Where technology might be headed -- honestly labeled.

Humanoid robots performing real paid work (not demos) would be the clearest signal yet that general-purpose robotics has crossed from research to industry.

More verified multi-site commercial pilots succeeding, falling hardware costs, AI model improvements in manipulation and planning.

Safety incidents, manufacturing bottlenecks, the gap between demo footage and independently verified deployment numbers.

This estimate would be proven wrong if no additional independently-verified paid commercial deployment (beyond Figure AI's existing one) appears within the stated window.

Restoring a lost sense is a fundamentally different and harder claim than cursor control -- success would meaningfully expand what BCIs are understood to be capable of.

Successful early Blindsight trial results, faster regulatory pathways for compassionate-use devices.

The inherent difficulty and safety bar of implanted neurotechnology, slow multi-year clinical trial timelines.

This estimate would be proven wrong if the Blindsight trial is delayed past 2026 or fails to demonstrate any restored visual perception.

Custom silicon is a direct response to the cost of serving frontier models at scale -- it reshapes who controls the AI supply chain.

Continued inference-cost pressure, chip-partner manufacturing capacity.

Chip design and fabrication lead times, which run in years even once a team is staffed.

Already largely confirmed as a current trend (window: NOW) -- would only be invalidated if multiple major labs publicly abandoned their in-house chip efforts.

The compliance conversation for AI companies operating in the EU has moved from "what will the rules say" to "the rules are live."

Early enforcement actions setting precedent, other jurisdictions following the EU's framework.

Further political pressure for delays, as already happened once with the Digital Omnibus package.

Already confirmed as current fact (window: NOW) -- would only be invalidated if the EU formally suspended the transparency rules it already began enforcing.

Would change how people interact with digital information day-to-day, not just how they play games.

A lighter, cheaper mainstream form factor; a genuinely useful always-on software layer.

Persistent comfort/battery-life limitations, unclear everyday use case beyond gaming and specialized training.

This estimate would be proven wrong if no mainstream (non-gaming, non-enterprise) AR/VR use case reaches meaningful adoption within 3-5 years, or if the category instead shrinks.

Practical quantum advantage would reshape cryptography, materials science, and drug discovery.

A breakthrough in error correction, sustained multi-lab funding.

The error-correction problem has repeatedly proven harder than early timelines suggested.

This estimate would be proven wrong if a verified, independently reproduced practical quantum advantage is demonstrated before 5-10 years, or if progress stalls entirely beyond 10 years.

Would materially change commuting, logistics, and urban planning.

Regulatory approval expanding city-by-city, falling sensor/compute costs.

Safety incidents shifting public and regulatory sentiment, remaining edge-case reliability gaps.

This estimate would be proven wrong if personally-owned, fully driver-optional vehicles remain unavailable at mass-market scale past 5 years, or if a major safety incident triggers a broad regulatory rollback.

The gap between an agent that can chat and one that can reliably act is the current frontier of AI capability.

Better tool-use reliability, longer context and memory, industry-standard safety patterns for autonomous action.

Reliability failures in complex multi-step tasks, the same human-approval gating Project KAI itself uses for anything consequential.

This estimate would be proven wrong if agentic reliability plateaus and industry-wide human-approval gating remains mandatory for all consequential actions past 1-2 years.