On September 14, 2026, Elon Musk circulated Tim Urban's 2015 Wait But Why cartoon again — the one where machine intelligence crawls along the bottom of a chart, then rockets past "Human" and "Einstein" while stick figures flip from "Haha adorable monkey tricks" to "The fuck??" Musk said the drawing hits the target, argued that AI is improving rapidly, and implied we are already riding the steep part of that curve. The moment is less a new technical announcement than a cultural reset: a decade-old explainer is back at the center of a week already crowded with AGI declarations, superintelligence debates, and Musk's own endorsement of industry pacing.
If you build with frontier models daily, the useful question is not whether Musk liked a meme. It is whether exponential takeoff framing should change how you ship agents, evals, and governance this quarter — or whether it is mainly alignment politics dressed as a graph.
TL;DR — what people are asking
| Question | Direct answer |
|---|---|
| What did Musk share? | A repost of Tim Urban's "Reality" intelligence curve from Wait But Why's 2015 AI Revolution series — flat progress, then a near-vertical line past human-level labels, with stick-figure reactions. |
| Did Musk say anything new? | He reinforced a long-standing view: improvement is fast, the risk window is narrow, and society is psychologically unprepared — not a new xAI benchmark or product drop. |
| Is the chart evidence? | No. It is a thought experiment about how humans misread exponentials. It is not measured IQ, not a scaling-law fit, and not a dated prediction that 2026 must be the knee of the curve. |
| Why viral again now? | September 2026 already has frontier-model launch fatigue, CEO AGI headlines, and safety coordination news — Urban's joke names the feeling that capability jumps faster than institutions adapt. |
| What should builders do? | Plan for jagged superhuman pockets (coding, math, research assistance) without assuming domain-general superintelligence; keep evals and guardrails on agentic workflows; ignore label wars unless they change your compliance surface. |
| Where is the original art? | On Wait But Why — Part 1 of "The AI Revolution: The Road to Superintelligence" (January 2015). Do not re-host the copyrighted cartoon; link out. |
What the Wait But Why curve actually argues
Tim Urban's 2015 series did something textbooks rarely manage: it made recursive self-improvement and fast takeoff legible to readers who had never opened a reinforcement-learning paper. The famous panel Musk reposted is not a forecast with error bars. It is a pedagogical insult to human intuition.
Urban's setup assumes two ideas that were already mainstream in AI safety circles but alien to general readers:
- Intelligence can be treated as a single axis (a simplification experts still argue about).
- Once a system can improve its own cognition, progress might look flat for a long time and then vertical — because each increment of intelligence helps produce the next increment faster.
The stick figures are the point. Humans are good at extrapolating linear change (slightly better phones each year) and bad at exponential change (the same "small" weekly gain compounding into something categorically different). The first caption — treating early AI as cute parlour tricks — is how Urban dramatizes under-reaction. The second caption is over-correction panic once the curve is obviously leaving human band.
Writers such as Vikram Chandra and many others have since said that essay changed how they thought about AI — not because it proved a date, but because it replaced vague "smarter computers someday" with a picture of discontinuity. Musk reposting it in 2026 is an attempt to put that picture back in front of a public that has spent months arguing whether GPT-6 Astra is "AGI," "math superintelligence," or "expensive autocomplete with great launch demos."
Musk's repost in context — same curve, new news cycle
Musk has used exponential intelligence rhetoric for years — from OpenAI co-founding-era warnings through xAI's Grok launches and interviews where he attaches dates to machine superiority. explainx.ai covered his 2031 "all human intelligence combined" timeline as a specific, hold-loosely forecast tied to media appearances, distinct from any single cartoon.
September 2026 adds fresh anchors that make Urban's 2015 panel feel prophetic even when the underlying metrics are messy:
- Launch-week spectacle for frontier models, followed by practitioner pushback on reliability in daily coding and agent workflows — the same "day one magic / day two friction" cycle Jensen Huang's AGI post sat inside.
- Label inflation: "AGI," "superintelligence," and "alien mind" metaphors competing in public threads while labs' own safety essays call for pacing the frontier rather than declaring victory.
- Musk's own policy pivot: endorsing Anthropic's pacing framing days before re-sharing Urban's panic panel — tension between slow down verification rhetoric and speed up capability marketing is not new, but it is visible in a single week's timeline.
None of that proves the Wait But Why curve's vertical segment started in 2026. It proves the story still mobilizes attention: a simple drawing beats a fifty-page AGI-to-ASI pathway paper in share count, even when the paper is what safety researchers actually cite.
2015 vs 2026 — what aged well and what did not
What aged well
- Psychology of underestimation. Teams still ship agent features as "assistants," then discover emergent failure modes when tools chain — Urban's stick figures are a stand-in for production incidents, not lab leaderboard moves.
- Attention to takeoff narratives. DeepMind's 2026 "From AGI to ASI" work explicitly lists scaling, paradigm shifts, recursive improvement, and multi-agent collectives as pathways — all descendants of conversations Urban popularized, with more citations and fewer jokes.
- Crossover into culture. When novelists and journalists reference Wait But Why, they are signaling that AI risk is no longer niche — the same dynamic Musk exploits when he reposts the panel to hundreds of millions of followers.
What did not survive contact with reality unchanged
- Single-axis intelligence. Real systems are jagged: superhuman on some benchmarks, brittle on others. explainx.ai's superintelligence debate separates domain mastery from Bostrom-style broad superiority — a nuance Urban simplified for clarity.
- Imminent datelines. 2015 readers often walked away with a vague sense that transformation was around the corner. A decade later, transformers rewrote the economy, but human-level general agents remain uneven — fast progress plus long tails of failure, not a clean vertical line on one chart.
- Monolithic "AI" actor. 2026's frontier is many models, many labs, tool ecosystems, and agent harnesses — not one self-improving ghost in a box. Recursive improvement shows up in coding agents and eval loops, not only in a mythical singleton.
That gap matters for builders. If you plan as if one vertical line arrives next quarter, you over-rotate on moonshots. If you plan as if nothing ever changes, you under-invest in evals and access control while competitors ship agent loops that compound small weekly gains.
"Steep curve" — capability, hype, or both?
Musk's claim that we are on the steep part can mean three different things. They are not interchangeable.
| Meaning | What would count as evidence | What we actually have in September 2026 |
|---|---|---|
| Compute + algorithm scaling | Published scaling laws, training FLOPs, sustained benchmark lifts across diverse tasks | Strong but uneven — math and coding spikes with persistent reliability gaps elsewhere |
| Economic adoption | Revenue mix, task automation measured in hours saved, enterprise deployment breadth | Large and real, but distributed — not a single discontinuity event |
| Existential risk slope | Demonstrated autonomous self-improvement closed loops at frontier capability | Contested — RSI is discussed seriously (recursive self-improvement explainer) but not shown as Urban's cartoon implies |
Practitioners on the ground often report something closer to compound incrementalism: each model generation makes the previous generation's agent harness more valuable — more context, better tool use, cheaper inference — which feels like a curve bending upward even when no single number proves takeoff.
That is compatible with Musk's repost and with skeptic replies that current systems "still fail simple tasks." Both can be true: median capability rises while tail risk and tail magic both expand — jagged intelligence again.
What people are asking — alignment, politics, and builders
Is Musk using the cartoon to push regulation or speed?
Musk's messages oscillate — warn about extinction risk, then ship faster models; endorse pacing essays, then celebrate singularity memes. The Wait But Why repost is risk-framing, not a bill text. For operational policy, the September 2026 news with teeth is embedded evaluators and industry coordination, documented in Musk, Altman, and Hassabis reactions — not a stick figure panel.
Does the repost help non-technical readers?
Yes, with a caveat. Urban's original series is still one of the best on-ramps to why alignment researchers worry about fast transitions. The caveat: readers who stop at the cartoon may think AGI already happened because launch demos looked impressive — the same misread Huang's "AGI has arrived" post exploited. Pair memes with how to read benchmarks before making architecture decisions.
Should startups change roadmaps because Musk reposted a 2015 chart?
Only if your roadmap already ignored agent security, eval debt, and model swap risk. The cartoon is a reminder that discontinuities happen in perception before they happen in your Gantt chart — not a signal to freeze shipping. If you sell B2B automation, the actionable takeaway is contractual human review, logging, and capability-specific SLAs, not metaphysical panic.
How does this connect to superintelligence law?
Congressional proposals such as the Ban Artificial Superintelligence Act try to legislate thresholds Urban drew as a line on a cartoon. The enforcement problem is the same: definitions lag jagged capability. Musk sharing Urban does not clarify statutory language; it raises salience — which can accelerate both safety work and performative politics.
A builder checklist — exponential story, linear engineering
Use Urban's curve as scenario planning, not as a release schedule:
- Task-level evals, not headline labels. Before you trust a model on a customer workflow, run your own suite — the same discipline explainx.ai recommends when launch benchmarks change post-release.
- Assume harnesses compound. The steepest felt slope in 2026 is often agents + tools + memory, not raw next-token loss. Invest in loop engineering and stop conditions.
- Separate demo from deploy. Launch-week videos are stick-figure bait — adorable monkey tricks. Production is the second panel — unexpected failures at scale.
- Track policy that changes access. Pacing commitments may eventually affect release timing across labs; monitor corporate statements, not only cartoons.
- Read the primary explainer. Send teammates to Urban's official Wait But Why series for depth; use explainx.ai for 2026-specific fact checks and comparisons.
Why Tim Urban's essay still beats most 2026 threads
Quality explainer writing ages because it teaches mental models, not because it nails dates. Urban's series walked readers through narrow AI, general AI, why speed matters, and why alignment is hard — with citations to Bostrom, Yudkowsky-era debates, and mainstream scientists, wrapped in jokes that travel.
Musk did not need to invent a new metaphor in September 2026. He needed to re-anchor a distracted audience on discontinuity the week after pacing politics and launch hype. Vikram Chandra and others were right that the 2015 post changed conversation; Musk's repost is proof the conversation still has not converged on definitions, only on feelings about the slope.
For explainx.ai's read, the honest synthesis is this: capabilities in 2026 are seriously impressive and seriously uneven; institutions are slower than labs; exponential framing is useful for risk communication and bad for sprint planning unless you translate it into evals, access control, and concrete task metrics.
Related on explainx.ai
- DeepMind's four pathways from AGI to ASI — research-grade takeoff paths beyond a single cartoon axis
- Has AI reached superintelligence? The Astra debate, defined — why jagged capability breaks one-line "AGI" claims
- Jensen Huang says "AGI has arrived" with GPT-6 Astra — another September 2026 label fight, fact-checked
- Musk, Altman, and Hassabis react to "Pace the Frontier" — Musk's policy endorsement days before the Wait But Why repost
- Musk predicts AI beats all human intelligence combined by 2031 — dated superintelligence rhetoric from the same speaker
- What is the AI singularity? Musk's August welcome explained — soft vs hard singularity definitions
- Paul Graham's 1980 AGI test, revisited — a concrete threshold thought experiment, older than Urban's curve
- How to read AI benchmarks — turn hype curves into decisions you can defend in production
Sources: Tim Urban, "The AI Revolution: The Road to Superintelligence, Part 1", Wait But Why, January 2015 · Public Musk repost and commentary circulating September 14, 2026 · explainx.ai coverage linked above for GPT-6 Astra, pacing reactions, and superintelligence definitions
Cartoon descriptions refer to Tim Urban's published Wait But Why artwork (copyright Wait But Why / Tim Urban). explainx.ai links to the official essay and does not host the image. Capability claims and executive statements are accurate as of September 14, 2026, and may change as models, benchmarks, and corporate commitments evolve.