The Word Warp Drive

Speed comes from three multiplying trends. Hardware improved about 1.4x/year over decades and is slowing. Training compute spending has grown near 4x/year since 2010. Algorithms often deliver similar performance on much less compute, roughly 3x/year efficiency gains. Multiply them and you get today's curve.

Scaling laws made that spending rational: error fell predictably as compute, data, and knobs rose (Deck 1, Section 2). For users, equivalent capability has trended toward ~10x cheaper per year (task variance high). Exponentials still hit walls—power, fabs, data, or capital. Data is the one already being routed around: the readable internet is finite, so labs increasingly train on synthetic data, text one model writes for the next to learn from. The question is where first.

Flagships set the ceiling, but most work does not need ceiling performance. As costs fall, yesterday's flagship capability fits smaller hulls. Four drivers: distillation, quantization, mixture of experts, and specialization.

Command rule: route each task to the smallest vessel that clears the bar, and reserve flagships for high-stakes judgment. On-device shuttles are instant/private, runabouts are cheap/fast, flagship handles irreversible missions.

Moravec's paradox still bites: symbolic tasks that feel hard to humans can be easier for machines, while toddler-level perception and dexterity remain difficult. Evolution spent far longer on sensing and motor control than abstraction. LLMs learned words first because text data was abundant; there is no internet-scale touch dataset.

Robotics must generate data via simulation, teleoperation, and video. Then similar next-token machinery predicts motor actions (VLA models). It works, but on a slower clock because physical examples consume real time. Self-driving's path from demos to paid rides took about two decades.

The frontier is jagged: strong legal prose, weak letter counting; exam wins, weak freshness without context. Benchmarks saturate and get replaced. Even Turing-test claims triggered argument, not consensus.

Forecasts diverge because AGI lacks one definition. Every date implies a chosen bar. Some bars are already met, some contested, some clearly unmet with frozen knobs (Deck 1, Section 2). Better signals than headlines: task horizon and price trends.

Three plausible futures dominate discussion. Plateau: progress bends and mostly gets cheaper/reliable. Steady climb: current doubling continues and routine work shifts to agents. Fast takeoff: systems improve systems and institutions lag. Reasonable people hold each; nobody knows which wins.

Stoic move: separate control from uncertainty. You can't set the curve; you can set habits that work in all timelines—define done, verify evidence, use small checked steps, hold human boundaries, know your domain, route to smallest sufficient model, and keep practicing.

Words that appear in future-of-AI arguments, with deck addresses:

Four rules hold across all timelines:

• Read curves, not headlines. Multipliers, cost trends, and task horizon are sourced measurements; demos are bets paying off.
• Route to the smallest sufficient ship. Capability flows downhill; reserve flagship use for judgment-heavy work.
• Bodies move slower than words. Robotics progress is data-constrained and on a slower clock.
• Prepare similarly in every timeline. Specify, verify, iterate small, keep human boundaries, know domain, keep learning.

That's the full ship: one guessing core, answer costume, action crew, and command habits riding an uncontrollable curve. You cannot steer the curve; you can steer the ship.