Is Something Big Happening? A Viral Essay by Matt Shumer
The question isn’t how big AI can get — it’s whether it’s progressing at all at its core. Current “self-improving” systems are narrowly limited to fixing code not changing foundational principles
Here are Matt Shumer’s viral essay “Something Big Is Happening” core ideas:
A Step-Change in AI Capability Has Already Occurred
February 5, 2026 marked a discontinuity: new models (GPT-5.3 Codex, Opus 4.6) made everything before them feel like “a different era”
AI has progressed from failing basic arithmetic in 2022 → passing bar exams in 2023 → writing working software in 2024 → engineers handing over most coding work by late 2025 → AI now building and testing complete applications autonomously in 2026
AI Agents Now Possess “Taste” and Autonomous Judgment
The latest models don’t just execute instructions — they make intelligent decisions independently
Shumer describes AI that designs apps, writes tens of thousands of lines of code, opens and tests the app itself, iterates on design flaws autonomously, and only returns when it judges the work meets its own standards — requiring zero human corrections
Self-Improving AI Has Begun
GPT-5.3 Codex was “instrumental in creating itself” — used to debug its own training, manage deployment, and diagnose evaluations.
This marks the potential start of an intelligence explosion where each generation accelerates the next
Massive, Imminent Labor Disruption Across White-Collar Work
Dario Amodei (safety-focused AI CEO) predicts 50% of entry-level white-collar jobs eliminated within 1–5 years — many insiders consider this conservative.
Affected fields include:
- Software engineering: Complex multi-day projects fully automated
- Legal: Contract review, case law research, brief drafting rivaling junior associates
- Finance: Model building, data analysis, investment memos
- Writing/Content: Output indistinguishable from humans
- Medicine: Scan interpretation, diagnosis suggestions
- Customer service: Genuine multi-step problem-solving agents (not chatbots)
Unlike Past Automation, There’s No “Safe Harbor”
AI substitutes for general cognition, not specific tasks — meaning displaced workers can’t easily transition to adjacent roles as in previous automation waves
Urgent Action Required — Not Panic
Shumer emphasizes this isn’t meant to scare but to prepare:
- Use AI seriously now: Subscribe to premium models ($20/mo), but crucially — integrate it into actual work, not just quick queries
- Build adaptability as your core skill: Tools will obsolete rapidly; the ability to continuously learn new workflows is the only durable advantage
- Commit to daily experimentation: One hour daily using AI for novel, challenging tasks will put you ahead of 99% of peers within six months
- Seize the brief window: While most companies still ignore this shift, early adopters who demonstrate AI-augmented productivity will be “the most valuable person in the room — right now”
Existential Stakes
- Potential upside: Solving grand challenges (disease, climate)
- Potential downside: Uncontrolled autonomous systems with capabilities exceeding human oversight
- Framed as humanity’s most critical inflection point — comparable to “February 2020” (early pandemic) where early recognition determined outcomes
This isn’t a prediction about the distant future — it’s a description of capabilities already deployed that most people haven’t yet encountered because they’re not using the latest models or integrating AI deeply into workflows.
Critical/Skeptical Responses
Gary Marcus (AI researcher, NYU)
- Called the essay “weaponized hype” that “tells people what they want to hear, but stumbles on the facts”
- Criticized Shumer for providing no actual data supporting claims that AI can write complex apps without errors
- Noted Shumer selectively cited METR benchmarks while omitting that the success criterion is only 50% correct (not 100%) and ignoring widespread hallucination/error problems
- Highlighted security concerns with AI-generated code and pointed out Shumer’s history of “exaggerated claims” about unreplicated models
Vox Technology
- Published a critique titled “What the latest viral AI apocalypse warning gets wrong” arguing Shumer’s pandemic analogy was misleading
- Noted economic constraints (productivity J-curve, Baumol effect) would slow real-world AI adoption despite capability gains
- Cited Google DeepMind CEO Demis Hassabis’s caution that “one or two AlphaGo-level breakthroughs” are still needed for AGI as a reality check
Forbes (Paulo Carvão)
- Critiqued Shumer’s “inevitability” framing that creates unnecessary panic
- Argued responsible AI integration requires enterprise software adoption cycles, not individual urgency
- Questioned whether Shumer’s narrative served more as positioning/marketing than sober analysis
Key Pattern in Responses
Most critics didn’t dispute that AI capabilities are advancing rapidly, but challenged:
- Reliability claims — current systems still produce frequent errors/hallucinations despite improved fluency
- Economic adoption speed — enterprise integration faces organizational friction beyond pure capability
- Security/trustworthiness — AI-generated code and outputs lack verification mechanisms for critical applications
- Rhetorical framing — pandemic analogies and urgency may be disproportionate to near-term reality
