Essay · February 2026

The Day It Becomes Obvious

How AI went from party trick to the most consequential technology in history — and why most people haven't noticed yet.

~18 min read

Something extraordinary is happening with artificial intelligence, and most people don't know it yet.

Not "extraordinary" in the way tech companies always say their products are extraordinary. Not another crypto, not another metaverse, not another overhyped Silicon Valley fever dream. Something more like the early days of COVID — where a small group of people who understood exponential growth were watching the numbers with mounting alarm while the rest of the world carried on as normal.

This essay makes a specific, testable prediction: sometime between late 2026 and mid-2027, AI will have its "March 11, 2020" moment — a week or two when it becomes suddenly, viscerally undeniable to the general public that this technology is reshaping the economy and the world, whether they like it or not.

On March 11, 2020, within the span of about six hours, the WHO declared COVID a pandemic, Tom Hanks announced he'd tested positive, the NBA suspended its season mid-game, and Trump announced a European travel ban. The day before, most Americans thought of COVID as something happening in other countries. The day after, everything was different. Not because the virus had changed — but because the accumulated weight of evidence finally broke through collective denial all at once.

AI is approaching an equivalent moment. Here's why.

· · ·

What Actually Happened While You Weren't Looking

If you last paid serious attention to AI when ChatGPT went viral in late 2022 and concluded it was a clever but unreliable chatbot — a party trick that hallucinates and can't do math — your mental model is roughly three years and several revolutions out of date.

The benchmarks tell a clear story

The standard academic tests that were supposed to measure AI capability have been almost entirely conquered. Graduate-level science questions: AI scores 92.6%, above the 69.7% scored by human PhD-holding domain experts. The hardest competition mathematics: 100%. Medical licensing exams, bar exams, PhD-level reasoning — all essentially solved.

But those benchmarks were created for a world where AI was much weaker. More telling are the tests specifically designed to resist AI progress — created in the last year or two with the explicit goal of being "AI-proof":

FrontierMath, a set of research-level mathematics problems so hard that most professional mathematicians can't solve them, went from under 2% to roughly 40% correct in about thirteen months. SWE-bench Verified, a test of real-world software engineering tasks pulled from actual open-source projects, went from about 33% to roughly 80% in eighteen months. Humanity's Last Exam, created by thousands of experts contributing their hardest questions, has already been pushed to around 45% — within its first year of existence.

These aren't incremental improvements. These are massive leaps on tests specifically designed to be hard enough that this wasn't supposed to happen yet.

The metric that matters most

A nonprofit research organization called METR developed what may be the single most useful measurement of AI progress: the autonomous task time horizon. The concept is simple: take a set of real software engineering and reasoning tasks, measure how long each one takes a skilled human professional, then see which tasks an AI agent can complete on its own without human help.

14.5 hrs
The length of human-expert tasks that the leading AI model can now complete autonomously — with 50% success rate

As of February 2026, the most capable AI model (Claude Opus 4.6, made by Anthropic) can independently complete tasks that take skilled humans about 14.5 hours, with a 50% success rate. For tasks it can complete 80% of the time — the threshold of "reliable enough to actually count on" — the time horizon is about one hour.

That one-hour number might not sound dramatic. But here's the thing that changes everything: this metric has been doubling every seven months, consistently, for six years. Not "improving gradually." Doubling. Exponentially. Every measurement, across every model generation. No plateau. No slowdown.

See the METR Time Horizon Graph
The single most important chart for understanding AI progress. Toggle between log and linear scale to see both the consistent exponential trend and the dramatic hockey-stick shape.
View interactive chart at metr.org →

To make this concrete, here's what the trajectory looks like for the 80% reliability threshold — the level at which companies can genuinely restructure around AI, using the conservative seven-month doubling rate:

Projected 80% Reliability Time Horizon (Conservative 7-Month Doubling)
DateReliable Time HorizonWhat That Means
Feb 2026~1 hourShort, well-defined tasks
Sep 2026~2 hoursMeaningful subtasks
Apr 2027~4 hoursHalf a workday
Nov 2027~8 hoursA full workday of autonomous work
Jun 2028~16 hoursMulti-day projects
Jan 2029~32 hoursA full work week

An important caveat: the time horizon doesn't mean "the AI works nonstop for eight hours like a human." It means the AI can be given a task that would take a skilled professional a full day, and it will return correct, usable results roughly 80% of the time. One human overseeing several AI agents, each handling day-long tasks, is a massive productivity multiplier — and it's a structure companies will find irresistible.

The most recent data actually suggests the doubling time may be accelerating to around four months. On that faster schedule, reliable full-workday autonomy arrives by early 2027 instead of late 2027.

AI coding is the canary in the coal mine

If you want to see the future arriving in real time, look at software development. It's being transformed right now, in public, with receipts.

41%
of new code industry-wide
is now AI-generated
95%
of code written by the team that builds Claude Code is written using Claude Code

An AI tool called Claude Code — essentially an AI software engineer you interact with through a command line — has become arguably the most transformative professional tool released in years. At Coinbase, over 40% of daily code is now AI-generated. One developer publicly documented landing 259 pull requests — over 40,000 lines of new code — in a single month, 100% written by AI. A team used AI to build a working C compiler capable of compiling the Linux kernel, for about $20,000 in API costs.

A researcher at OpenAI put it bluntly: "Programming always sucked. I'm glad it's over. 100% of my code is being written by AI."

This isn't happening in some research lab. This is happening at publicly traded companies, shipping real products, right now.

The money is not theoretical

This is not a speculative bubble built on promises. The financial numbers are real:

$14B
Anthropic annual revenue
(10× growth year-over-year)
$20B
OpenAI annual revenue
(tripled in one year)

Enterprise AI spending hit $37 billion in 2025 — a 22× increase from $1.7 billion just two years earlier, the fastest enterprise software category expansion in history.

$660B+
Combined AI infrastructure spending planned by the five biggest tech companies in 2026 alone. That's roughly the GDP of Switzerland.

This level of spending creates its own gravity. Companies that have committed hundreds of billions to AI infrastructure need to show returns on that investment. That means deploying AI to replace labor costs — not as a future possibility, but as an urgent business imperative.

· · ·

Why Most People Still Don't Get It

Given all of this, why is the dominant public sentiment still roughly "AI is overhyped"?

The product most people use is not the frontier. The average person's experience of AI is the free version of ChatGPT, which is several generations behind the cutting edge. It's like judging the future of aviation by looking at a paper airplane. The tools actually reshaping industries — Claude Code, GPT-5.2 Pro, enterprise agent deployments — are used by a relatively small number of professionals and developers.

AI still makes dumb mistakes. Frontier AI models can build working compilers and find security vulnerabilities in code, but they can also confidently misread a chart or hallucinate a citation. This feels contradictory, but it's exactly what you'd expect from a technology that's superhuman in some dimensions while still catching up in others. The dumb mistakes are more memorable and more shareable, which means public perception systematically underweights what AI can do.

People are pattern-matching to previous hype cycles. Crypto, the metaverse, Web3, self-driving cars — the last decade has been littered with technologies that were supposed to change everything and didn't. It's entirely rational to be skeptical when Silicon Valley says "this time it's different." But the exponential curves in AI capability are unlike anything in any of those previous cycles.

Workers are hiding their AI use. Anthropic's own research found that 69% of workers who use AI at work conceal it from their employers. When AI makes you more productive, the rational individual response is often to pocket the efficiency gain as free time rather than advertise it. This means the economic impact is larger than it appears from the outside.

The gap between what AI can actually do and what the general public believes it can do is wide and growing wider — because the technology is improving faster than public perception is updating.

This is precisely the condition that produces sudden phase changes in collective awareness: when accumulated reality finally breaks through accumulated complacency, all at once.

· · ·

What the Experts Are Saying

This isn't just the opinion of AI enthusiasts. Here is where the people who lead the major AI labs and research organizations have placed their estimates:

2027
Dario Amodei, CEO of Anthropic — expects AI with intellectual capabilities matching Nobel laureates across most disciplines by late 2026 or early 2027.
2028
Shane Legg, co-founder of Google DeepMind — has maintained a 50% probability estimate for artificial general intelligence by 2028 for over fifteen years.
2029
Daniel Kokotajlo, former OpenAI researcher and prominent forecaster — places his median for full coding automation at 2029. His 2021 predictions about 2025 proved remarkably accurate.
Oct 2027
Metaculus, a large forecasting platform with a strong track record — community median prediction for "weakly general AI."

Note the spread: roughly 2027 to 2029, from people who are building the technology and have strong incentives to get their forecasts right. The public "wake-up moment" doesn't require full artificial general intelligence — it only requires enough visible economic disruption that denial becomes untenable. That threshold is lower, and arrives earlier.

· · ·

The Prediction

Will there be a specific moment — a week, a month — when AI stops being something people have opinions about and becomes something that has undeniably changed their lives?

The most likely trigger (~75% probability)

The conditions for a sudden public wake-up are assembling. The 80% reliability threshold goes from "one hour" today to "a full workday" within roughly 12–18 months. Hundreds of billions in infrastructure spending demand returns. Companies that invested heavily in AI through 2024–2025 will be reaching full deployment through 2027.

The trigger will almost certainly be negative, not positive. Positive AI milestones don't create mass awareness because they're too abstract. What creates mass awareness is personal relevance — your industry is affected, your company is restructuring, your kid can't find an entry-level job.

The most likely scenario: a marquee employer outside of tech — a major bank, consulting firm, or professional services company — explicitly cites AI while cutting a large, visible slice of white-collar headcount. A second follows within days. Meanwhile, a consumer-facing AI capability is circulating that gives ordinary people a visceral understanding of why it's happening. This probably plays out over a quarter rather than a single day, but the narrative crystallization — the moment when "AI is changing everything" becomes the default assumption rather than a controversial claim — will feel sudden.

Central Estimate
Spring to Summer 2027
With meaningful probability extending from late 2026 through end of 2027. Roughly 25% chance it arrives earlier, 25% chance the transition is gradual with no single dramatic moment.

The alternative: no single moment (~25% probability)

There's a genuine possibility that AI disruption unfolds like climate change — gradually, unevenly, deniably. Each quarter, a few more companies announce efficiency gains. Unemployment ticks up in certain sectors. Entry-level knowledge work dries up slowly. But it never crystallizes into a single dramatic week. Unlike COVID, AI disruption doesn't force binary institutional decisions on hard deadlines.

Even in this scenario, the underlying transformation is the same. The question is only whether the public realization arrives in a concentrated burst or a slow burn.

· · ·

Addressing the Skeptic

"AI has been 'about to change everything' for years and it hasn't."
The difference is in the data. Crypto never had an exponential capability curve validated across six years of independent measurements. The metaverse never generated $34 billion in combined annual revenue from two companies. Self-driving cars never saw 41% of their industry's output being AI-generated. The closest historical analogy isn't previous tech hype — it's the early internet, where skepticism was widespread and reasonable right up until it wasn't.
"AI still makes mistakes, so it can't really replace people."
Humans make mistakes too — that's why we have editors, code review, quality assurance, and supervisory structures. The economic question isn't whether AI is perfect; it's whether one human supervising five AI agents produces more output than five humans working alone. At current capability levels, in domains like software engineering, the answer is increasingly yes. And the capability is doubling every few months.
"The people saying this are the ones selling it."
Fair point — lab CEOs have incentives to hype their products. But the METR time horizon data is produced by an independent nonprofit that has incentives to be conservative (overstating capabilities would undermine their credibility on AI safety). The benchmark results are independently verifiable. The revenue figures are public. And the skeptics have their own incentive problems — many AI critics have built careers on the "it's all hype" position and face reputational costs from updating.
"The productivity gains haven't shown up in the economic data."
True as of early 2026, and a legitimate puzzle. Partial explanations: workers are hiding their AI use. Enterprise deployment cycles run 12–18 months, so the wave of adoption starting in 2024 hasn't fully translated into restructuring. And we're in the brief window where AI boosts individual productivity but organizational structures haven't caught up — the same lag that occurred with electrification of factories, which took decades to reach full potential even after the technology was proven. The lag will be much shorter this time because AI is software, not physical infrastructure.
· · ·

Why This Matters Even If You Don't Work in Tech

The sectors most immediately affected are knowledge work broadly: software engineering, data analysis, legal research, financial analysis, content creation, customer service, administrative work, and consulting. If you or someone you know works in any of these fields, the impact is not theoretical and not distant.

But the ripple effects extend further. If AI can reliably handle day-long autonomous tasks by 2027 or 2028:

Entry-level knowledge work becomes scarce. The traditional path of "start in a junior role, learn on the job, get promoted" gets disrupted when AI handles junior-level tasks more cheaply and often more reliably. This hits new graduates first and hardest.

The labor market restructures around AI supervision. The most valuable skill becomes "directing and quality-checking AI output" rather than "doing the work yourself." This rewards different abilities than the current system selects for.

Service costs drop while displacement rises. Tasks that currently require expensive professionals — legal review, financial planning, diagnostic support — become dramatically cheaper. Genuinely good for consumers. Disruptive for the professionals in those fields.

The pace of change itself accelerates. When AI can do more of the research and development work, the feedback loop between "better AI" and "faster AI research" tightens. This is already happening — AI models are being used extensively in their own development process.

· · ·

The Bottom Line

We are living through the early stages of the most consequential technological transformation in at least a generation. The evidence for this is not speculative — it's in the benchmark data, the revenue figures, the infrastructure spending, the expert predictions, and the lived experience of millions of professionals already using these tools daily.

The gap between what AI can actually do and what the general public believes it can do is wide and growing wider. This is exactly the condition that produces sudden phase changes in collective awareness.

My best estimate: the general public's "oh, this is real" moment arrives between spring and summer of 2027, most likely triggered by a visible cluster of AI-driven workforce changes at major employers, combined with consumer-facing capabilities that make the abstract feel concrete.

None of this is certain. Exponential trends can break. Deployment challenges, regulation, or technical barriers could slow things down. There's perhaps a 10–15% chance that current trends decelerate enough to push any reckoning further out.

But the direction is clear. And the time to start paying attention was yesterday.