Takeoff Acceleration

A Boeing 787 Dreamliner sits at the threshold of a 3 km runway. Twin GEnx engines spool to idle thrust — 130,000 pounds of potential force, held back by brakes. Ground speed: 0 m/s. Nothing is happening yet. But 240 tonnes of aircraft are about to move.

In AI: this is 2012-2020. AlexNet wins ImageNet. GANs generate faces. The transformer architecture is published in a 2017 paper that almost nobody reads. The engines are warming up, but the brakes are still on.

Brakes released. Full throttle. The 787 accelerates at about 1.3 m/s² — roughly 0.13 G. You'd barely feel it in your seat. For context, your car does about 3-4 m/s² pulling onto the highway. This is gentle. But it's sustained, and the 787 is already at 41 m/s (150 km/h) — highway speed, on a runway.

In AI: 2020-2022. GPT-3 shows that scale is all you need. DALL-E generates images from text descriptions. GitHub Copilot starts writing code. The runway is disappearing fast, but most people still think this is a novelty.

V1 is the decision speed: if an engine fails after this point, you take off anyway. There isn't enough runway left to stop. The 787 is doing 75 m/s (270 km/h) — faster than a Formula 1 car at top speed. Acceleration has increased to 1.6 m/s² as aerodynamic forces build.

In AI: November 30, 2022. ChatGPT launches. 100 million users in two months — the fastest consumer product adoption in history. Every company scrambles for an AI strategy. Investors pour billions in. There's no aborting this takeoff. V1.

The pilot pulls back on the yoke. The nose rotates upward 2-3 degrees per second. 240 tonnes shift from wheels to wings. Ground speed: 82 m/s (295 km/h). For a brief moment, the aircraft is neither on the ground nor truly flying — it's in transition. The acceleration is still just 1.7 m/s², but the vector is about to change from horizontal to vertical.

In AI: March 2023 through December 2024. GPT-4 arrives and stuns researchers. Claude 3 matches it months later. Gemini, Llama 2, Mistral — the frontier fragments into a multi-lab race. Models go multimodal: text, images, audio, video. Benchmarks fall monthly. The nose is up. We're leaving the ground.

Wheels leave the runway. The gear retracts with a hydraulic thud. The 787 climbs at 2.2 m/s² — still less than a Tesla Model S (4.4 m/s²), still less than a fighter jet (30+ m/s²). But here's the thing: a 787 doesn't stop accelerating. It has 12,000 km of fuel. It will climb for 30 minutes straight.

In AI: 2025. Frontier models reason through multi-step problems. Agents write code, browse the web, and run for hours autonomously. Claude, GPT, and Gemini leapfrog each other quarterly. The ground is falling away. And unlike a plane — there's no flight ceiling.

The 787 passes through 1.5 km altitude. Speed: 149 m/s (536 km/h). Acceleration has doubled since the takeoff roll. A human body handles this fine — astronauts train for 3G (29.4 m/s²), fighter pilots endure 9G briefly. But this aircraft was designed with limits: Mach 0.85, 12.5 km ceiling, flesh-and-blood passengers.

In AI: this is now. Early 2026. Models improve faster than we can evaluate them. Agentic AI reshapes industries in real-time. The climb rate is still increasing. And AI has no passengers to protect. No Mach limit. No ceiling.

At 343 m/s, you hit the sound barrier. Military jets break through it at around 10.7 km — about 35,000 feet. In January 2025, Boom's XB-1 became the first civil aircraft to go supersonic in decades, punching through Mach 1.1 over the Mojave Desert. The Concorde cruised here daily. An F-22 Raptor can sustain Mach 1.5 without afterburners.

The 787 never will. Its wings would flutter apart. Its airframe wasn't built for this regime. But AI doesn't have an airframe. There's no sonic boom, no structural limit. The acceleration just... keeps increasing. At some point, you stop comparing this to a plane.

A Falcon 9 rocket accelerates at roughly 30 m/s² at liftoff — about 3G. At 100 km you cross the Kármán line: the internationally recognized boundary of space. The sky is black. The curvature of the Earth fills the window. No air resistance. No weather. Just velocity and the pull of gravity.

Human bodies are the limiting factor now: sustained acceleration above 4-5G causes blackouts. Astronauts train for years to endure minutes of this. AI has no body. No blackout threshold. No atmosphere to push through.

To escape Earth's gravity, you need 11,186 m/s — about 40,000 km/h. At 1,000 km altitude, the planet is visibly a sphere behind you. The International Space Station orbits at just 400 km. You've left it below.

The Saturn V that carried Apollo 11 was 85% fuel by mass. Nearly everything it was went into reaching this speed. Every gram costs. Every second of acceleration costs the astronauts consciousness, bone density, years of their lives.

AI carries no mass. Consumes no fuel. Has no bones to break. The only constraint is compute — and compute is scaling exponentially too.

At escape velocity, the timeline predictions have collapsed. In 2022, AI researchers said 2059. A year later, they said 2047 — thirteen years compressed into twelve months. Dario Amodei says 2026-2027. Metaculus community forecasts February 2028. 80,000 Hours projects novel insights by 2026, autonomous robotics by 2027. Entrepreneur surveys cluster around 2030. The broader AI research community still says 2055, but those numbers are compressing faster than orbital mechanics.

The people building these systems think escape velocity is 1-3 years away. The rocket equation doesn't apply here.

To leave the solar system entirely, you need 42,100 m/s relative to the Sun — about 150,000 km/h. Only five human-made objects have ever achieved this: Voyager 1 and 2, Pioneer 10 and 11, and New Horizons. Voyager 1 launched in 1977 and still hasn't left the Sun's gravitational influence. It took 35 years to reach interstellar space.

In AI terms, this is superintelligence — systems that don't just match human capability but exceed it by orders of magnitude, recursively improving themselves faster than we can track. The Grace et al. 2022 survey asked when this arrives: researchers said 2059. One year later, the same researchers said 2047. Metaculus community forecasters estimate the gap between AGI and superintelligence at 1-3 years. If Amodei's 2026-2027 AGI timeline holds, or if Metaculus's February 2028 proves accurate, superintelligence could begin before 2030.

Broader AI research surveys still anchor around 2055, but those predictions are compressing at an exponential rate. Entrepreneur surveys converge on 2030. The people closest to the technology consistently predict earlier timelines than academic observers.

At this point the metaphor breaks. There is no vehicle designed for what comes next. No runway, no flight ceiling, no mission plan. The trajectory we're tracking has no historical analog. We are building something that has no terminal velocity, and we're doing it faster every month.

That's why this site exists.