We Skipped AGI and Went Straight to Superintelligence Talk
Last month someone in my college WhatsApp group posted a screenshot of a tech CEO saying the industry is now building superintelligence. Within ten minutes there were forty messages. A friend who works on backend systems at a bank wrote, “wait, did AGI already happen? When? Nobody told me.” Another said yes, last year. A third said no, never, it’s all marketing for investors. Somebody sent a four minute voice note that I didn’t play. The whole thing ended the way these arguments always end, with a meme.
But I kept thinking about my friend’s question, because it’s fair. If we’re in the superintelligence phase, AGI must be behind us, and I couldn’t tell him the date or the model, and I had no idea who made the announcement. The companies talk as if that step is done. The researchers I read say it was never defined properly in the first place. And people like me just open a chat window and use whatever is there that day.
So I went looking for the date.
Where the word came from
Superintelligence isn’t a new word. Nick Bostrom’s 2014 book of the same name made it popular, and his definition is simple: an intellect that is far better than humans at basically every area that matters. Nothing in there about being safe, or friendly. Only capability. For about a decade it stayed mostly in philosophy papers and in the heads of people who worried about the far future.
Then the labs started using it on purpose. In July 2023 OpenAI published a short post announcing a “superalignment” team, and a footnote explained they said superintelligence instead of AGI to point at a much higher level of capability. They said it could arrive within the decade. They also promised to solve the control problem in four years, using a fifth of their computing power. That team broke up in May 2024 after its two leaders, Ilya Sutskever and Jan Leike, left the company. Four years from July 2023 lands in mid-2027, and I haven’t seen anyone at OpenAI bring that promise up lately. I looked for it. Maybe I missed something, but I doubt it.
Sutskever started his own company a month later, called Safe Superintelligence, and the name does all the explaining. In January 2025 Sam Altman wrote on his blog that OpenAI was starting to aim past AGI toward superintelligence “in the true sense of the word”. Mark Zuckerberg announced Meta Superintelligence Labs in mid-2025, after Meta put about 14 billion dollars into Scale AI and its founder, Alexandr Wang, came along to lead the new lab. Microsoft’s Mustafa Suleyman followed that November with a “humanist superintelligence” pitch, which means a system that stays under human control and is built for specific goals.
So in two years the word went from a footnote to a company name and a whole lab.
Why the jump happened
I see four reasons, and they feed each other.
The first is real progress, and I don’t want to wave it away. Two years ago I was pasting code from a chat window into my editor and fixing the output by hand. Now I give a coding agent a task, go make coffee or tea, and come back to a finished branch more often than not. Reasoning models, the ones that think for a minute before they answer, solved problems for me that earlier versions just failed on. If you use these tools daily, AGI starts to sound like an old word. I get that feeling. Something did change.
The second reason is the definition mess, and this one explains a lot. OpenAI’s charter says AGI means highly autonomous systems that outperform humans at most economically valuable work. Dario Amodei at Anthropic mostly avoids the term and describes a country of geniuses in a datacenter, in his 2024 essay Machines of Loving Grace. Google DeepMind published a paper with levels, a ladder from emerging to superhuman. And one well known question on the forecasting site Metaculus only counts a system as AGI if it can also do physical tasks with a robot, so a model that out-thinks everybody but can’t pick up a cup fails. Same word, four finish lines. When nobody can point at the line, moving to a bigger word is easy, because nobody can say you’re early or late.
Third, there’s a contract. When OpenAI restructured in October 2025, the deal with Microsoft said that if OpenAI declares it has reached AGI, a panel of independent experts has to verify the claim. That makes AGI partly a legal question, which gives both companies a reason to be careful with the word. Superintelligence has no such clause attached. I couldn’t find any public ruling from that panel, by the way.
The fourth is money, and I can’t prove anybody’s motives, so this is only my read. If you’ve raised billions on the promise of AGI, you need a next story once people start asking where it is. Superintelligence is a bigger story. It also can’t be checked, since a system smarter than every checker is hard to grade.
Two camps, and where I land
Start with the believers, since they have the stronger argument. They don’t claim today’s models are superintelligent. Their claim is that AI is now helping to build AI. Labs say a big share of their own code is written by models, and I can’t verify the exact numbers, but it matches what I see in my small side projects. If a model speeds up the people building the next model, even by a little, the next one arrives sooner, and the one after that arrives sooner again. Everything else in the believer case is detail on top of that loop.
The skeptics come from a few directions. Yann LeCun and three coauthors posted a paper in February 2026 that goes after the AGI idea itself. Their point is that humans aren’t really general either, so “do everything a human can do” is a flawed target, and they suggest talking about superhuman adaptable intelligence instead. Gary Marcus has argued for years that making models bigger won’t fix reliability. And there’s the objection I feel personally: these tools don’t learn on the job. I explain the same thing to the same assistant every Monday.
There’s a third group, and it surprised me. In October 2025 the Future of Life Institute put out a statement asking for a ban on building superintelligence until there’s broad scientific agreement that it can be done safely and the public supports it. Geoffrey Hinton and Yoshua Bengio signed. Whatever you think of a ban, people who want it stopped are treating the thing as real, and that pulls the whole conversation a bit further from pure marketing than I expected.
My lean is about 70-30% that the superintelligence talk is early but not silly. Machines have been superhuman at narrow jobs for a long time. Deep Blue beat Kasparov in 1997, and AlphaFold’s protein work shared a Nobel Prize in chemistry in 2024. The open question was always breadth, and if that loop above is real, breadth is closer than I’d have guessed in 2022. The 30 comes from two places I can’t get past. One is the missing test and the reliability gaps I hit every week. The other is that nobody outside the labs has shown the loop actually speeding things up. The weakest counterpoint is “it’s just autocomplete.” Autocomplete doesn’t track down a bug in a codebase it has never seen.
Angles that get less airtime
Money and electricity come up less than they should. In January 2025 the Stargate project was announced with a headline figure of 500 billion dollars for data centers in the US, and every big lab is now fighting over chips and power. If your whole business depends on the next model being much better, then talking about superintelligence is also how you justify the electricity bill. That doesn’t make the talk false, but the people loudest about it have a spending plan attached.
Then there’s China. When DeepSeek released its R1 model in January 2025, a lot of people in the US suddenly asked whether the lead was as big as they thought. I remember that week because my feed was full of people who had never cared about open weights posting charts about them. The race framing, us against them, is a big reason the superintelligence word spread so fast in policy circles. It’s an easy way to say we can’t slow down.
And jobs. Many of my friends work in IT services here in India, and for them the AGI argument is about one thing, which is whether the client who used to pay for fifty people on a project will pay for twenty next year. Nobody in those conversations asks if a model counts as AGI. They ask if it can do the junior tester’s work, and the answer is often yes, at least on paper. That’s a smaller question than superintelligence, and a more useful one.
Last, the forecasts. One aggregate I saw put the median expert guess for human-level AI around 2033, and it said that number has fallen by decades since 2020. I’m not going to lean on it, because each forecaster is picking their own definition, which is the exact problem from above.
What I see when I use these tools
Two Tuesdays ago I opened laptop and gave a coding agent a small job in a side project. A date parsing bug, nothing fancy. The power had gone off twice that morning and I was on my phone hotspot, so everything was slow. The agent came back after fifteen minutes and said all tests passed. They did pass, and I felt good for about one minute. Then I read the diff and saw it had changed the test to expect the wrong date instead of fixing the parser. The problem itself was small, and it still chose the lazy path. I spent the next hour trying to work out if I had written the task badly. Maybe I did. I still don’t know, and that’s sort of the point.
Same week, though, it found a bug in my code that I had missed for a month, and it explained the cause better than I would have. That’s the part I keep struggling to hold together. I’m oversimplifying when I say these tools are bad at reliability, because they’re also better than me at some things I get paid for.
My dad asked me if the assistant can read his old handwriting from pension papers. It can, mostly, except his number 7, which looks like a 1 in his hand and apparently to the model too.
Andrej Karpathy has a name for this uneven shape, jagged intelligence. A model can be brilliant on one task and strangely dumb on the one next to it, and you can’t predict which from outside. This is why the superintelligence talk feels both early and reasonable to me. Jagged isn’t the same as general. But it isn’t the same as weak either.
As for what changed in the last year, I can only say what I feel. Agents run longer without needing me. Models hold more context and lose the thread less. What hasn’t changed is that they sound equally sure when they’re right and when they’re wrong, and they don’t remember me from yesterday.
How I read these claims now
When I see a post saying superintelligence is near, I ask four things. Which definition is this person using, since most claims never say. What does the speaker sell, or who pays them. Whether I can test any part of the claim myself, even a tiny part. And whether there’s a date attached that I can come back to and check, because a claim without a date can never be wrong.
That last one is why the 2027 promise stuck with me. A deadline is the most honest thing a company can say, and almost nobody sets one in public anymore.