Matt Shumer's 'Something Big Is Happening' Went Viral. Here's What He Got Right, What He Got Wrong, and Why Millions Couldn't Stop Reading It.

On February 9, Matt Shumer published a 5,000-word essay called “Something Big Is Happening” on his personal blog. Within 24 hours, it had tens of millions of impressions and tens of thousands of likes on X. Matt Walsh shared it. The Washington Post’s Megan McArdle wrote about it. Fortune, Inc., and Mediaite covered it. Your uncle probably texted it to you.

The essay’s thesis: we are living through a COVID-level inflection point for artificial intelligence, most people don’t realize it yet, and the window to prepare is closing fast. Shumer argues that recent model releases — GPT-5.3 Codex and Claude Opus 4.6, both launched on February 5 — represent not incremental progress but a categorical leap. He cites METR research showing AI task completion doubling every seven months (possibly accelerating to four), quotes Dario Amodei predicting 50% of entry-level white-collar jobs could vanish within five years, and points to OpenAI’s own documentation stating that GPT-5.3 was “instrumental in creating itself.”

It’s a compelling read. It’s also a document that rewards careful scrutiny — because some of the claims hold up under pressure, some don’t, and the author himself carries baggage that’s worth understanding before you take the essay at face value.

Who Is Matt Shumer?

Shumer is the co-founder and CEO of OthersideAI, the company behind HyperWrite, an AI writing and autocomplete tool. He studied at Syracuse University’s Whitman School of Management and founded two companies before college — Visos (VR for medical applications) and FURI (sporting goods). He’s an active angel investor in AI infrastructure and created the open-source GPT-Prompt-Engineer tool. He has a genuine track record in the space and a sizable following.

He also has a credibility problem.

In September 2024, Shumer released Reflection 70B, an open-source model he claimed achieved unprecedented benchmark results. Independent evaluators couldn’t replicate the performance. Some found evidence the model was essentially routing queries to Anthropic’s Claude 3.5 Sonnet — a commercial API being passed off as an open-source breakthrough. Tom’s Guide documented the full controversy. Shumer later said he “got ahead of himself” without fully explaining the discrepancies.

This doesn’t automatically invalidate his essay. But it means his claims about the current state of AI should be evaluated on their sources, not on his authority — which is exactly what we’re going to do.

The Claims That Hold Up

The METR Data Is Real — and Genuinely Striking

The most substantive claim in Shumer’s essay is the one with the best empirical backing. METR (Model Evaluation & Threat Research) published research in March 2025 showing that the “time horizon” of AI — the task duration at which an agent succeeds 50% of the time — has been doubling approximately every seven months over six years. Their January 2026 update shows this trend continuing, with recent data suggesting possible acceleration to every four months.

The latest numbers are concrete. Claude Opus 4.5 hits 50% success on tasks that take human experts roughly 5.3 hours. GPT-5.2 manages 6.6 hours. These are tasks that involve multiple steps, tool use, and sustained reasoning — not parlor tricks.

The exponential trend line, when projected forward, implies AI agents capable of completing multi-day tasks autonomously within a year or two. That projection is what’s powering much of the urgency in Shumer’s essay and in broader industry discourse.

The Self-Building Model Claim Is Documented

Shumer cites OpenAI’s documentation stating GPT-5.3 Codex was “instrumental in creating itself.” This isn’t speculation — it comes directly from OpenAI’s announcement and system card. The Codex team used early versions to debug training runs, diagnose evaluation results, manage deployment infrastructure, optimize serving stacks, and dynamically scale GPU clusters during launch.

NBC News and The New Stack both covered this independently. The feedback loop is real. Whether it represents the beginning of recursive self-improvement or just a very good coding assistant being used by its own engineering team is a matter of interpretation — but the factual claim checks out.

The Amodei Quotes Are Sourced and Consistent

Dario Amodei’s prediction about 50% of entry-level white-collar jobs traces to an Axios interview from May 2025, where he warned that “AI could wipe out half of all entry-level white-collar jobs and spike unemployment to 10% to 20% in the next one to five years.” He repeated the thesis on 60 Minutes and CNN, and expanded on it in his January 27, 2026 essay “The Adolescence of Technology,” where he described each AI data center cluster as having the brainpower equivalent of “50 million Nobel Prize winners” and framed the technology as “the single most serious national security threat we’ve faced in a century.”

Shumer isn’t fabricating these quotes or pulling them from obscure sources. This is Anthropic’s CEO, on the record, repeatedly.

The Claims That Don’t Survive Scrutiny

The METR Graph Is Probably the Most Misunderstood Chart in Tech Right Now

MIT Technology Review published a piece on February 5 titled “This is the most misunderstood graph in AI,” and it directly addresses the exponential extrapolation that Shumer and others are making from the METR data.

The problems are specific and technical:

Time is not difficulty. A one-hour data entry task and a one-hour strategic planning task are completely different cognitive challenges. Models excel at pattern-matching within well-defined domains. Extending a trend line from “can do 6-hour coding tasks” to “can do 6-month business strategy” requires a leap that the data doesn’t support.

Data sparsity at the frontier. In the 1-4 hour range where the most dramatic recent results appear, METR’s original dataset had “remarkably few samples.” Trend lines drawn from a handful of successful long tasks create false confidence in the extrapolation.

Domain specificity. The strongest results concentrate overwhelmingly in software engineering — a domain with formal logic, verifiable outputs, and massive training data. Extrapolating to legal research, medical diagnosis, or financial modeling involves assumptions the data can’t validate.

50% success rate is not deployable. A system that fails half the time on four-hour tasks cannot be autonomously deployed in any professional context where errors carry consequences. The gap between “50% benchmark success” and “reliable enough to replace a human” is enormous.

The COVID Analogy Falls Apart on Inspection

M.G. Siegler’s response essay in Spyglass makes the sharpest critique of Shumer’s framing: “The virus didn’t need a Substack.” COVID-19 didn’t require persuasive blog posts to demonstrate its reality. It overwhelmed hospitals. It shut down borders. The evidence was visible, immediate, and impossible to deny.

If AI were genuinely at a COVID-equivalent inflection point, you wouldn’t need 5,000 words of persuasion to convince people. The displacement would already be obvious in unemployment numbers, in the collapse of specific service industries, in visible institutional failure. Instead, what you have is a set of impressive benchmarks, some dramatic quotes from industry leaders, and a projection curve that requires you to assume past exponential trends continue indefinitely — which, historically, they almost never do.

The Recommendations Contradict the Urgency

Here’s the tell that even Shumer may not fully believe his own framing: his practical advice. If 50% of entry-level white-collar jobs are genuinely at risk within five years, the appropriate response isn’t “spend an hour a day experimenting with AI tools.” It’s “fundamentally restructure your career, your savings, your education, and your risk exposure.” The modest, self-help-book-level suggestions — try paid AI models, build savings, pursue your passions — sit uncomfortably alongside claims of civilizational disruption.

The Conflict of Interest Nobody’s Ignoring (But Not Enough People Are Weighting)

Tim Rice of the Daily Wire put it bluntly: “Guy who sells AI tools warns us AI is coming for every job and tells us the only way out is to pay for AI tools like his.”

This isn’t an ad hominem dismissal — it’s a structural observation about incentives. Shumer runs HyperWrite. His investors benefit from AI adoption acceleration. His audience growth benefits from viral AI narratives. Dario Amodei runs a company valued at $60 billion whose valuation depends on the assumption that AI will reshape entire industries. OpenAI’s self-building model narrative supports their pitch to enterprise customers and government regulators alike.

None of this means they’re lying. It means every claim from every source in this ecosystem should be weighted against the financial incentive behind it. At Davos in January 2026, multiple CEOs from outside the AI industry publicly disagreed with Amodei’s timeline, arguing that actual enterprise AI deployment is slower, messier, and more limited than the models’ benchmark performance would suggest.

What’s Actually Happening

Strip away the hype and the counter-hype, and here’s what’s concretely true as of February 2026:

The models are genuinely better. Claude Opus 4.6 and GPT-5.3 Codex represent meaningful capability jumps, particularly in sustained agentic work, code generation, and long-context reasoning. This isn’t marketing — the benchmark improvements are independently verified and the practical differences are visible to anyone using these tools daily.

The feedback loop is tightening. OpenAI using its own model to debug its training pipeline is a notable milestone, regardless of whether you call it “self-improvement” or “a good coding assistant.” Anthropic has acknowledged similar dynamics internally. The distance between “AI helps build AI” and “AI builds AI” is still significant, but it’s measurably shorter than it was a year ago.

Job displacement is happening, but unevenly and slowly. Entry-level coding tasks, first-draft content generation, routine legal research, and basic data analysis are all areas where AI is already reducing headcount in specific organizations. But “reducing headcount at some companies” and “eliminating 50% of entry-level white-collar jobs” are separated by an enormous gap filled with institutional inertia, regulatory friction, integration costs, and the stubborn complexity of real-world workflows.

The pace of progress is legitimately unprecedented. Shumer’s timeline from 2022 to 2026 — arithmetic failures to passing the bar to writing production software to self-referential model development — is factually accurate. Whether this pace continues, accelerates, or hits diminishing returns is the trillion-dollar question nobody can answer honestly.

The Real Lesson of 33 Million Views

The most interesting thing about “Something Big Is Happening” isn’t whether Matt Shumer is right about the timeline. It’s that millions of people read a 5,000-word essay about exponential AI capability curves and job displacement projections in a single day. That doesn’t happen because someone wrote a good blog post. It happens because a critical mass of people — professionals, knowledge workers, people who write code and contracts and reports for a living — already feel the ground shifting under them and are looking for someone to articulate what they’re sensing.

Shumer’s essay resonated not because it revealed new information, but because it organized existing anxieties into a coherent narrative with specific data points. Whether those data points fully support the conclusion is almost secondary to the cultural signal: tens of millions of people are worried enough about AI displacement to spend 20 minutes reading about METR benchmarks on a Sunday afternoon.

That anxiety is worth taking seriously, even if the specific predictions turn out to be wrong on timeline. The models are getting better fast. The feedback loops are tightening. The companies building these systems are, for the first time, openly warning about their own products’ displacement potential. Whether the inflection point is 2026 or 2030 matters less than whether you’re paying attention at all.

Sources: