We keep asking whether the machines have caught up to us. The better question is caught up to us at what — because the honest answer is already strange, uneven, and very American.
Sometime in the last couple of years, a quiet thing happened: the people who build artificial intelligence stopped agreeing about what they had built. One famous chip executive said flatly that we have already reached artificial general intelligence.1 A roomful of Stanford faculty predicted, just as flatly, that there will be no AGI this year at all.4 Both camps are looking at the same models. Both are serious. They simply disagree about where the finish line is — and whether the runner has crossed it.
That disagreement is the real story, and it's a more interesting one than the headlines. The truth, as far as I can read it, is that we are no longer in the land of "narrow" tools that each do one trick — but we are also nowhere near a single machine mind that thinks across everything the way a person does. We are somewhere odd in the middle. I've come to think of it as Segmented AGI: intelligence that has already blown past the best humans alive in a handful of narrow lanes, while remaining clumsy, forgetful, and strangely childlike the moment it steps outside them.
This essay maps that middle ground. First, what today's AI actually is under the hood. Then why the experts can't agree — including the part nobody likes to say out loud, which is that some of the loudest voices have billions of reasons to talk a certain way. Finally, the horizon beyond AGI: self-improving machines, the race between America and China to build them first, and the genuinely open question of whether that finish line is a promised land or a cliff.
A note on wordsYou'll see a lot of initials in this piece — AGI, ASI, HLMI, HRSRAI. I define every one of them in the glossary at the end. Where a term is my own shorthand rather than the field's, I say so.
Strip away the marketing and a modern chatbot is doing something almost embarrassingly simple, billions of times very fast.
The systems people mean when they say "AI" today — ChatGPT, Claude, Gemini, and their kin — are large language models, or LLMs. They are a kind of generative AI: software trained on a staggering amount of text (and now images, audio, and code) to learn the statistical shape of human expression, then asked to produce more of it.6 At their core they are prediction engines. Given everything written so far, they estimate the most likely next chunk of text — a word, part of a word — and then do it again, and again, until an answer has assembled itself.
A language model doesn't "know" the answer. It scores the likeliest next word — then commits, and repeats.
Out of that humble loop comes something that genuinely feels like reasoning. The newest "reasoning" models pause to work through problems step by step, score graduate-level science questions in the same range as human PhDs, and resolve real bugs inside real software repositories.9 They write passable poems, draft contracts, debug code, and explain quantum mechanics to a ten-year-old without breaking a sweat. The breadth is real, and it is new.
But the same systems will also confidently invent a court case that never happened, lose track of what they said three paragraphs ago, and flunk a problem a child would find obvious — because, critics argue, they are matching patterns in their training rather than building a stable model of how the world actually works.3 That gap, between dazzling fluency and brittle understanding, is the crack that the entire AGI debate falls into.
Fluency is not the same as understanding — but it is a remarkably good costume for it.
There is no agreed-upon definition of "intelligence." So there can't be an agreed-upon scoreboard.
Artificial general intelligence has been the field's holy grail since its founding — and the field has never settled on what it means.1 Roughly, AGI is software that can handle essentially any cognitive task a person can, at least as well as an average capable adult. But "intelligence" itself has resisted clean definition since Alan Turing dodged it in 1950 by proposing his famous imitation game instead. If you can't define the target, you can't say when an arrow has hit it. AGI is a moving goalpost, and everyone is holding a different tape measure.
To escape the yes-or-no shouting match, Google DeepMind proposed a tiered scale — five levels of generality, from "emerging" up to "superhuman." It's now widely used, and it's the cleanest way to see why people talk past each other.
Equals or slightly beats an unskilled human across many tasks. Today's mainstream chatbots live here — sparks of generality, but inconsistent.
Reaches the median skilled adult on most cognitive tasks. Not yet achieved broadly — though 2026 models brush it in coding and math.
Top 10% of skilled adults. Reached in isolated lanes; nowhere near it across the board.
Top 1% of humans across most cognitive domains. What most people picture when they say "real AGI."
Beats 100% of humans at everything, including things no human can do at all. Purely theoretical today.
By this framework even the strongest 2026 models rank as "emerging" overall — Level 1 — while hitting Level 3 or 4 in narrow slices like coding.5
Plot the serious voices and they cluster into a few distinct schools of thought. Here is the lay of the land, stated as fairly as I can manage — including the camp I find myself in.
By any reasonable human standard, they argue, current LLMs already are general intelligence. No single human knows everything either; breadth plus depth across language, math, science and reasoning is enough. The debate, they say, only sounds shocking because people confuse AGI with superintelligence.
Who: Four UC San Diego scholars writing in Nature; Nvidia's Jensen Huang.2
Not AGI yet, but close — a few years out. The labs building the models forecast a "country of geniuses in a datacenter" and systems smarter than a Nobel laureate across many fields by 2026–2027.
Who: Anthropic's Dario Amodei & Jack Clark; OpenAI's Sam Altman; DeepMind's Demis Hassabis (more cautious).8
LLMs are sophisticated statistical mimicry, not minds. Real general intelligence needs a different architecture — one that models the world rather than autocompleting it. A 2026 survey of academic AI researchers found more than 84% doubt that today's neural networks alone can get us to AGI.
Who: Gary Marcus, Yann LeCun, Melanie Mitchell, and most of academia.3
The yes/no question is the wrong one. Functionally, AI is already reshaping work and is flatly superhuman in a growing list of narrow domains — while remaining sub-human at integration, memory, and judgment. The spectrum isn't a single dot. It's a jagged profile. (My framing — defined below.)
Who: the "functional AGI" view in the trade literature10 — and me.
Here is the uncomfortable part. The people telling us how close AGI is are, very often, the people selling it.
There is nothing sinister about a founder believing in their product. But when the same handful of executives both define the milestone and profit enormously from the world believing the milestone is near, a reader is entitled to a raised eyebrow. The pioneer Yann LeCun puts it bluntly: don't take CEOs at their word, because they have a vested interest in promoting the capabilities of what they sell.11 The critic Gary Marcus has asked, just as bluntly, whether a lab chief forever placing AGI "two to three years away" might simply be talking up his company's valuation.3
The incentives cut in two directions at once, and that's what makes them slippery. Hype the upside and you raise money, recruit talent, and justify a sky-high valuation. Hype the danger — "this could end the world" — and you cast your company as the responsible adult who must be the one to build it, while making the technology sound powerful beyond measure. Both stories sell. Neither is necessarily false. But both pay.
What's on the table in 2026As OpenAI moved toward a blockbuster public offering, reporting surfaced that its CEO held personal stakes — reportedly worth over $2 billion — in dozens of companies that had discussed or struck deals with the lab he runs. The U.S. House Oversight Committee opened an investigation, and several state attorneys general weighed in.12
And watch what happens when a prediction meets a balance sheet. For a year, leaders warned that AI would erase vast swaths of white-collar work — one CEO floated wiping out half of entry-level office jobs. Then, as IPOs approached, the same voices softened: maybe automation will expand the work people do after all.13 Perhaps they simply updated on new evidence. Perhaps a doom forecast reads badly in a prospectus. A careful citizen can hold both possibilities at once.
When the same person draws the map and sells you the trip, read the map twice.
None of this means the technology is fake — it plainly isn't. It means the timeline and the framing are marketing as much as measurement. Which is exactly why I'd rather look at what the machines verifiably do than at what their makers promise they'll do next quarter.
If you stop asking "is it AGI?" and start asking "at what, exactly?", the fog lifts — and a clear, jagged shape appears.
Here is what I mean by Segmented AGI. In certain narrow domains, machines did not merely catch up to us — they left us behind, completely and uncontroversially. In 2016 a program beat the world's best Go player a decade before experts thought possible. In 2020, AI cracked a 50-year-old problem in biology by predicting the shapes of proteins; the work earned a Nobel Prize and a database now covering essentially every protein known to science. Newer models already surpass it.14 AI systems now forecast medium-range weather competitively with the great physics-based supercomputer models. In chess, poker, and competition-level math and coding, the best human on earth is no longer the best player on earth.
That is real superintelligence — segmented into slivers. Now look at the same systems on the open road of ordinary life: remembering yesterday, learning continuously, navigating a kitchen, knowing when they're wrong. There, they're shaky. Plot it all on one chart and you don't get a dot creeping toward a line. You get a skyline.
It isn't a dot approaching a finish line. It's a skyline — towers next to empty lots.
Segmented AGI explains the whole confusion. Camp A is staring at the towers and calling the city built. Camp C is standing in an empty lot insisting nothing's here. Both are describing the same skyline from different street corners. The interesting questions for the next few years aren't "are we there yet?" — they're how many more lanes go superhuman, how fast, and what happens to the work and the world as they do.
Beyond AGI sits the milestone that actually keeps researchers up at night — and it isn't just a smarter chatbot.
I'll borrow my own shorthand for it: HRSRAI — High-Reasoning, Self-Recursive Artificial Intelligence. The field's name for the engine underneath it is recursive self-improvement (RSI): an AI capable enough to meaningfully improve the next version of itself. The mathematician I. J. Good sketched the idea back in 1965 and called the result an "intelligence explosion" — each generation designing a smarter successor, faster than humans could follow.15 Run that loop long enough and you arrive at artificial superintelligence (ASI): a mind that exceeds the best of all of humanity, in every domain at once.16
For sixty years this lived in the realm of theory. What changed by 2026 is subtler and more real than the sci-fi version: AI is now materially accelerating its own development — under human supervision, in production, at scale. One leading lab reported its engineers writing many times more code with AI help than two years earlier (while cautioning that raw code volume overstates the true speed-up).17 Fully autonomous self-rewriting — an AI improving itself with no human in the loop — remains speculative.18 We are not in the intelligence explosion. We are, arguably, standing near the fuse.
The danger was never a smarter chatbot. It was a chatbot that builds the next chatbot.
This is where the story turns distinctly American. Washington's animating fear is simple: that China reaches transformative AI first.19 When the U.S. and Chinese presidents met in Beijing in 2026, an AI rivalry hummed beneath everything. But — and this is the nuance most headlines miss — the two countries may not even be running toward the same finish line.
Industry estimates put Chinese frontier models roughly six to eighteen months behind America's best — "nanoseconds," in one chip executive's telling.20 Analysts at Brookings caution that "winner-take-all" may be the wrong frame entirely: the two may simply run side by side for years.
Suppose the optimists are right and a true HRSRAI / ASI arrives. What would it mean for our species? The honest answer is that we don't know — but the shape of the bet is clear enough to write down.
It's worth saying plainly: ASI does not exist today, and a respectable share of researchers doubt the intelligence explosion will ever ignite at all.16 The Singularity — the point where machine progress slips past human comprehension entirely — may be a real horizon or a permanent mirage. I lean skeptical that we're close. But "we don't know" is not the same as "we're safe," and a country building the fuse should probably also be building the fire extinguisher.
The fight is about the ruler, not the machine. With no agreed definition of intelligence, "have we reached AGI?" can't have one answer. Experts looking at identical models land in genuinely different camps.
Today's AI is a prediction engine with astonishing breadth and brittle depth — superhuman fluency wrapped around a shaky grip on the real world.
Read the prophets' incentives. Many of the loudest forecasts come from people who profit whether they hype the upside or the doom. That doesn't make them wrong — it makes them marketing.
We're in Segmented AGI. Already superhuman in narrow lanes, sub-human across the open road. A skyline, not a dot. The real question is how fast new lanes top out.
The next horizon is self-improving AI — and the U.S.–China race toward it is on, even if the two nations are chasing different prizes. The promise is immense; so is the unsolved problem of keeping it aligned with us.
I started writing songs about this country because America is at its best when it's curious, generous, and a little brave all at once. The machines we're building will test all three. We don't get to choose whether the technology arrives. We do get to choose whether we meet it as informed citizens or as a credulous audience. Read the map twice. Then keep walking.
Where each term sits on the road from tools to superintelligence — and, where experts can be pinned down, where they think we stand today.
Every claim above is traceable. Tap through and judge for yourself — that's the whole point.
American Spotlight · "The Measured Machine" · by M. S. McKenzie · for the curious citizen