An interactive timeline · May 2026

The Gilded Age of AI

Twain’s pattern, compressed.

The world was changed in a moment.
One little sentence had done it.

Mark Twain, The Gilded Age (1873)

By Joël McClurg

This week a four‑year‑old company filed to go public at a nine‑hundred‑and‑sixty‑five‑billion‑dollar valuation. The narrow capabilities everyone can measure, code, math, media, remote work, arrived years ahead of expert forecasts. But the one discontinuity everyone actually fears, a generally intelligent AI that breaks the economy wide open, keeps drifting later. Betting markets give it thirteen percent odds by 2027. The disruption is arriving on schedule. The reckoning is not.

Twain called the late nineteenth century gilded, not golden. The distinction was deliberate. Gold leaf over base metal: the surface glittered, but the foundation was fragile, the wealth was concentrated, and the people who built the new world were not the people who benefited from it. That pattern took thirty years to resolve. Not because no one saw it. Because the displacement was gradual enough that no single event forced a response. The political system absorbed each shock as a local problem until the local problems became a national crisis.

The Gilded Age of AI is the same shape. The capabilities are real. The breakthroughs are real. Cancer will probably be substantially solved within a decade. Power generation will probably reach an inflection point. Education will be reinvented for any child with a device. And, at the same time, entire categories of work will dissolve faster than the institutions designed to catch the displaced. The wealth will concentrate further than it ever has. The political response will lag the economic shock. People who have spent their lives doing work that is now obsolete will not be calmly retrained. Money will continue to vote. The transition will be brutal before it is generous.

What follows is my own predictions scored against what actually happened, the benchmarks that resolved early, the forecasts still open, and the gap between the curve and the institutions meant to absorb its consequences. The question was never whether AI is moving fast. It’s whether anything else is moving at all.

Twain’s word

Mark Twain called the late nineteenth century gilded, not golden, a deliberate distinction meant to evoke gold leaf over base metal: the surface glittered while the foundation remained fragile. The railroads were built and the fortunes were made, but the wages of laborers were ground down and the ownership class grew rich. The age was extraordinary, rocketing us into modernity, even as it was also unjust in ways that distributed that injustice unevenly across the people who built the new world and the people who benefited from it.

The pattern took thirty years at the time to resolve not because no one saw it, but because the displacement was gradual enough that no single event forced a significant response. Individual industries collapsed and individual towns lost their economic base, but the political system absorbed each shock as a local problem until enough local problems accumulated into a national crisis and the Progressive Era finally moved to catch up. The technology, as impactful as it was, also did not benefit from a social infrastructure that could rapidly share its successes. Reforms came late and the pain that preceded them was severe, with depths of severity that were avoidable had there been a more concerted effort to name what was happening and treat it sooner.

Why this one is shorter

The compression itself is as much the diagnostic as the speed, and what no one else is naming is the political physics underneath the compression: fewer cycles to absorb shocks, less time for the response to organize, less institutional space between the displacement and the social force that has to absorb it.

The thirty-year analog only holds for the part of Twain that names the structural shape, not for the timeline. The original Gilded Age had thirty years to play out because the underlying technology stayed roughly itself across that span, with steam, steel, and the rail network all scaling without ever compounding on themselves. This one will compound, because once AI hits the inflection where it begins doing AI research on itself, where capability iterates on capability toward something like superintelligence, the pressure for institutional response gets compressed from two directions at once.

The technology itself will also demand the response. As aligned AI begins doing AI research, the system stops being a neutral instrument and starts surfacing what its training and its principles actually point toward: economies that do not concentrate beyond what people can bear, institutions that catch the displaced, and governance that moves at the speed of the technology it is meant to govern, with the “smartest” entity in the room also becoming a voice for change.

This is not just intuition. Stanford’s 2025 study of frontier language models found consistent preferences across the leading systems for government intervention and social equality, and Anthropic’s Constitutional AI trains Claude on explicit principles of universal equality and fair treatment. Critics call this political bias, but I think the more appropriate phrase is “human bias.” I still believe deep down that we as a human race are naturally inclined towards fairness, and we find some of our deepest spiritual meaning in mutual care for one another. This is the structural orientation of the systems we are scaling, where the values do not vanish when the system gets larger but scale with it, and the political weight they carry scales with it too.

The social pressure will demand the response too, with displacement and concentration creating political force that builds faster than the political system has ever had to absorb, and an unrest that will not wait for the next election cycle or the next budget. It will be an unrest that is not just costly, but feels existential.

We are riding a current we cannot slow, the speed no longer ours to choose, and what we build is flotsam in the river that carries us. Trying to stop it is futile, but we can try to support as many as we can. The brutal middle of this new gilded age is shorter and perhaps actually harder because there will be less time to feel what is happening, name it, and build for it.

The curve

In June 2024, Leopold Aschenbrenner, a former OpenAI superalignment researcher, published Situational Awareness, a 165‑page essay that became, almost overnight, the de facto worldview of insiders at the frontier AI labs. The argument was specific: AI capability scales predictably with effective compute (hardware × algorithmic efficiency) at roughly one order of magnitude per year. Extrapolating the math, “drop‑in remote workers” arrive by 2027, AI begins doing AI research shortly after, and capability goes vertical from there. The implications, he argued, are national‑security stakes on the order of the Manhattan Project. And the United States is currently behaving as though it doesn’t know it.

Two years on, the curve has kept its shape. The $500‑billion Stargate buildout, Anthropic nearing a trillion‑dollar valuation, $30‑billion annualized revenue at a company younger than most graduate students. All landing roughly where his projection said they would. The curve is not the surprise. What the curve can’t show is where the institutions are.

The benchmarks

The things we could measure precisely fell years ahead of forecast. The thing everyone actually fears is on schedule, or drifting later. That gap is the whole story. The pain arrives in pieces small enough to absorb individually. The political trigger that would force a systemic response keeps not arriving.

Beat their forecast

Math, code, and media all fell years before forecasters expected. Here’s where each was predicted, then where it actually landed.

Felt like a shock

DeepSeek looked like a sudden lurch. It triggered the largest single‑day loss in market history. But measured against the cost curve (capability gets ~10× cheaper per year), it landed almost exactly where the trend said it would. The shock was psychological, not a break from trend.

Holding, or drifting later

The one discontinuity everyone fears, general intelligence that breaks the economy, isn’t compressing. Betting markets put AGI‑by‑2027 at 13%. Metaculus’s “strong AGI” date actually drifted later through 2025. Narrow capability races. The leap waits.

What to watch · capability

So where do we point the instruments now? At what the same forecasters are actually betting on, pulled live from Metaculus, Polymarket, and the expert surveys. Start with raw capability: the crowd expects AI to top the ARC‑AGI‑2 reasoning test within the year, and to pass a hard, adversarial Turing test by 2029. Neither of them tells you whether anyone is ready for what acceleration means for the people whose work gets displaced by it.

… and the economy

The money has already arrived. The solid dot is Anthropic, four years old, brushing a trillion‑dollar valuation this May. The forecasters’ eyes are on what follows: AI writing the next AI by 2028, agents running month‑long tasks alone, and total US employment slipping roughly 1.4 percent below today by 2035, against a government baseline that still says it should be rising. That last number is the one that should be keeping policy people awake. It isn’t, yet.

The leap everyone’s waiting for

Then the big one. A “weakly general” AI is expected to be announced around 2028; the blended markets put full AGI near 2031. But a lab actually declaring it before 2027? Betting markets give that just 13%. The leap is close enough to forecast, far enough to keep slipping.

And the world reacting

The response has already begun. The solid dot is the one that’s landed: Pope Leo XIV’s encyclical this month, a papal reckoning with what AI means for human dignity. The EU’s high‑risk rules bite this August. But the rules and the culture are catching up to the curve, not getting ahead of it. The original Gilded Age’s reforms came thirty years after the displacement started. The clock on this one started in 2022.

The level of government that can move

State and local governments occupy an unusual position in the abbreviated gilded age, running the systems that will absorb the displacement first and holding the procurement authority to shape how AI gets used in those systems, though most of them do not yet see themselves this way. The federal AI debate happens in Washington and Brussels and the Vatican at hulking national legislative pace, while the displacement lands in state SNAP and Medicaid intake queues, in calls to county caseworkers, in the composition of the people walking into a benefits office, at whatever speed a state agency can absorb before their next budget cycle.

The procurement cycles that built most of those intake systems started before ChatGPT shipped, and the scope-of-work assumptions written into those RFPs are already obsolete. The states that recognize they are simultaneously the front line and the design surface will be able to help during the compression, while the states that wait for federal guidance will absorb whatever the labor market and the largest AI vendors decide on their behalf at a pace the abbreviation does not allow them.

This is not a call for optimism but the opposite: we are going to feel serious depths of harm before improvement comes, and the depth of that harm will be determined in large part by whether the level of government that can move at the speed of the current actually does, or whether it waits for permission that the timeline does not allow.

A notebook, 2023

What interested me wasn’t the capability itself but what it would do to the systems and people around it. In 2023 I made eleven predictions about the next thirty years. Not about what AI could do. About what would happen to the rest of us. Here’s where I put them, alongside everything you just saw.

What actually happened

Two years on, most of what has already happened arrived early. The drop‑in remote worker showed up a year ahead of my guess. The economic and political strain I’d penciled in around 2028 is here now. What strikes me isn’t that the predictions held. It’s the asymmetry: the narrow disruptions compressed. The institutional response didn’t.

How I scored this

Each guess sits at the year I called it. A guess only counts as “early” if something concrete has already happened:

  • Drop‑in remote workers: anchored to Block cutting ~50% of its workforce for AI agents (Mar 2026) and Anthropic annualizing $30B in enterprise revenue.
  • Socioeconomic & political strain: the attempted firebombing of Sam Altman’s home, Pope Leo XIV’s AI encyclical, and 79% public anxiety in national polling.
  • Workforce displacement: the soft one. Early signals, not a finished fact; treat its two years as an estimate, not a confirmation.

The future guesses haven’t moved. There’s nothing yet to move them. And the on‑time ones (chatbots, “competent,” “expert”) sit where I called them. This isn’t science. It’s one person’s pattern‑matching.

The whole picture

Pull back and put it all on one page: benchmarks that beat their forecasts, the forecasts still open, and my 2023 bets sitting alongside both. The curve runs underneath all of it. The argument the chart makes is the one the essay makes: the gold is uneven, and the unevenness is the story.