The Compression of Time in the AI Era

AI is “compressing time” in a way that changes what work can be accomplished per hour—even if the clock doesn’t speed up. Humans experience time as increasingly scarce because the volume of tasks, updates, and communication keeps expanding. AI, by contrast, doesn’t “feel” time passing; it converts compute gains into faster throughput. The practical result is a widening gap between two things that scale at different rates: raw intelligence and the ability to maintain intent over long stretches.

That mismatch matters because today’s biggest limitation for AI agents isn’t raw capability in the moment—it’s staying aligned with a goal over time. Maintaining intent across days or weeks is still difficult, so the near-term roadmap for agents is cautious. A common projection is that by 2026 an AI agent might be able to work on a task for about a week. For many organizations, that’s meaningful because real projects often require months of strategic alignment and context retention. Humans are generally better at holding onto intent and remembering what matters; AI agents are more prone to forgetting or drifting, even when they perform impressively in shorter bursts.

The intelligence-versus-intent gap is also why “agent progress” can feel slower than model progress. Intelligence scaling is often described as steep and vertical—models get smarter quickly as compute rises. Intent scaling, however, moves more slowly. Yet compute advances keep increasing what AI can do within a fixed time window, effectively expanding the amount of “work” that fits into an hour. The transcript frames this as time expanding for AI work capacity, not time shrinking for the clock.

A concrete case study comes from robotics and simulation work associated with NVIDIA. Jim Fan’s “physical touring test” reframes the classic touring test: it’s not enough for an AI to pass as human in conversation; it must also operate in physical space with human-like competence. The example scenario is a messy room after a hackathon—pizza boxes, beer cans, scattered items, and obstacles like a dog or a tennis ball. The challenge is imagining a robot that can autonomously clean the entire space end-to-end, without direction, safely and reliably. That bar remains far out of reach in real-world robotics.

But simulation can compress training time. In a Sequoia talk context, Fan describes taking what would normally require 10 years of training and compressing it into roughly two hours by using a simulated environment that allows massive parallelization. The key idea is that training can run far faster when the system isn’t constrained by real-world speed limits.

From there, the argument returns to agents and time. If an agent can complete a task that usually takes months in a matter of hours—provided it has the right tools and the task scope is well defined—then limited intent-over-time might still be sufficient for meaningful outcomes. The transcript suggests agents will function like highly capable interns: they can execute large portions of well-scoped work quickly, while humans retain responsibility for architecture, validation, and boundaries. Current examples like Devin are treated as early previews, including known failure modes such as token limits, task ambiguity, and overreach. The bottom line: compute-driven intelligence gains can make “short intent windows” productive, but only when humans design the scope and autonomy carefully—an operational shift that many teams will have to learn quickly as chips and throughput keep accelerating.