The Concentration of Attention and Intelligence

Executive Summary

Two Concentrations, One Loop

For thousands of years, attention was governed by the rhythms of daily life — family, work, community, season. That governance held in principle until the smartphone, and it has eroded sharply since. Highly personalized machines now compete for human attention at a scale and precision no prior medium has matched.

The intelligence shift has been equally abrupt. Frontier AI models can now outperform most humans on standardized IQ measures — a threshold crossed in less than four years of exponential capability growth. The same underlying force driving both — cheap, powerful computation at massive scale — has concentrated machine intelligence into a handful of systems controlled by a small number of firms. The two trends are not parallel. They are a single self-reinforcing loop, and we are only beginning to understand what it costs to live inside it.

Section 1

A History of Attention

Human attention has always been worth competing for. For most of history, the competitors were each other — a storyteller, a preacher, a friend, a market. The exchange was reciprocal: attention freely given, something offered in return. The trouble begins when the mechanisms of capture become more powerful than any one person's capacity to resist them.

Attention in the Square

In ancient Greece and Rome, public spectacle — theater, gladiatorial contests, the agora — was the dominant medium for shared attention. These events were scheduled, physical, and finite. Attention was borrowed for an afternoon and then returned. For most of human history, leisure attention was local, voluntary, and bounded. The village storyteller, the traveling minstrel, the church service — all required the audience to choose to show up. Attention was largely governed by the rhythms of daily life and social obligation, and when the event ended, it was over.

Attention in the Living Room

The first industrial-era mass-attention technology was print — newspapers and magazines, which by 1900 reached tens of millions of homes. Radio deepened the shift: by the 1930s, receivers sat in living rooms across America, delivering scheduled entertainment directly into domestic life. Television accelerated the trend dramatically. By 1960, the average American household watched 5+ hours of television daily. Entertainment had fully migrated indoors — and with it, attention became sedentary, home-based, and largely passive. You watched what the networks scheduled.

Attention in the Pocket

Personal computing and the early internet briefly reintroduced agency: you navigated to websites, chose what to search, built your own media diet. But the smartphone changed the geography of attention permanently. The screen moved from the living room to the pocket — and from the pocket to the bedside table. For the first time in history, the attention-capture infrastructure was present at every waking moment: at meals, in bed, in line, mid-conversation. Attention was no longer something you lent to a screen at scheduled times. It had become continuously available to whoever could reach it.

Section 2

Active vs. Passive Screen Time

Screens entered the living room in the 1950s, migrated to the computer desk in the 1980s and 90s, and arrived in the pocket in the 2000s. But across that journey, a more important distinction emerged — not where the screen was, but how it was being used. There are really two forms of screen consumption, and they are not equivalent.

Active Consumption

Active consumption is screen time with an agenda. Browsing a specific topic, writing an email, reading an article you sought out, creating a document, building something, searching for an answer — these are all forms of active engagement where the user is in the editorial seat. The screen is a tool, and the user is directing it toward a purpose. Active consumption tends to reach natural stopping points: the task is done, the question is answered, the email is sent.

Passive Consumption

Passive consumption is screen time without an agenda. The user opens a feed not to find something specific, but to be entertained — to see what's there. In this mode, the platform makes the editorial decisions. A recommender system predicts what the user is most likely to engage with next and serves it automatically, one item after another, without pause or natural end. The user isn't really choosing; they're receiving.

For a deeper look at how recommender systems work mechanically — and how they shape adolescent formation — see our study on What Shapes Children.

Negative Impacts of Passive Time

The shift from active to passive screen time is not abstract. It is producing a documented set of effects at population scale:

• Attention fragmentation. Average session lengths on recommender platforms are shorter than user-directed consumption, but sessions are more frequent and harder to terminate voluntarily. The Microsoft Research "attention span" study (2015) documented the first sustained decline in measured attention span coinciding with smartphone penetration.
• Emotional amplification. Content that evokes strong emotion — particularly anxiety, outrage, and social comparison — is consistently over-represented in recommender outputs relative to user-stated preferences. Internal documents from multiple major platforms have confirmed this divergence between what users say they want and what engagement optimization surfaces.
• Formation effects. For adolescents, whose identity and worldview are actively being shaped, the shift to recommender-directed screen time means that a substantial share of formative content is now curated by an engagement optimizer rather than by parents, teachers, or peer communities.
• Sleep displacement. The presence of the recommender in the bedroom — enabled by the smartphone — has measurably compressed sleep duration and quality. The infinite scroll has no natural end; the clock does not.

Section 3

A History of Machine Intelligence

Machine intelligence has exploded in capability at a pace that has surprised even the researchers building it. Peak supercomputer performance grew roughly 350,000× between 2000 and 2024. But raw compute is only part of the story — the more consequential shift was the transition from machines that could simulate physical phenomena to machines that could approximate human cognition. The transformer architecture (2017) and large language models marked the turning point: for the first time, systems could reason, write, and respond in ways that were practically indistinguishable from human output at scale.

Frontier AI models can now outperform most humans on standardized IQ measures — a threshold crossed in less than four years. No prior technology has replicated the cognitive capabilities humans considered uniquely their own this quickly, or at this scale.

For the full arc of machine versus human intelligence — including the IQ crossover data and what it means — see our companion study on Machine vs. Human Intelligence.

Section 4

Putting Intelligence to Work — The Human Equivalent Frontier

Raw compute is one measure of AI capability. A more honest measure is what AI agents can actually do: the length and complexity of tasks they can complete autonomously, with no human in the loop. As of May 2026, the 4-hour milestone is already behind us. The table and chart below project when each subsequent milestone falls, under three growth scenarios.

The implication is not that AI replaces every human worker. It is that the line between what requires a human and what does not is moving — on a predictable schedule, and faster than most institutions are prepared for. Within the planning horizon of most organizations, tasks that today require weeks of expert judgment will fall inside the autonomous frontier. The question is no longer whether the line moves. It is what kind of life remains on the human side of it.

Section 5

The Flywheel

The two concentrations — intelligence and attention — are not parallel trends. They are a feedback loop. Advances in machine intelligence manifest first as better recommender systems: more personalized, more predictive, harder to disengage from. Better recommenders capture more attention. More attention generates more revenue — Google and Meta alone produced over $300 billion in advertising in 2024. That revenue flows into AI infrastructure, which produces the next generation of models, which produces the next generation of recommenders. The cycle repeats.

Each turn tightens it. The firms with the most compute build the best recommenders. The best recommenders generate the most revenue. The most revenue funds the most compute. The result is a self-reinforcing system, tightly held by a small number of firms, that concentrates machine intelligence and human attention at the same time — and grows more concentrated with every iteration.

Each turn tightens the loop. The cycle has no off-switch — and the firms holding the wheel are the same in every position.

Section 6

Conclusion and What We Can Do

The concentration of attention and machine intelligence are the two most consequential structural shifts of our era — and they are not separate forces. They are a single self-reinforcing loop, accelerating together, controlled by a small number of firms, and touching nearly every waking hour of modern life. Seeing the loop is the first step toward responding to it.

This is not an argument against technology. Intelligence tools are extraordinary, and the ability to direct our attention freely remains ours to keep. But the systems shaping both are optimized for engagement and capability growth — not for human flourishing. The question is no longer whether these technologies will keep advancing. They will. The question is whether we will invest the same energy in shaping their impacts as we do in building them. For practical next steps, see the EconFaithAI Solution Map.

No single actor reverses these trends alone. But every actor has leverage — and the loop has no off-switch we did not build ourselves.

Appendix A

Methodology

Attention measurement

Leisure screen hours per day is the per-adult average of discretionary (non-work) time spent on any screen-based medium, drawn from BLS ATUS, Nielsen Total Audience Reports, Common Sense Media, Kaiser Family Foundation, and Pew Research. Pre-1950 values are reconstructed from radio-listening surveys (BBC/CBS audience research, RCA penetration data).

Recommender-driven hours is the subset of leisure screen time spent on platforms where an algorithmic recommender determines what the user sees next. The share is derived by mapping each platform's daily-active-user share of leisure screen time to whether its dominant interaction mode is algorithmic. A user browsing Wikipedia (search-driven) does not count as recommender; the same user watching auto-play YouTube does. The dominant recommender platforms by era:

• 1996–2005: Early Yahoo personalization, basic Google ranking, Netflix DVD recommendations. Cumulative share < 10%.
• 2006–2010: YouTube recommendations, Facebook News Feed (EdgeRank 2009), Pandora/Spotify radio. Share rises to ~25%.
• 2011–2016: Instagram, Snapchat, Pinterest, Twitter algorithmic timeline, Netflix streaming. Share rises to ~48%.
• 2017–present: TikTok For You Page, Instagram Reels, YouTube Shorts — all major platforms algorithmic-first. Share crosses 50%.

Intelligence measurement (Rmax)

Rmax is the maximum sustained performance measured by the LINPACK benchmark, reported in PFLOPS (10¹⁵ floating-point operations per second). The benchmark is the official metric of the TOP500 list, maintained since 1993. This report uses the #1 system's Rmax from the November list of each year, starting in 2000.

LINPACK measures dense linear algebra (fp64), which is relevant to AI training but underestimates real-world AI compute — AI accelerators have much higher fp16/bf16 throughput than their LINPACK numbers suggest. The trajectory is directionally correct; the order-of-magnitude growth is real.

Appendix B

Limitations

On attention

1. Recommender share is a judgment call. A platform like Spotify has both user-initiated playlists (non-recommender) and Discover Weekly / auto-play (recommender). The split is estimated, not directly measured. A ±5-point error band on the recommender share is appropriate.
2. Leisure screen time includes overlapping consumption. Multi-tasking inflates the headline number; a user watching TV while scrolling a phone counts as both. The ~7 hours/day figure is best read as "media exposure," not "discrete attention units."
3. Per-adult averages obscure distribution. Teens average closer to 9 hours/day (CSM 2021), older adults closer to 4 hours/day. The aggregate number is a population mean.
4. Pre-1950 data is reconstructive. Radio-listening surveys were less standardized than modern Nielsen data. Pre-1950 values should be read with wider error bars.

On intelligence (Rmax)

1. LINPACK is a single benchmark. Real AI training uses a different mix of operations (mostly matrix multiplication in lower precision). LINPACK PFLOPS overstates fp64 capability vs AI-relevant fp16/bf16 capability; AI accelerators have much higher fp16 throughput than their LINPACK numbers suggest.
2. Rmax tracks #1 system, not total compute. Aggregate global compute is much larger and grows faster than the #1 system. The Rmax curve is the leading indicator, not the total.
3. Pre-2000 values are not shown. TOP500 begins in 1993; values before 2000 are not visually meaningful on the chart's scale.
4. The cost-and-concentration story is not captured by Rmax alone. A more complete picture would also show $-per-PFLOPS over time (which has fallen by ~10,000× since 2000) and the geographic/institutional concentration of where these systems are operated.

What this report does not claim

• Does not claim recommender systems are inherently harmful (some — Netflix's recommendations, Spotify Discover — are widely liked).
• Does not claim supercomputer concentration is inherently harmful (compute drives medicine, weather, science).
• Does not predict trajectories beyond 2024.

What this report does claim

• The share of leisure screen time driven by algorithmic recommenders crossed 50% in 2017 and reached 56% in 2024.
• Peak supercomputer performance grew roughly 350,000× from 2000 to 2024.
• The two trajectories are linked — recommender capability is funded by and made possible by the underlying compute infrastructure.

Appendix C

References

Attention measurement

• Bureau of Labor Statistics. American Time Use Survey.
• Nielsen. Total Audience Report, quarterly series.
• Common Sense Media. (2021). Census of Media Use by Tweens and Teens.
• Rideout, V. J., Foehr, U. G., & Roberts, D. F. (2010). Generation M2. Kaiser Family Foundation.
• Pew Research Center. (2024). Teens, Social Media and Technology.
• Gomez-Uribe, C. A., & Hunt, N. (2015). The Netflix Recommender System: 80% of Watch Time. ACM TMIS.

Intelligence (Rmax)

• TOP500 Supercomputer Sites. Semi-annual list since 1993.
• Dongarra, J. J., Luszczek, P., & Petitet, A. (2003). The LINPACK Benchmark: past, present and future. Concurrency and Computation: Practice and Experience.
• Strohmaier, E., et al. (2005). Recent trends in the marketplace of high performance computing. Parallel Computing.

Companion projects

• TechConcentration — wealth concentration of the tech sector.
• TechReligiousProfile — religious composition of tech leadership and workforce.
• ChildhoodInfluence — share of formative influence on children.
• MachineVsHumanIntelligence — frontier-AI IQ trajectory and population scale.
• FruitOfSpiritTech — spiritual audit of 124 technologies.
• Companion visualization at concentrations

Appendix D

Project Files