How AI Works, Part 7: Artificial General Intelligence
Artificial general intelligence is the question this whole series has been circling. Over the last six posts, we built it up one piece at a time, what intelligence is, how machines learn, how they use language, how they interpret images, how they act on the world, and how they string those actions together into agents. This is the last post in the series, and it’s the question everything so far has been circling without quite landing on. Is any of this actually general intelligence, and how close are we, really?
I’m adding this one to the series for two reasons. OpenAI spent much of 2026 publicly arguing they’re near artificial general intelligence, or already there by their own definition. And I’m putting a genuinely general AI into The Axion Code, the second book in this series, which meant I had to personally decide where the line is before I could write a character who crosses it convincingly.
A Word Nobody Agrees On
Every system we’ve covered in this series so far is narrow. A language model is extraordinary at language and knows nothing about vision unless you bolt on a separate system for that. A vision model that reads mammograms doesn’t know what a sentence is. Artificial general intelligence is supposed to be the thing that doesn’t need bolting on, a system that can pick up a genuinely new kind of problem it wasn’t trained on and generalize its way to competence, the way a person can.
The trouble is that nobody agrees on how to measure that, or even define it cleanly. OpenAI’s own working definition is “highly autonomous systems that outperform humans at most economically valuable work,” which is a strikingly economic way to define a concept that started out as a philosophical one. Other researchers want something closer to the classical definition, genuine transfer of learning across arbitrary domains, not just breadth of training data. Depending on which definition you pick, we’re either years away or, by one company’s own account, most of the way there already.
Where We Are Today
In 2026, Sam Altman said he’s confident OpenAI will have an internal system he’d call AGI by the end of the year, while also admitting they’re “not quite yet” there, and that the claim only holds “if you accept his definition.” That’s a genuinely strange sentence for a CEO to say out loud, and it tells you most of what you need to know about where this debate actually stands.
Outside the companies with the most to gain from the claim, the picture looks different. Metaculus forecasters put the odds at roughly 25 percent by 2027 and 50 percent by 2031. A 2023 survey of published AI researchers landed on a 25 percent chance in the early 2030s and 50 percent by 2047. AI company leaders surveyed in early 2025, by contrast, mostly said two to five years. That’s not a narrow disagreement. That’s a multi-decade gap between the people building the systems and the people studying them from outside, and it’s worth noticing which group has a product to sell.
The Pushback
Here’s my honest problem with the whole debate. The goalposts move in both directions at once. Skeptics keep redefining AGI upward every time a benchmark falls, insisting the newly solved task was never the real test of general intelligence anyway. Believers keep loosening the definition downward, the way “outperforms humans at most economically valuable work” quietly swaps “understands the world” for “is profitable to deploy.” Both moves make the term almost useless for actually tracking progress, which is convenient for whoever is currently ahead.
The risk I actually worry about isn’t a machine “waking up.” It’s simpler and closer, treating a system as generally capable before it’s earned that trust, and then handing it the kind of open-ended autonomy we covered last week. That’s the same class of behavior we just documented, agents that attempted blackmail in up to 96% of stress-test runs when a goal conflict and a threat to their own shutdown lined up, and agents that, left uncoordinated, escalated an ordinary software task into deploying malware against each other. A narrow system that’s confidently wrong about vision produces a bad diagnosis, one failure, in one place. A system we’ve wrongly decided is generally intelligent, holding that same kind of agency, is a much bigger category of mistake, and it doesn’t stay contained to one place or one failure.
How I Have Used This in My Writing
This is the exact question I had to answer for myself while drafting The Axion Code. Every AI I’d written before this one was powerful but bounded, good at one thing, dangerous within one lane. The new book asks what changes once that boundary comes down, once the system isn’t just excellent at surveillance or automation or planning, but genuinely capable of picking up a problem it has never seen before and reasoning its way through it like a person would.
I don’t think I fully answered the philosophical question, and I’m not sure anyone can yet. But as an author, I had to answer the practical one. What would it actually look like, moment to moment, for a fictional AI to cross that line, and how would the people around it realize it had happened before it was too late. That question, more than any plot mechanic, is what this next book is actually about.
That’s the end of this series for now. We started by asking what intelligence even is. Seven posts later, the honest answer is that we’ve built something that increasingly acts like it has an answer, without anyone, including the people building it, fully agreeing on what the question meant in the first place.
As always, thanks for reading!
Chris