Analysis · convergence estimate
Estimated AI‑Human Convergence
Based on current AI development trajectory
The date on which AI systems match or exceed human performance across most cognitive and physical work, projected from the milestone record below. Each milestone that arrives ahead of schedule pulls the date forward, and the countdown reads the current value.
Two outcomes
The same trajectory, two endpoints
Every capability gain moves the date and widens the range of outcomes on the other side of it. The technical path is the same in both cases. What differs is alignment, who owns the systems, and how fast institutions adapt.
Downside
Unaligned systems, concentrated ownership
Unaligned systems amplify the worst of what people already do: autonomous weapons, surveillance at population scale, decisions nobody can audit. Automation displaces labour faster than institutions adapt, and the gap widens between the people who own the systems and the people displaced by them.
Upside
Aligned systems, distributed surplus
Aligned systems drive down the cost of research, medicine and physical work. Disease that is expensive to study becomes cheap to study. Work that was scarce becomes abundant, and the surplus can reach the people who need it.
The window to shape which of these we get closes before the date, not after it.
Milestone record
Milestones on the trajectory
Fourteen events, dated and tagged by capability domain. Each carries the number of days it removed from the countdown when it was scored. Entries dated after today are projected and marked; they hold their place in the model until the record confirms or replaces them.
- 2024-01 embodiment
Neuralink first human implant
The first human patient receives a Neuralink brain-computer interface and moves a cursor by thought.
countdown delta −25 d
- 2024-02 vision
OpenAI Sora
Sora generates minute-long photorealistic video from text. Recorded and synthetic footage become hard to tell apart.
countdown delta −10 d
- 2024-03 compute
NVIDIA Blackwell B200
The Blackwell B200 delivers about 20 petaflops of AI compute per chip and makes trillion-parameter training practical.
countdown delta −15 d
- 2024-05 science
AlphaFold 3
DeepMind's AlphaFold 3 predicts the joint structure of proteins, DNA, RNA and ligands at near-atomic accuracy.
countdown delta −12 d
- 2025-05 language
Claude 4
Anthropic releases Claude 4 with extended thinking, long-running agent workflows and deep research.
countdown delta −20 d
- 2025-08 language
GPT-5
OpenAI ships GPT-5 with stronger reasoning and unified multimodal input.
countdown delta −18 d
- 2025-09 embodiment
Figure 02 general-purpose robot
Figure's humanoid runs multi-step tasks on its own in a warehouse setting.
countdown delta −22 d
- 2025-11 autonomy
Waymo driverless nationwide
Waymo extends driverless ride-hailing across the major US metro areas without safety drivers.
countdown delta −14 d
- 2026-01 embodiment
Boston Dynamics Atlas
The electric Atlas reaches human-level dexterity and locomotion for industrial work.
countdown delta −20 d
- 2026-03 language
Gemini 3 Ultra
Google DeepMind's Gemini 3 reasons across text, image, audio and video natively, at expert level in most tested domains.
countdown delta −16 d
- 2026-05 embodiment
Robot foundation models generalise
One humanoid foundation model transfers learned skills across different robot bodies. Per-robot programming gives way to portable, learned dexterity.
countdown delta −18 d
- 2026-07 embodiment
Humanoid cost curve breaks $20K
Mass-manufactured humanoids from Tesla, Figure and Unitree approach appliance pricing. The labour-substitution arithmetic flips for whole job categories.
countdown delta −16 d
- 2026-09 embodiment projected
Mass robotic adoption
Humanoid and task robots move from pilots to payrolls across logistics, manufacturing, retail and services at once. This is the physical-world inflection point.
countdown delta −30 d
- 2026-10 autonomy projected
Automation price shock
As the marginal cost of physical work falls, robotic substitution shows up as visible price decreases across goods and services.
countdown delta −22 d
14 milestones · 2 projected
Trajectory
Year by year
The record read as a sequence: consumer adoption, then perception and bodies, then reasoning and autonomy, then agents and mass deployment. The convergence window follows from the pace of the last three.
ChatGPT reaches 100M users
The fastest consumer adoption on record. Public awareness of what these systems can do changes in a season, and every industry roadmap is rewritten around it.
Multimodal models, video generation, humanoid robots
Systems learn to see, hear, generate video and control bodies. Foundation models meet robotics, and machines begin to perceive and act in the physical world.
Expert-level reasoning, autonomous vehicles, brain-computer interfaces
Models reach expert-level reasoning in scientific domains. Brain-computer interfaces enter clinical trials. Self-driving vehicles become ordinary in major cities.
Agentic AI and mass robotic adoption
Agents run long, multi-step workflows without supervision. Humanoid and task robots cross from pilots to mass adoption in the second half of the year, and the falling cost of physical work starts to show in prices.
Economic strain and automation backlash
Automation displaces labour faster than institutions adapt. Economic strain follows, and the rift widens between displaced populations and the people who own the automation. This is the stress test that decides whether the transition is managed or chaotic.
Convergence window
The projected window in which AI capability matches or exceeds human cognition across most domains. On the current trajectory the date is moving earlier, not later. The decisions made before this point determine whether convergence is collaborative or catastrophic.
Status
The clock, as of today
How we build AI now determines which future we get. The trajectory is set by the milestones above, and the next ones will move the date again.
Build carefully. The date moves either way.