Tracking humanity’s path to AGI

Where value accrues when the marginal cost of cognitive work approaches zero.

• Productivity. Gains concentrate in tasks with fast, cheap verification. Where checking the work costs as much as doing it, the gain largely disappears.
• Employment. Displacement arrives task by task rather than job by job, which makes it hard to see in aggregate statistics until it is well advanced.
• Inequality. Returns flow to compute, data and distribution. Each is more concentrated than the labour it substitutes for.

Acceleration within human-framed questions, and the unresolved question of who frames them.

• Discovery rate. Structural biology and materials screening show real compression of the hypothesis-to-candidate loop.
• The physical bottleneck. Experiments do not accelerate because the proposer did. Wet-lab throughput, not idea generation, now binds.
• Replication. Validation of AI-originated findings is thin relative to the volume of claims, and the literature is biased toward successes.

Augmentation and displacement occurring simultaneously in the same occupations.

• Task decomposition. Jobs are bundles of tasks. Automation unbundles them, and the residual bundle is often worse work, not less work.
• Verification labour. Checking machine output is emerging as a distinct and undervalued category of work.
• Entry-level erosion. The tasks used to train junior practitioners are the most automatable, which threatens the pipeline that produces senior ones.

Epistemics, education and the cost of producing plausible content falling to zero.

• Epistemics. When generating a convincing claim costs nothing, provenance becomes the scarce good and verification infrastructure becomes load-bearing.
• Education. Assessment designed around the production of text stops functioning. Replacing it is slow institutional work.
• Access. Expert-level guidance becomes broadly available — the clearest unambiguous benefit in this section.

Compute as strategic infrastructure and the coordination problem it creates.

• Compute concentration. Frontier training runs are feasible for a small number of actors, making capability a function of infrastructure access.
• Export control. Hardware restriction is the primary lever available to states, and its effects are measured in years.
• Military application. Autonomy in targeting and decision support advances faster than the doctrine governing it.

Oversight, control and the failure modes that do not announce themselves.

• Oversight. Human review is the assumed safeguard, and it degrades exactly as capability rises — precisely when it matters most.
• Silent failure. Systems more often produce plausible wrong work than halt. This is the property that makes autonomy expensive.
• Concentration. The nearer-term risks are about who controls capable systems, not about the systems themselves.