The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock

TEXT START: Large language model (LLM) agents can solve complex interactive tasks, yet reasoning and tool calls take time.

The Dissection

The paper turns latency from an implementation nuisance into the central evaluation variable. RealtimeBench keeps six classes of games running during inference, exposing the gap between capability measured in paused environments and capability under a live clock.

TORSI then splits intelligence across timescales: a slow teacher studies failures offline, writes reusable instructions, and a fast student executes them without teacher calls during gameplay. The rise from 13.51 to 24.19, across all six game types, demonstrates policy and instruction compression—not general autonomy.

The Core Fallacy

The paper treats temporal organization as if it solves the deeper problem. It solves a bounded deployment bottleneck: how to convert slow deliberation into fast action where situations contain repeatable structure. It does not show that reusable instructions survive open-ended novelty, distribution shifts, adversarial changes, or radically different environments.

Under Discontinuity Thesis logic, the result is worse than the paper’s framing suggests. The teacher is not a human safeguard; it is another AI process. Slow cognition is being compiled into cheap, fast execution. That does not preserve human productive participation. It expands the domains in which human reaction time and deliberation become economically irrelevant.

Hidden Assumptions

• The six game types meaningfully represent broader real-time environments.
• A higher 0–100 progress score corresponds to economically valuable competence.
• Offline teacher work and repeated attempts are available at acceptable cost.
• Instructions that transfer across three other models will remain robust under changing conditions.
• Benchmark improvement reflects durable competence rather than narrow policy compilation.
• Real-world environments contain enough recurring structure for this two-timescale loop to function.

Social Function

Classification: partial truth, transition management, and ideological anesthetic.

The partial truth is real: paused benchmarks overstate agent competence when timing matters. The transition-management function is equally clear: keep expensive reasoning outside the live loop and deploy a cheaper fast actor in its place. The anesthetic lies in the phrase “real-time intelligence,” which foregrounds technical progress while concealing the labor consequence.

This is not empty copium. The engineering result is credible within the supplied evidence—and therefore more dangerous. It offers a mechanism for turning increasingly capable but slow systems into operational substitutes.

The Verdict

This paper is a benchmark correction and a deployment recipe, not an escape from the Discontinuity Thesis. It shows that slow cognition can be precompiled into fast action. That is a temporary engineering moat for the system, not a human survival moat.

If generalized, TORSI strengthens Cognitive Automation Dominance and accelerates Productive Participation Collapse. The live clock does not save the worker. It disqualifies whoever cannot act before the state changes.