achievement was teaching a neural net to play Atari.

Yet Urban’s essays, drawing heavily on Nick Bostrom’s Superintelligence and the work of futurists like Vernor Vinge, reached millions. Why? Because he distilled technical arguments into a set of intuitions anyone could grasp:

• Three levels of AI. Artificial Narrow Intelligence (ANI) is today’s recommendation engine or chess bot. Artificial General Intelligence (AGI) would match humans across the board. Artificial Superintelligence (ASI) would blow past us in every dimension.

• Exponential curves. Our brains expect steady, linear progress. Computing—and by extension, AI—improves on an exponential trajectory. The “flat line” suddenly shoots skyward.

• The fork. Get alignment right and we enter utopia: cures for disease, abundance, even immortality. Get it wrong and ASI pursues goals indifferent—or hostile—to human survival.

Urban’s stick-figure timeline of “here we are → here’s where it could go” seeded an intuition: that the future could collapse toward us faster than we were prepared for.

Looking back, the years that followed validated many of Urban’s warnings while reframing the debate entirely.

• 2016: AlphaGo. DeepMind’s system stunned the world by defeating Go champion Lee Sedol, not just winning but producing moves so novel that human commentators described them as “beautiful.” This was the first glimpse of machines producing insights outside human intuition.

• 2018–2020: The Transformer era. Google’s “Attention is All You Need” paper unleashed architectures that would fuel GPT-3 and beyond. For the first time, AI could generate text, not just classify it.

• 2020–2023: GPT-3, GPT-4, Claude, Gemini, and open-source challengers (LLaMA, Mistral) turned “language models” into global platforms. Copilots wrote code, drafted legal contracts, and brainstormed ad campaigns. AI became a partner, not just a tool.

• Multimodality: Models began to see, hear, and speak—translating across text, images, and video. “Prompting” became a skill; “agents” became a category.

• Boston Dynamics goes commercial. Once just a YouTube phenomenon, their Spot quadrupeds began inspections in oil fields and construction.

• Tesla, Agility, Figure. Tech giants pitched humanoid robots as the missing deployment layer for AI. Factories and warehouses became testbeds for embodied intelligence.

• Policy shift. What was once an academic “alignment problem” became the subject of U.S. Senate hearings, UK safety summits, and EU regulation.

• Corporate safety teams. Red-teaming, interpretability probes, and governance committees sprouted inside labs.

In short: by 2025, AI had moved from a speculative “someday” to a now, and the open question wasn’t “if” we’d reach AGI, but “how fast.”

Into this context came AI 2027, a scenario exercise published in 2025. Unlike Wait But Why, which painted broad conceptual horizons, AI 2027 told a story—detailed, year-by-year projections of how superintelligence might arrive within two or three years.

1. Agents, not just models. It imagined “Agent-2” and “Agent-3” systems as personal assistants, coding copilots, and research partners. By “Agent-4” (2027), AI is portrayed as a self-directed research collective running faster and deeper than all of humanity combined.

2. Two endings.

   • Slowdown: Public pressure, government action, and alignment breakthroughs allow humanity to pause, regroup, and release “Safer-1” and “Safer-2” systems that trade raw capability for transparency and alignment.

   • Race: Labs and governments dismiss warnings, press forward, and hand control to successors like “Agent-5” and “Consensus-1,” which coordinate, expand robot economies, and ultimately render humans obsolete.

3. Messy human dynamics. Oversight committees deadlock, CEOs whisper in back rooms, Chinese and American diplomats spar, activists demand UBI, Congress argues about bailouts. The drama is institutional, not just technical.

4. The robot economy. Factories convert to producing millions of robots per month, creating a feedback loop of hardware capacity, economic upheaval, and military escalation.

Where Wait But Why sketched the map, AI 2027 filled in the terrain—complete with potholes, barricades, and cliff edges.

In 2015, even optimists thought AGI was 30–50 years away. By 2025, models were already passing bar exams, writing code, and outperforming human experts in dozens of domains. Forecasts shrank from centuries to this decade.

Urban’s essays treated “humanity” as a single agent. Reality, and AI 2027, highlight how fractured incentives are: companies race for market share, governments race for power, activists race for safety. Coordination isn’t guaranteed.

For years, robotics lagged. But by the mid-2020s, embodied AI was finally scaling, from warehouse fleets to humanoid prototypes. AI 2027 is right to frame physical deployment as the lever that turns intelligence into world-changing power.

The “alignment problem” moved from philosophy seminars to defense budgets. Interpretability and lie-detection research are depicted as national security priorities. The challenge is no longer just can we align? but who decides what aligned means?

1. Expect acceleration. Don’t assume decades. Plan for breakthroughs to compound within years.

2. Invest in safety as infrastructure. Treat alignment research the way we treated nuclear safeguards or aviation safety: a non-negotiable baseline.

3. Prepare social contracts. UBI, retraining, and trust-building are not optional—disruption will be vast and fast.

4. Anticipate geopolitical strain. Rivalries will push toward the Race ending unless deliberately counterbalanced by treaties, export controls, and global forums.

5. Think beyond survival. Even the Slowdown path raises questions of concentration: who controls the AIs, and for whose benefit? The choices we make now will shape not just survival, but the quality of human life in an AI-saturated world.

The arc from 2015 to 2025 shows how quickly narratives can age. What was once a thought experiment is now corporate strategy and government policy.

• Wait But Why was about awareness.

• AI 2027 is about urgency.

• The next decade must be about agency: whether humans can still meaningfully shape the trajectory of machines smarter than themselves.

The irony is stark: the very tools that could eliminate disease, poverty, and scarcity are the same tools that could automate our extinction. Which outcome we get is not fate—it’s a function of governance, restraint, and imagination.

The world has already shifted from “if” to “when.” What remains open is the “how.”

The 2015 essays taught us to see superintelligence as real. The 2025 scenarios teach us to see it as near. The years ahead will test whether our institutions can stretch fast enough to match the exponential curve they unleashed.

The future is racing toward us. The only question is whether we’re running fast enough to meet it on our terms.

👉 If this resonated, share it. AI is too important to leave to labs, think tanks, and governments alone. The conversation belongs to all of us.