The Bandwidth Gap.

Frontier compute doubles every 5.2 months. Agent task horizons double every seven months and may now be doubling every four. Human conscious cognition, the part that decides, prioritises, and judges, has held steady at roughly ten bits per second for as long as it has been measurable. The capability stack now scales orders of magnitude faster than the people supervising it.

This paper makes four claims. First, the bottleneck of AI-augmented work has moved from the model to the operator, and the empirical record now supports this in a way it did not eighteen months ago. Second, the most consequential gap is not at the sensory or motor periphery but at what neuroscientists have begun calling the inner brain: the slow, deliberate layer where attention, judgment, and commitment happen. Third, closing this gap is an engineering problem, not a self-help one. It calls for a new class of artefact, the cognitive primitive, whose job is to filter, prioritise, and decide alongside the operator rather than for them. Fourth, the differentiator that will matter is not maximal capability per agent but maximal operator leverage per unit of attention spent. Zeno Center calls the property this protects cognitive sovereignty. It is the operating thesis of the lab.

Zeno's first deliverables are inline cognitive primitives for AI product teams, agent developers, and operator-heavy organisations, instrumented for multi-agent supervisor cognitive load. This paper sets out the evidence, the lineage, and the agenda.

AI compounds. Human attention does not.
We build the cognitive layer that closes the gap.

This is a thesis about scarcity. For most of computing's history the scarce resource was machine cycles, and most of the field's progress can be read as a story about making them cheaper. That regime is ending. Frontier language model training compute has been doubling every 5.2 months since 2020 and overall growth is running at roughly 4.5x per year (Epoch AI, 2026). The length of tasks an autonomous agent can complete with reasonable reliability has been doubling on a seven-month cadence since 2019, with the trend appearing to accelerate to roughly four months in the most recent twelve (Kwa et al., 2025; METR Trend Tracker, 2026). The supply of capability is no longer the constraint on AI-augmented work. The supply of human attention is.

This claim is older than it looks. Herbert Simon stated it in 1971 in a now-famous passage on the design of organisations: a wealth of information creates a poverty of attention (Simon, 1971). What is new is that the wealth of information has been transformed into a wealth of agentic output. The artefact pressing against the operator is no longer the news feed. It is a workforce of legible, plausible, often wrong machine collaborators. The economic question of the next decade is whether the operator can keep up. The operating thesis of Zeno Center is that the tools that close this gap will define the next platform of knowledge work, and the companies that build them will hold a position structurally analogous to the one personal computing held in the 1980s and the cloud held in the 2010s.

A picture is sufficient. On one axis, time. On the other, capability. The AI curve is a near-vertical exponential. The human curve is a flat line that has been flat for fifty thousand years. The gap between them is the operating environment of every knowledge worker alive today.

The two curves are not an analogy. They are the result of the only two empirical regimes we have direct measurements of: the scaling laws of frontier models and the throughput of conscious human cognition.

On the AI side, three numbers anchor the trend. Training compute for notable models grows 4.5x per year and for frontier language models doubles every 5.2 months (Epoch AI, 2026). Inference cost on equivalent capability falls by roughly 50x per year at the median, halving roughly every two months (Epoch AI, 2026). Time-horizon-of-task, the longest task an agent can complete with 50% reliability, doubled every seven months from 2019 to 2025, with recent data suggesting acceleration to roughly four months (Kwa et al., 2025).

On the human side, the numbers are older and more stable. Working memory holds approximately four chunks at a time, revised down from George Miller's classic seven (Miller, 1956; Cowan, 2001). Sustained attention on a single screen lasts an average of forty-seven seconds before a switch (Mark et al., 2016; Mark, 2023). Recovery from an interruption takes twenty-three minutes and fifteen seconds (Mark, Gudith and Klocke, 2008). And the load-bearing figure for this paper, recently re-derived in Neuron by Zheng and Meister: human conscious behavioural throughput averages around ten bits per second (Zheng and Meister, 2025).

The compression ratio is approximately 100,000,000 to 1.Sensory intake (~10⁹ bits/s) versus deliberate cognition (~10¹ bits/s).

The paradox is not that we are slow. It is that we have built a peripheral nervous system, our senses and our motor systems, that handles roughly a billion bits per second, and a central deciding layer that handles ten. Whatever the inner layer is doing, it is not bandwidth-matching.

A second arithmetic is worth doing in plain sight. Microsoft's June 2025 Work Trend Index special report finds that the average knowledge worker now experiences 275 interruptions per day, roughly one every two minutes during the standard nine-to-five (Microsoft WTI, 2025). Mark, Gudith and Klocke have established that recovery from a single interruption takes twenty-three minutes (Mark, Gudith and Klocke, 2008). If both numbers held independently and additively, recovery alone would consume approximately 105 hours of working time per day. They cannot both be true at full strength. The implication is therefore the only available one: the contemporary knowledge worker is operating in a state of chronic partial recovery, never reaching baseline cognitive engagement before the next interruption arrives. This is the operating environment AI is being deployed into.

Zheng and Meister's contribution is not the figure itself. The slowness of conscious cognition has been observed at least since Miller (1956). Their contribution is a clean conceptual split. The brain has an outer layer, which handles perception, motor control, and reflex, and an inner layer, which handles deliberate thought. The outer brain is fast, parallel, and subconscious. The inner brain is slow, serial, and the seat of attention.

This split clarifies the AI augmentation question. Most existing AI tooling targets the outer brain. Voice transcription, autocomplete, image recognition, retrieval, and search are all outer-brain prosthetics. They feed perception or extend motor output. They are not where the cognitive load of AI-augmented work lives.

The load lives in the inner layer. It lives in the moment when an operator reads three drafts from three agents and has to decide which one matters. It lives in the moment when a developer reviews a diff that compiles, looks reasonable, and is wrong in a way only the human can catch. It lives in the moment when a product manager has eight Slack threads, four agents running, two meetings queued, and forty seconds before the next interruption.

The inner brain is the bottleneck, and almost nothing in the current AI tool stack is designed for it. The dominant paradigm of 2024 to 2026 has been to give the operator more outputs to inspect, more agents to supervise, more drafts to choose between, more interfaces to switch across. Each of these moves expands outer-brain bandwidth and consumes inner-brain bandwidth. The category most analogous to the inner brain in the current stack is the prompt, which is also the category most under-engineered.

If this analysis is right, the next decade of AI-augmented productivity gains will not come primarily from larger models. It will come from artefacts that are designed for the ten-bit channel.

The first wave of empirical work on generative AI showed dramatic productivity gains.

These numbers are real. They are also incomplete. The same Dell'Acqua paper that reported 12% to 40% gains within the AI frontier reported a 19 percentage point drop in correct solutions on a complex managerial task selected to lie outside the frontier. That drop is the pivot. Each of the celebrated productivity studies was generated under conditions that systematically favour AI: short-horizon tasks, in-frontier problems, novice-skewed populations, lab settings with low task ambiguity. None of them was a fair test of the population that does most of the world's high-value knowledge work.

The fair test arrived in July 2025. METR ran a randomised controlled trial on sixteen experienced open-source developers working on their own repositories. The repositories averaged ten years of development and more than one million lines of code, and each developer had on average five years and 1,500 commits of personal authorship on the codebase they worked in. Each task was randomly assigned an AI-allowed or AI-disallowed condition. The numbers are best read in order:

Before the study, developers forecast that AI would make them 24% faster.

After the study, developers reported feeling 20% faster.

The measured result was a 19% slowdown.

Becker, Rush et al., 2025

The gap between perceived productivity and measured productivity was thirty-nine percentage points. The developers were not just slower. They believed they were faster while being slower. This is a metacognitive failure, not a productivity failure, and it is the single most important finding in the AI productivity literature to date.

Three observations from the METR data deserve to land. First, the effect was robust across twenty alternative explanations the authors tested. It is not an artefact of the specific tools (Cursor Pro and Claude 3.5/3.7 Sonnet at the time), the task selection, or the developer population. Second, the one outlier in the dataset, a developer with more than fifty hours of Cursor experience, did show a positive speedup. The implication is that there is a high skill ceiling on getting useful work out of these tools, and most operators are nowhere near it. Third, the dominant time sink reported by participants was cleaning up AI-generated code, not generating it.

The METR result does not refute the earlier productivity studies. It triangulates with them. The honest synthesis is this: AI substantially accelerates novices on short, in-frontier tasks; it can slow experts on long, ambiguous, off-frontier tasks; and across the population it does so in ways that the operators themselves cannot accurately perceive.

The metacognitive failure is the deeper finding, and it has been replicated across populations and tasks. Lee, Sarkar, Tankelevitch and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers across 936 first-hand examples of AI use at work. The headline result: higher confidence in the AI predicted less critical thinking; higher confidence in oneself predicted more (Lee et al., 2025). Sarkar's companion framing put it in three words: when copilot becomes autopilot (Sarkar et al., 2024). Stadler and colleagues showed that LLM-assisted students achieve cognitive ease at the cost of depth (Stadler, Bannert and Sailer, 2024). Bastani and colleagues ran a randomised study of GPT-tutored students who scored worse than controls on subsequent unaided tests (Bastani et al., 2024). Kosmyna and colleagues at the MIT Media Lab measured EEG connectivity during AI-assisted essay writing and found brain network engagement weakest in the LLM-assisted group, persistently so even after the assistance was withdrawn (Kosmyna et al., 2025).

The MIT Media Lab paper, with senior author Pattie Maes, named this pattern cognitive debt. The Zeno thesis is that cognitive debt is not a side effect to be flagged and managed. It is the central design constraint of the next decade of knowledge work, and the artefact category that pays it down does not yet exist.

Cognitive debt is not an argument against AI. It is an argument for instrumentation, and for a new class of tool.