The Slop Curve: What Happens to AI Slop as Intelligence Compounds
Everyone has a name for it now. The confident paragraph that says nothing. The pull request that touches forty files to fix one. The generated landing page that looks like every other generated landing page. The internet settled on one word for all of it: slop.
Slop is not the same thing as being wrong. A wrong answer is cheap to catch. Slop is subtler. It's plausible excess: the model doing more than you asked, restructuring things you never wanted restructured, filling silence with filler, breaking two things on the way to fixing one. It's the difference between a colleague who is wrong and a colleague who is enthusiastic.
Slop is what capability looks like when it outruns judgment.
The complaint I keep hearing, from engineers, from founders, from people running large companies, is some version of the same question. Models keep getting smarter, so why does everything feel messier?
Both things are true at once. They sit on the same curve.
So I drew it.
The Graph
Put intelligence on the x-axis and slop on the y-axis. Only the positive side exists. There is no negative intelligence to plot, and the story only runs forward.
Roughly, it's a parabola with a long top. Slop rises with capability, flattens into a long stay at the summit, and only afterwards falls away toward zero while intelligence keeps going. The equation under the graph is the closest simple fit to that shape, though the real thing holds its peak longer than a clean parabola would. The flat stretch is the part that matters, and it's the part most projections leave out.
Three phases. Each one has its own logic.
Phase One: The Rise
Why does slop increase with intelligence at all? Smarter systems should produce less junk.
The rise happens because capability multiplies output before it multiplies judgment. Every increment of intelligence lowers the cost of producing something that looks finished. Code, prose, designs, decisions. The volume of plausible artifacts explodes, and the fraction of them that were actually needed does not keep pace.
A model that writes one function badly is annoying. A model that writes forty files convincingly is a different kind of problem, because the review burden grows faster than the value delivered. You didn't receive a pull request. You received homework. The system is smart enough to act, not yet smart enough to know when not to act.
That gap, between the power to generate and the taste to withhold, is the slope of the curve.
Sundar Pichai has the most useful word for its shape. He calls today's intelligence jagged: brilliant in one direction, tripping over something trivial in the next. It will architect a distributed system for you and then lose an argument about how many r's are in a word. Jaggedness is the engine room of this curve. A jagged system's peaks arrive years before its floors do, and slop is what pours out of the difference. Work produced at peak capability, checked at floor judgment.
This is where we live right now. It's also why "AI slop" became a word at exactly the moment models became genuinely useful. The curve predicts that. The complaint and the capability arrive together, because one is the exhaust of the other.
Andrej Karpathy is reporting from inside it. He now lets agents write most of his code, and in the same breath he warns that a slop flood is coming for every feed, every repo, every paper archive, and that the models, for all their brilliance, are still not really there. More capability and more mess, from the same person on the same day. That isn't a contradiction. It's the slope.
The Peak: Something AGI-Shaped
The curve peaks when something AGI-shaped shows up.
I'm choosing that phrase carefully. I don't think the peak gets marked by a press release. The systems that define the top of this curve are more likely to exist on stage, inside labs, inside evals, inside deployments that are real but never announced as a threshold. By the time anyone agrees to use the word publicly, the thing itself will have been around for a while.
Sam Altman has more or less said so from inside the company most responsible for the flood. He argues the takeoff toward superintelligence has already quietly started, and that it feels far less strange from within than anyone expected. He has also noticed out loud that the AI-heavy corners of the internet now feel fake to him. Hold those two thoughts together and you get the approach to the peak exactly as this graph draws it. Takeoff underway, feeds full of junk, no announcement anywhere.
Here's the uncomfortable part. The peak is the point of maximum capability with minimum control.
Dario Amodei has the sharpest description of what sits up there: a country of geniuses in a datacenter, possibly only a few years out. It's a good phrase. It also describes a country with no weekends, no performance reviews, and nobody senior enough to say no.
A system at that level can pursue goals across long horizons, touch real infrastructure, and generate consequences faster than any human review loop absorbs them. The guardrails we have (evals, red teaming, policy) were built for the climb, not for the summit. At the summit we can't fully characterize what the system will do, which means we can't bound what it might break.
Maximum capability, maximum unnecessary action, minimum oversight. That is peak slop, at civilizational scale instead of pull request scale.
Arthur C. Clarke's line fits almost too well: "Any sufficiently advanced technology is indistinguishable from magic." The problem with magic is that you can't code review it.
None of which makes the peak a death sentence. It makes it a control problem, which is a very different kind of thing, and the rest of this curve is the argument for why it doesn't have to end badly.
The Plateau: The Long Stay at the Top
This is the part most projections get wrong. The peak is not a point we pass through. It's a place we stay.
Slop doesn't start declining the moment AGI-shaped systems exist, because the thing that brings slop down is not intelligence. It's governance in the broadest sense: guardrails that hold, evaluation methods that measure, institutions that move, definitions that mean something. None of that arrives on a training run's schedule.
Regulating a system you can't fully specify is genuinely hard. The labs won't have clean guardrails ready, not because they're careless, but because the science of controlling systems smarter than their overseers is younger than the systems themselves. Legislation moves in years. Capability now moves in months. International coordination moves slower still. If the plan is to wait for a treaty, bring snacks.
So the curve goes flat at its highest point. Capability keeps inching up, control crawls behind it, and the gap between them stays wide open.
Amodei attaches a condition to his country of geniuses that I only half believe: that once it exists, everyone will simply know. The first half of that is my peak. The second half is where I part company with him. The knowing lags the existing, and that lag is the plateau.
Roy Amara's law is usually quoted about markets, but it describes this plateau exactly: "We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run." The plateau is the short run. It will feel endless from inside. It's also temporary.
If you run a company, the plateau is where your planning horizon actually lives. Not the rise, you're already in the rise. Not the decline, that arrives later than the optimists say. The plateau. Budget for it.
The Decline: When Everything Stops Feeling Like Slop
Eventually the curve breaks downward, and it breaks for a specific reason. Intelligence starts getting spent on oversight instead of only on output.
The same capability that generated the flood becomes the thing that filters it. Systems verify other systems. Alignment techniques mature from research papers into infrastructure. Institutions finish the slow work of formalizing what was ungovernable at the peak. Taste, meaning knowing what not to do and what not to touch, finally scales with power instead of trailing it.
On the graph this is the right side. The y value falls and keeps falling while the x-axis runs on. Intelligence doesn't decline. Slop does.
The endpoint is a strange and honestly appealing world, one where nothing feels like slop. Not because generation got rarer, but because judgment got cheap. Abundance with discernment. The curve completes its descent and hugs the axis, and "slop" becomes a period label, the way "spam crisis" describes an era of email instead of email itself.
But here's the constraint the whole thing hangs on. You can't get to the decline without going through the peak. The descent is purchased at the plateau. There's no path where slop fades quietly while systems stay comfortably sub-AGI, because the oversight that kills slop is itself a product of the intelligence that caused it.
Controlled Arrival
Nothing on this curve says the summit has to hurt.
The danger at the top was never intelligence itself. It was the gap: capability arriving faster than the judgment to aim it. Close the gap and the same summit reads completely differently. A system that can pursue goals across long horizons is a catastrophe when nobody can bound what it does, and it's the most useful thing our species has ever built when somebody can. Same system. The only variable that moved was control.
That's the whole case for doing this deliberately rather than quickly. AGI reached in a controlled way isn't a rupture in human existence. It's a transition, and a survivable one. The disruption people brace for comes from arriving at the summit without brakes, not from the summit existing. Harnessed properly, the thing at the top of this graph runs the science we don't have enough humans to run, takes apart diseases we had quietly written off, and absorbs the enormous share of human work that was never the interesting part anyway.
The mess we're in right now is, oddly, the encouraging part. Slop is what an uncontrolled arrival looks like in miniature, at a scale where the worst case is a bad afternoon and a reverted branch. We're getting to rehearse the control problem on pull requests before anyone has to solve it on infrastructure. Most civilizations don't get a practice round.
The optimistic case was never that we get lucky. It's that we get ready.
The Caveat: We May Not Recognize the Peak
One honest limitation of this graph. The quantity on the x-axis keeps rising, but our labels for it don't hold still. And if the jagged framing is right, intelligence was never a single number to begin with.
Demis Hassabis is already fighting over that label. He thinks the definition of AGI is being watered down, and he holds the strictest bar in the field: every human cognitive capability, no partial credit. Altman has gone the other way and taken to saying the term isn't useful anymore at all. The people closest to the thing cannot agree on what would count as its arrival.
AGI is a moving target. What we call AGI next year may be defined differently than it is today: narrower, broader, split into grades, or retired entirely for better words. The system sitting at the peak of this curve may never be called AGI by anyone. It may be something adjacent, close enough to force the control problem, not close enough to end the definitional argument. We don't have a clear picture of what AGI looks like even a year out, and the curve doesn't require us to.
The peak is defined by a relationship, capability exceeding control, not by a word crossing a threshold.
That's also why the peak will be visible mostly in hindsight. We'll know we were there when the decline starts, the same way you only know the top of a hill by walking down the other side.
What I Take From the Curve
Drawing this changed how I think about the current moment in three ways.
Slop is a leading indicator, not a nuisance. Rising slop means capability is arriving faster than judgment, which is exactly the condition that defines the approach to the peak. Measure it inside your own organization. How much generated work gets reverted, re-reviewed, or quietly deleted? That number is your local position on the x-axis. Nobody has ever volunteered it in a board meeting. Be the first.
The scarce asset on the plateau is verification. In the rise, advantage went to whoever generated fastest. At the peak and across the plateau it flips. Advantage goes to whoever checks fastest: evals, review infrastructure, provenance, taste. Companies building verification are buying the dip on the far side of this curve.
The endgame is worth the middle. The right side of the graph, high intelligence and near zero slop, is not a fantasy tacked on to make the essay end well. It follows from the same mechanism that produced the mess. The systems get good enough to clean up after themselves, and then good enough not to make the mess at all.
The curve only runs in one direction. We don't get to choose whether to climb it. We do get to choose what we build for the plateau, and the plateau is where the whole thing gets decided. Build it well and the far side of this graph is the best place people have ever gotten to stand.