When do you want the alarm to go off?
If you are waiting for a single, cinematic moment when everyone agrees that AI has crossed a line, you may be waiting until after the line has already moved your job, your industry, and your government's options. The uncomfortable truth is that "it" rarely arrives with a bang. It arrives as a string of ordinary Tuesdays where the tools get quietly better, the best versions stay gated, and the people who should be planning are still arguing about last year's demos.
This piece is about what "it is happening" actually means in practical terms. Not a mystical prophecy, not a labor market manifesto, and not a promise that artificial general intelligence is definitely next year. It is about accelerating feedback loops in AI capability and AI research, and why the gap between what most people have tried and what frontier users are building with is now large enough to distort public intuition.
What "it" is, and what it isn't
When people say "the singularity," they often mean different things. Some mean mass unemployment. Some mean conscious machines. Some mean a hard takeoff where systems rapidly become smarter than humans at almost everything. Those are separate claims with separate evidence.
The narrower, more defensible claim is this. AI systems are getting better at the work that improves AI systems, especially software and research workflows. If that continues, progress can compound. The compounding is the story. Even if you are skeptical of the most extreme outcomes, compounding alone is enough to reshape timelines, markets, and policy windows.
Why most people think nothing is happening
A large share of the public experience of AI is still a free tier chatbot, a few short prompts, and a model tuned to be fast and cheap. That experience is often underwhelming. It hallucinates. It forgets. It feels like a toy. If that is your only exposure, the rational conclusion is that the hype is ahead of reality.
But the most important capability gains in the last two years have not been "the model can answer trivia better." They have been about reliability in longer tasks, tool use, code generation that survives contact with real repositories, and systems that can plan, execute, check their work, and iterate. Those gains show up most clearly when the model is embedded in a scaffold: a coding environment, a test harness, a retrieval system, a set of tools, and a workflow that turns a chat model into something closer to a junior team.
This is why two people can talk past each other. One has tried a thin interface and seen a clever autocomplete. The other has tried a toolchain and seen a machine that can ship meaningful chunks of work. Both are describing reality. They are just living in different parts of the distribution.
The quiet engine: feedback loops
The most important question is not whether AI is impressive today. It is whether AI is becoming a lever on its own improvement. There are three compounding inputs that matter more than any single benchmark screenshot.
First is compute. Training and serving frontier models requires enormous infrastructure, and the industry is still building. The relevant signal is not a press release about a new data center. It is sustained capital expenditure, long-term power contracts, and the supply chain reality that chips, networking, and energy are being locked in years ahead.
Second is algorithmic efficiency. Even when raw compute growth slows, better training methods, better data curation, better post-training, and better tool integration can deliver large jumps. This is the part that makes "we'll hit a wall" arguments hard to time. Walls exist, but efficiency improvements often route around them.
Third is automation of the work itself. If models materially speed up software engineering and research, then the same headcount can run more experiments, build more tooling, and iterate faster. That is the loop that turns a steady trend into a steeper one.
The METR-style question: how fast is the curve moving?
One of the cleanest ways to think about progress is to stop arguing about labels and measure time horizons. Take tasks that humans can do, measure how long they take, and then ask what fraction of those tasks an AI system can complete successfully at different time budgets. If a model can reliably do "one hour" tasks, that is not just a little better than doing "ten minute" tasks. It implies planning, persistence, and error recovery.
Time-horizon framing also clarifies why coding is a window into everything else. Code is legible, testable, and composable. If a system can take a vague goal, turn it into a plan, write code, run tests, debug failures, and ship a working change, you are not looking at a parlor trick. You are looking at a general pattern for automating complex work.
Why "sigmoid" arguments feel comforting, and why they can mislead
Many technologies follow an S-curve. Early progress is slow, then it accelerates, then it saturates as the easy gains are exhausted. This is a useful mental model, and it is often correct.
The problem is that people use "it will sigmoid" as a substitute for identifying the bottleneck. Saturation happens when constraints bite. Sometimes the constraint is physics. Sometimes it is economics. Sometimes it is data. Sometimes it is coordination. If you cannot name the constraint and show it tightening, "sigmoid" is not a forecast. It is a vibe.
AI has multiple dials still turning at once. Compute investment, efficiency gains, better tooling, and broader adoption inside high-leverage organizations can stack. When several curves rise together, the combined effect can look like a step change even if each individual curve is merely exponential.
The real gap: who has seen what
A small number of people use frontier systems in ways that expose their true strengths. They run them inside coding agents, research pipelines, and internal tools. They see the model fail, then watch it recover. They learn what it is good at, what it is brittle at, and how quickly the brittleness is shrinking.
Most lawmakers, regulators, and institutional leaders are not in that group. Many senior executives are not either. That creates a dangerous asymmetry. Decisions are being made by people whose mental model is anchored to the weakest public experience, while the most capable systems are being used by a narrow slice of industry to move faster and capture value.
This is not a conspiracy. It is a distribution problem. The best tools are expensive, gated, and require skill to use well. The result is that society's "average intuition" lags the frontier by more than it should.
So what would actually convince you? Draw your Rubicon
If you want to avoid being emotionally whipped around by every new model release, you need a personal trigger. A Rubicon is not "when AI gets scary." It is a concrete, falsifiable threshold that changes your behavior.
Here are Rubicons that are specific enough to matter, and boring enough to be real.
One is sustained autonomy on real work. Not a curated demo, but an agent that can take a ticket in a production codebase, make a change, write tests, pass CI, open a pull request, respond to review comments, and do this repeatedly with a low error rate. When that becomes routine, software stops being a human-only bottleneck.
Another is research acceleration you can feel in release cadence. If frontier labs and open ecosystems start compressing what used to be yearly leaps into quarterly ones, that is a sign that iteration speed is compounding. Watch not just model launches, but the surrounding ecosystem: tooling, evals, agent frameworks, and the speed at which yesterday's "state of the art" becomes a baseline feature.
A third is the internal adoption line. When teams that are paid to be skeptical, such as security engineers, reliability engineers, and compliance groups, start building workflows that assume AI assistance as default, the technology has crossed from novelty to infrastructure.
A fourth is the economics of serving intelligence. If the cost per useful unit of work keeps falling while capability rises, adoption will not be a debate. It will be a procurement decision. The moment AI becomes cheaper than waiting, it spreads through organizations the way cloud did, only faster.
How to take this seriously without losing your mind
There is a temptation, when you feel the curve, to swing between denial and doom. Neither helps. The practical stance is disciplined attention.
Start by upgrading your own sample of reality. If your view of AI is based on a free chatbot and a few casual prompts, you are judging a jet engine by listening outside the airport fence. Try a modern coding assistant inside an IDE. Try a tool-using model with a real task and a test harness. Give it a messy document set and ask it to produce something you can verify. The point is not to be impressed. The point is to calibrate.
Then separate capability from deployment. A model can be powerful and still not be widely used because of cost, risk, integration friction, or governance. Those frictions matter, but they are not permanent. Many are engineering problems, and engineering problems tend to get solved when the incentives are large.
Finally, decide what you will do if your Rubicon is crossed. For individuals, that might mean investing in AI literacy, learning to supervise agents, and choosing roles that are closer to problem definition than routine execution. For organizations, it means building evaluation capacity, setting policies for safe use, and preparing for a world where speed becomes a competitive weapon.
The uncomfortable possibility: the "moment" is a lagging indicator
People ask when they should start "kicking and screaming." The honest answer is that if you wait until the change is obvious to everyone, you are choosing to react at the slowest possible time. The better question is when you should start paying for better information, demanding clearer benchmarks, and insisting that decision-makers experience the tools they are regulating.
Because the most dangerous part of exponential change is not that it is fast. It is that it feels slow right up until the day it doesn't, and by then the only thing left to debate is who got to steer.