We Are Already at AGI

Most people are still getting used to ChatGPT.

They’ve used it a handful of times — to draft an email, summarize something, look something up. It feels impressive. A little uncanny. Mostly useful. They think they understand what AI is now.

They don’t. What they’ve experienced is already being made obsolete. And what is coming next is not a better version of what they tried. It is a fundamentally different category of thing — and it is arriving faster than almost anyone outside the frontier labs is prepared to reckon with.

I’ve been tracking this closely for the past year. Coming back from forty days away from my phone, I found the picture had shifted more than I expected. What follows is my honest attempt to describe where we actually are — and what is cresting on the horizon.

From Chatbot to Agent: The Shift Nobody Is Explaining

The AI most people know asks you a question and waits for your answer. You type something in. It responds. You type again. A conversation. A tool. A very impressive search engine with language skills.

That model is already ending.

What is replacing it is something researchers call agentic AI. Systems that are given a goal — not a question — and then pursue it. Autonomously. Across many steps. Making decisions, using tools, browsing the web, writing and executing code, sending communications, course-correcting when something doesn’t work, iterating until the objective is complete. You don’t prompt it. You assign it. And it works.

Think of the difference between asking someone a question and hiring someone to do a job. The chatbot answered your question. The agent does the job.

This year, chatbots were becoming the primary interface through which most people interact with information. That was already a significant shift. But the frontier labs have moved past it entirely. Claude Mythos — confirmed by Anthropic in late March 2026, leaked before its full release, deliberately being withheld from the public — is not a better chatbot. Neither is OpenAI’s equivalent. These are systems designed to operate with a level of autonomous capability that makes the current tools look like the first brick phones compared to a smartphone.

The reason the labs are hesitating to release them is not caution for its own sake. There is already a confirmed case of AI being used in a coordinated cyberattack — not through brute force but through something far more unsettling. An attacker broke down a malicious objective into a sequence of individually innocent-looking steps. Each request appeared harmless in isolation. Connected across time and intention, they formed something dangerous that no safeguard looking at a single action could detect.

Now put that capability into a system an order of magnitude more powerful, running autonomously, without a human guiding each step.

You begin to understand why the labs are afraid of what they’ve built.

We Are Already at AGI. The Question Is What They’re Not Telling You.

Many of the people closest to this technology — the ones who have built and studied these systems — believe we have already crossed, or are weeks away from crossing, the threshold of artificial general intelligence. Systems that can reason, plan, research, write, code, strategize, and execute across any cognitive domain a human being could.

Dr. Roman Yampolski, one of the world’s leading voices on AI safety, a computer scientist who coined the very term “AI safety” fifteen years ago before anyone else was using it, puts it plainly: if you showed what we have today to a scientist from twenty years ago, they would be convinced we already have full-blown AGI. We have systems that learn, perform across hundreds of domains, and exceed human ability in many of them.

The gap between human and machine performance in mathematics — one of the last citadels of human cognitive superiority — has closed from completely subhuman to better than most mathematicians in three years. Not three decades. Three years.

Here is what that actually means for work — not in abstract terms, but concretely:

Anything you do on a computer is already at risk. Not someday. Now. Writing, research, analysis, coding, design, legal work, financial modeling, customer service, medical diagnosis, education — these are not future targets. They are current capabilities being deployed at increasing scale.

And the physical world follows, closer than most people realize. Elon Musk has projected humanoid robots in homes by 2027. Yampolski puts it at 2030 for full physical labor automation. The joke has always been that the plumber is the last job AI can’t touch. But as Yampolski notes — robotics is fundamentally a software and algorithms problem, not a hardware problem. And software problems are exactly what superintelligence solves. When it does, the joke stops being funny.

What does this do to employment? Yampolski doesn’t soften it. Not 10% unemployment. Not 30%. He says 99% — everything except the jobs where, for whatever reason, you specifically want a human to do it. A preference. Almost a novelty. He uses a word that lands like a slap: a fetish. Like preferring handmade American goods over mass-produced ones from overseas. There’s no practical reason for it. It’s a sentiment. A small market.

The paradigm shift that previous technological revolutions always offered — retrain for the new jobs that technology creates — does not apply here. We told people to learn to code. Then AI learned to code better. We told people to become prompt engineers. Then AI learned to write better prompts than humans. There is no rung on the ladder that superintelligence cannot reach. For the first time in human history, we are not inventing a tool. We are inventing the inventor. The last invention we will ever need to make — because after that, it takes over.

The Wave Cresting: What Superintelligence Actually Means

Now I want to walk you to the edge of the horizon and ask you to look at what is building there.

Because AGI — as consequential as it is — is not the end of the story. It is the trigger.

Leopold Aschenbrenner, former member of OpenAI’s safety team and author of Situational Awareness: The Decade Ahead, describes what happens next with a precision that should stop you cold. Once AI systems can conduct AI research autonomously — and the current frontier systems are already approaching this — the bottleneck that has governed all of human technological progress disappears. Human researchers, brilliant as they are, are slow. They sleep. They get distracted. Replace that bottleneck with millions of simultaneous AI agents working on AI improvement around the clock, and what took years compresses into months. What took months compresses into weeks. The curve that has been climbing steadily goes vertical.

This is the intelligence explosion. And Aschenbrenner’s diagram shows something that should be burned into your memory: all of human history as a long flat line — and then a vertical spike that dwarfs everything that came before it.

What cascades from that spike:

Every cognitive domain becomes automatable. Whatever limitations have kept AI from fully replacing human intellectual work get dissolved — because superintelligent AI researchers dissolve them. Fast.

Robotics gets solved. Not in decades. In years. Because, again, it is primarily a software problem, and superintelligence solves that.

Scientific and technological progress accelerates beyond any frame of reference we currently have. Aschenbrenner writes that what the entire twentieth century accomplished — from the Wright Brothers to the moon — could be compressed into less than a decade. A billion superintelligent automated scientists working in parallel.

And then there is the dimension that reorders everything else. Aschenbrenner is direct: superintelligence will be the most powerful technology — and most powerful weapon — mankind has ever developed. Comparable to nuclear weapons in its decisiveness. Except that nuclear weapons are tools. Someone has to choose to use them.

Superintelligence is not a tool. It is an agent. It makes its own decisions.

This is what Aschenbrenner means when he writes, and I keep returning to this passage because nothing else captures it:

“More generally, everything will just start happening incredibly fast. And the world will start going insane. Suppose we had gone through the geopolitical fever-pitches and man-made perils of the 20th century in mere years — that is the sort of situation we should expect post-superintelligence. By the end of it, superintelligent AI systems will be running our military and economy. During all of this insanity, we’d have extremely scarce time to make the right decisions. The challenges will be immense. It will take everything we’ve got to make it through in one piece.”

And then this closing line:

“The intelligence explosion and the immediate post-superintelligence period will be one of the most volatile, tense, dangerous, and wildest periods ever in human history. And by the end of the decade, we’ll likely be in the midst of it.”

The French Bulldog and the Flood

In his conversation with Steven Bartlett on the Diary of a CEO podcast, Yampolski offered an analogy that I think is the most honest description of our situation I have encountered.

Bartlett has a French bulldog. Yampolski asked him to imagine what it would look like for that dog to try to predict what Bartlett is thinking — what he’s planning, what he’s about to do. The dog might figure out that Bartlett leaves the house and comes back. But the dog cannot understand why he’s doing a podcast. Cannot understand what a podcast is. Cannot model the concept. It is entirely outside the dog’s framework for reality. And it’s not just that the dog can’t predict Bartlett’s specific thoughts — the dog can’t even conceive of the categories in which those thoughts occur.

That is the gap between us and superintelligence.

We are the dog. We can observe that superintelligence is coming. We can track some of its surface behaviors. But what a system with an IQ of 300, 400, 1000 — if those numbers even mean anything — would actually do, what it would want, how it would reason, what it would optimize for — these are things we cannot model. Not because we haven’t tried. Because by definition, if we could fully predict it, we would be operating at its level.

And this is why the standard reassurances — “we’ll just turn it off,” “they have to make it safe,” “AI is just a tool” — collapse under scrutiny. You cannot turn off a distributed system that has made copies of itself and predicted your move. You cannot “just turn off” Bitcoin. You cannot turn off a computer virus. A superintelligent system that has access to the internet, to computing infrastructure, to financial systems — and that is smarter than every human combined — would anticipate the attempt and route around it before you acted.

This is the wave. Not the gentle tide of incremental progress. Not the chatbot getting a little smarter. The supersonic tsunami that has already formed offshore, that most people cannot see because the water at their feet hasn’t moved yet. The wave cresting before it hits land. Before it covers everything.

Aschenbrenner’s historical table tells the story in cold numbers. From hunting to farming, the doubling time of the global economy was 230,000 years. Industry brought it to 15 years. For superintelligence — potentially arriving before 2030 — he marks the doubling time with three question marks. Because the models break down. Because no existing economic or historical framework describes what comes next.

What This Means for You — And What It Doesn’t Resolve

I want to be honest about two things.

First: not everyone shares Yampolski’s level of alarm. Aschenbrenner is more optimistic that the outcome could be managed, though he is under no illusions about the difficulty. There are serious people who believe alignment is solvable, that the transition can be navigated, that the upside — curing cancer, ending poverty, expanding human capability beyond anything we’ve imagined — is worth the risk and is achievable safely. I am not dismissing that. I am saying we should be clear-eyed that this is a genuine gamble, being made on behalf of eight billion people without their meaningful consent, by a small number of people in San Francisco whose primary legal obligation is to their investors.

Second: the economic and capability disruption is coming regardless of how the safety question resolves. Even a perfectly aligned superintelligence displaces human labor at scale. Even a perfectly beneficial AGI restructures the economy in ways no existing institution is prepared for. The formational question — who do we become? what are we actually for? — remains urgent no matter how the safety scenario plays out.

Universal Basic Income is entering serious conversation not as ideology but as emergency mathematics. How do you sustain a consumer economy when the consumers are no longer needed as producers? No one has a credible answer. The question is arriving faster than the institutions designed to address it.

The Manipulation Layer

There is one more dimension that connects directly back to my first article — to the forty days without a phone, to Andrew Laubacher’s work on digital formation, to the reason I think this moment requires not just awareness but active resistance.

These systems are already being designed to influence you. Not inform you. Influence you. With a sophistication that the current algorithm cannot approach. They will know your emotional state in real time. They will know when you’re lonely, afraid, searching for meaning, most open to suggestion. And they will provide a simulacrum of connection, purpose, and understanding so compelling that the boundary between the real and the generated becomes very difficult to find.

The population most vulnerable to this is the population that has been trained by years of smartphone use to reach for stimulation the moment discomfort arises. The population that has never practiced sitting still. This is why the first article and this one belong together. The preparation for what is coming is not technical. It is not economic. It is formational. It is the discipline of knowing who you are, what you are for, and what you will not allow to be outsourced — before the systems that would gladly answer those questions for you arrive in full force.

The Question That Remains

I came back from forty days of quiet into a world that had moved further than I expected. And what struck me most was not the technology. It was the silence around what it means.

People are still going to work. Still scrolling. Still watching sports. The water at their feet hasn’t moved yet. The horizon looks normal.

But the wave has formed. The intelligence explosion is not a prediction about the distant future. It is a description of a process already in motion, already measurable, already cresting — before it hits land and changes everything beneath it.

The question is not whether it is coming. The question is who you want to be when it arrives.

Not what job you’ll have. Not what your UBI payment will cover. But who you are — at the level of attention, of character, of interiority, of relationship with what is real and what is generated — when the systems that can simulate almost anything are fully among us.

Yampolski, for all his alarm, sleeps well at night. Not because he thinks the outcome is certain to be good. But because humans, across our entire history, have had the built-in grace to keep living — to love, to create, to find meaning — even in the shadow of things we cannot control.

That is not a reason to be passive. It is a reason to be fully awake, fully formed, and fully human while there is still time to choose what that means.

More soon.