I Think AGI Just Happened
I think we just saw the birth of AGI, and it's from the most unexpected place. At least for me.
I think it arrived in the form of a product from Anthropic called Claude Tag.
My definition of AGI, which is what I think most of us actually care about, is just:
An AI system that can replace an average knowledge worker.
The "G" in AGI stands for generality, and there's wide disagreement on how measure that. But one thing we know requires lots of human-level-or-above generality is the chaos of knowledge work. Even the most basic knowledge work role is full of dozens or hundreds of random mini-tasks in the course of a day or week.
So my thinking here is that if something has general enough intelligence to be a knowledge worker, that's a pretty solid standard for generality.
In November 2023 I wrote Why We'll Have AGI by 2025-2028:
My prediction is a 60% chance of AGI in 2025 and a 90% of AGI in 2028.November 2023
The argument for how we'd get there was systems, not components. We wouldn't need one magic model to be AGI by itself. A system of AIs working together toward a shared goal would get there first.
Then in March I wrote We Are Confusing Two Types of AGI, which split AGI into Hard AGI (the academic, Computer Science version) and Soft AGI:
An AI system implemented as a Product/OSS Project that emulates learning generality well enough to replace knowledge workers.
And that post made another specific call:
I believe we'll see one or more commercial products in 2026 or 2027 that do this.March 2026
I even described how that product would behave. It onboards like a regular human. It does orientation. It takes instruction from the manager. It interacts with the team. It works. It ships results.
That's the shape Claude Tag just showed up in. You install it into Slack, where knowledge work already lives, and you treat it like a coworker.
What it actually does
Here's how Anthropic's docs describe it:
Anyone in a channel can tag Claude into a problem and hand it work: reproduce a bug and open a pull request, turn a decision thread into a doc, assemble the state of a project.Claude Tag documentation
Here it is picking up a Sentry alert, reproducing the bug, opening the PR, and asking the code owner for review:
Here it is doing the boring recurring work, a Monday pipeline digest pulled from Salesforce on a schedule:
And in the webinar, Anthropic literally tells you to manage it like a person. Their words: "Brief it like a capable new hire: set the goal, let it own the process, verify the result."
At install you wire it into your repos, your docs, your data warehouse, your monitoring, and your ticketing. It runs in a sandbox hosted by Anthropic, and it works out of the channels your team already uses.
A GPT-4 moment for worker replacement
I think this is like a GPT-4 moment, but for worker replacement.
Think of it this way. All these companies have been trying to "replace" workers with AI, but they have no idea how to do that. They don't know what the work is. They can't give specific instructions to AI. AI doesn't have enough tools. The whole thing is not scaffolded well enough and integrated well enough with their work to actually be able to do it.
I think this is the first moment via this product where it's so deeply integrated, and all you have to do is have regular conversations with it like a co-worker.
It's actually doing all the Claude Code stuff that we're all so impressed with. It's doing that all on the backend by itself, so it's literally like you're talking to a co-worker.
That's the part that is actually going to make laying off like half the workforce possible.
Basically, up until now, what replacing people with AI meant was giving that work to somebody else who was really good with AI.
Now you're not giving it to another person who has Claude Code, who's like one of us and a total ninja with AI. Now you're actually just handing it to a co-worker, in this case Claude.
Anyway.
After reading all the Claude Tag docs, and watching a full Anthropic webinar on the project, I think this might be it.
The tangible thing I was waiting for was an actual PRODUCT being dropped that can replace a human worker.
Not like general tech. Most people can't use Claude Code. But something that you literally install and you can then treat it like a knowledge worker.
I think Claude Tag might be it.
Meaning, I'm 85% sure we just hit what I believe to be the best definition of AGI.
I guess the reason I'm tripping out about this is because I think everyone, including myself, was waiting for a technology upgrade to get AGI.
When in fact it might just be in the form of an integration.
I always knew this was going to come from a product, not a model change. I just thought it would be some startup—a cool name, a cool logo, and the whole basis of the company is a drop-in knowledge worker replacement. Turns out it's a sub-product inside the Anthropic platform.
And you can damn sure bet OpenAI will be there soon. Microsoft can do this too. Google can't, because their product management is paralyzed—they're out of this game until they fix that. That's what sets Anthropic apart: they are so good at being connected to what people need and rolling out product fast. Listen, ship, iterate. Listen, ship, iterate.
Absolute Insanity: We got AGI not from a new model or a sexy new startup, but from a Slack integration.
A single prompt
I've been thinking about this all day and it keeps re-freaking me out. Here's the crux of it in a single example.
Imagine there's an employee named Chris. He's been there for 13 years. He's basically an admin/coordinator for the company and helps with lots of different things, including putting together plans and summaries, organizing meetings, etc.
His managers have been mostly happy with him, but he's not the best at writing reports, and he needs to get lots of help when he does anything technical.
His new manager has just been onboarded, and he has a Claude Tag. Here's the prompt that he sends in:
We are about to fire Chris. I need you to: read every email he has received or sent over the last 13 years. Pull up all transcripts of all of his meetings. Look at every document he's written. Pull his entire Slack history. Do a deep analysis of what all tasks he was doing, which ones he did well, and which ones he did poorly. Then put together a set of scheduled tasks that perform those tasks far better than he was doing them, with the outputs being sent to the same places he was sending them, at the ideal time.
So, like, 37 minutes later, it comes back and says:
Chris was writing three different reports per week. He was sending an average of 48 emails per week. He was doing coordination on the following 18 meetings. He was doing a little bit of report analysis every month. And here are 37 things he was asked to do over his time here that he was never able to get good at. I have looked at all those tasks, confirmed that no one else is doing them, and I have scheduled all the work he was doing to happen at the ideal time, with the outputs being sent to the appropriate places, including scheduled tasks for all the things he was asked to do that he wasn't able to. Is there anything else that you need?
This is what Karpathy was so on about in that post he got so much shit over:
People roasted him for it, calling it overhyped marketing for a basic Slack bot. He shot back that it shouldn't be dismissed as "a mediocre Slack bot"—the hard part was never the bot, it's making an org-level agent actually work inside a company.
Everything I just described was already possible if you spent weeks or months assembling context throughout your company, and you had a bunch of Claude Code wizards who could go harvest all the stuff. They basically work for management and do a bunch of projects like this. All of it was possible a few months ago. But that's a lot of wizardry, a lot of context engineering, a lot of scheduled task management. All sorts of things have to go into making that happen, and that friction is the reason most companies haven't done it.
The difference here is that what I just described is a single fucking prompt inside of Slack.
Let me say this a different way. Some manager just replaced an employee in a single fucking prompt.
And it could very well be that the work produced by this new Claude instance will be far superior to the work Chris was doing.
And that's not even getting slightly silly with this yet. The moment that system comes back in 37 minutes and says, "Yeah, I just automated Chris," the very next thing they're going to ask is, "Oh, how about you do that for my entire department?" No, actually, why don't you spend some tokens and do it for the entire company?
Or: "Find me the people who are doing the least amount of work at the lowest quality, and/or are creating conflict within the organization, or acting against the organization's culture."
All without leaving Slack.
Again, doing all of this with a dedicated team of AI ninjas, writing their own scaffolding and building their own custom applications, is one thing. It's completely different when everyone in the company can do things like this from within the system they already work in, never having to leave and go into an "AI interface."
This surprised me. It's the exact same thing, but a hundred times more powerful, because it's riding on top of Slack and Claude, which are already heavily accepted inside of companies.
A different character of layoffs
People have been wondering why there haven't been more direct replacements of employees by AI.
The reason is because we haven't yet had a product that was general enough in the tasks that it could do, with low enough onboarding friction, to do actual work.
That just changed, and I think we're about to see a completely different character of mass layoffs due to AI.
This time it won't be in anticipation of AI getting good enough, but because somebody has already literally replaced them and has the ability to compare their work side by side.
So the question becomes how many tokens it will cost to replace a slightly below-average, average, or halfway-decent knowledge worker versus what their total comp was.
I'm guessing that for many, many jobs, the token cost will be somewhere between 1% and 50% of what they are paying the human worker.
And by the way, this is not because those workers were trying their best, being super creative, and giving the job their all every day. Absolutely not.
Most people who are average or below average at a particular role today are that way because they shouldn't be doing that job in the first place, because it's one of David Graeber's bullshit jobs, or because management sucks, or because the company is not inspiring in the slightest.
They clock in every day, dead-eyed, finding ways to do the absolute minimum and praying for the end of the day. There's not just memes about this. It's an entire culture of memes.
We all knew that most of these corporate jobs were horrible for the human soul. Right until the end of 2022, when something came along to take them away.
Think about what Claude will say about you
Here's the part everyone needs to sit with.
Once Claude is inside your company, you can literally ask it to figure out what Sarah or Raj or Chris actually does. Find every Slack message they sent. Every meeting they sat in—read the transcripts. Every email in and out.
And not just the volume of work. The impact. What tasks are they actually doing? How many new ideas have they had? Have they come up with new product ideas? Do they encourage others? Are they a nurturing type who helps out across the team? Do they support the culture or quietly undermine it?
AIs understand all of these things. The only thing they need is inputs.
Yes, the quality of the prompts and the criteria will matter, and some companies will get that wrong and it'll be bad. But in general I think this locks in fairly quickly, because it's in companies' interest to keep the best possible people and get rid of the worst.
Companies will literally be able to say: this is what a model employee looks like for us. They produce this type of value. They have this type of attitude, this type of energy. They come up with new ideas. They help customers. They add signal to the organization—to our procedures, to how we do things. They're not simply executing tasks.
Because that's the question now: what can you do that Claude cannot?
In pretty much all cases, that domain is encouraging others, collaborating with creativity, coming up with new ideas for products and features, producing delight for customers. AI will get really good at delight in the SLA sense—fast response times, high-accuracy emails, great reports. But coming up with its own ideas based on a deep understanding of what a customer or a coworker needs? That requires internal fire. Internal vibration.
Humans have that. It comes from their experiences and passions and curiosity—from seeing bad products and unhappy customers and wishing the world were different than it is today. Wishing a product existed that doesn't exist yet. We could probably prompt something like that into an AI, but it's going to be unnatural. It's not going to be the same as from a human. At least not anytime soon.
The universal employee
It gets even simpler than that, because you no longer need to be an expert in Claude Code or AI harnesses. Even that is being abstracted away now. The only thing that matters is your ideas that don't exist yet in the world. New things to build. New products, new features.
So you hire for energy. You hire for positivity. You hire for creativity. You hire for culture fit. And 1 plus 1 equals 7—except each of those ones has hundreds or thousands of Claudes backing them, building designs, building prototypes, creating reports, doing testing.
Product merges with engineering. Engineering merges with design. Everyone is a vision person. Everyone's an idea person. Everyone's a PM. Everyone's an engineer. There will obviously still be hardcore engineers and hardcore designers, but everyone at the company becomes one of these universal employees. Super well-rounded. Super AI-native.
The people in danger
The people in the most danger are the people who have been trained not to have ideas. The people who have been trained to execute other people's ideas their whole life. That's what they were taught in school. That's what they were trained for. And that's what they're doing.
This is the whole point of my Human 3.0 thing—to encourage them to light up their inner self, their younger self that was curious and interesting and asking questions, before it got beaten out of them by peers and school and work.
The type of person who's going to thrive in this world is somebody who up until now has only been able to execute on a tiny, minuscule fraction of their ideas. They want to make movies. They want to write novels. They want to make music. They have ideas for new ways of doing healthcare for elderly people, for robots that help people with declining mental faculties.
The vast majority—99.999%—of people's ideas have never been attempted. Never encouraged, never possible, because of the giant stack of requirements between the idea and reality: money, connections, networking, timing, programming skills, design skills, a million different things needed to produce something end to end.
That is just no longer the case.
Inside a company, that manifests as the people with ideas for how to grow the thing—sales campaigns, marketing campaigns, new features, new products altogether. All those ideas now become possible.
If you've heard me talk about the K-shaped recovery before—where the people in a good position shoot up into orbit and the people in a bad position fall to the ground—well, that just got massively magnified.
Which is exactly what we knew was going to happen with AGI.
And it's now upon us.
Everyone reading this needs to be thinking very carefully about what Claude will say inside your company when it reads every single email and every single Slack message, and asks how aligned you are with the culture, how many new ideas you're putting out, and what you're actually doing that provides value AI can't.
This isn't a prediction. I'm describing a reality that's already happening.
We need to transition ourselves, as quickly as possible, from executors of other people's ideas to creators of ideas of our own.