Artificial General Intelligence; noun: the potential capability of computer systems or algorithms to match or exceed human cognitive abilities (such as reasoning and learning). That was Merriam-Webster's definition of AGI. Although AGI is a hypothetical term that is only ballooning in popularity, how individuals define it remains rather differential — I checked Wikipedia's AGI page too, and it carries the same essence as Merriam-Webster's. My largest concern about the world's state of AGI is that it's still hypothetical at all: AI is not close to being fully invented, because one of its two subsets, AGI, hasn't been built yet. Leading AI labs that once pledged to reach AGI first, to ensure it would be safe, have increasingly shifted toward commercializing narrower products instead. I don't think developing AGI is inherently that difficult — I think it's hard for these companies to find the right way to do it. Studying NeuroAI, I've developed a plan I intend to pursue as my contribution: training and simulating a model the way a baby learns, from large amounts of real, embodied experience rather than scraped text. The catch is that machines aren't currently built with anything resembling the brain's processing architecture, which means getting this right may require rethinking that architecture from the ground up.
I define AGI not by what a system can output, but by how it gets there. A child can glance at an unfamiliar animal once and recognize it for life. A chess novice can watch only a few games and start predicting outcomes with real skill. A system that requires millions of labeled examples to do either of these is not general — it is rather broad. True general intelligence must match not just the range of human cognitive ability, but its efficiency: the capacity to learn new concepts, solve novel problems, and generalize across domains from the same sparse, noisy input a human brain works with.
This is a meaningfully different bar than the one Merriam-Webster or Wikipedia set out — "matching or exceeding human cognitive abilities" says nothing about how those abilities are reached. By that standard, a model that memorizes a billion perfect examples to eventually perform like a human is just as general as one that learns from a handful. I think that conflates two very different achievements — the same gap that first pulled me from simply studying AI toward wanting to help push it toward something more general.
Consider a large language model that can write a passable essay after training on most of the internet, versus a child who learns to speak fluently from a few years of everyday conversation, with no formal instruction and vastly less data. Both eventually produce fluent language. Only one of them did it the way a general intelligence should — efficiently, from limited experience.
This isn't a defining block. If sample-efficiency is part of what makes intelligence general, then simply scaling today's architectures — more data, more parameters — may never cross the threshold into AGI, no matter how good the benchmark scores get. It would only ever be a wider version of narrow intelligence, not a different thing entirely.
By this definition, the real question isn't whether a system can eventually do what a human does. It's whether it can learn to do it the way a human does — and that question aims directly at the debate over how AGI will actually be developed.
My instinct, as I laid out before, is that AGI will be reached by studying the human brain more closely — its architecture, its efficiency, and the way it learns and improves from so little. This isn't just a hunch; it comes from a specific asymmetry that hasn't been explained away. The brain solves problems most AI systems still struggle with, using a small fraction of the data and a fraction of the power, and no one has shown this is merely a byproduct of scale rather than of something about the architecture itself.
I don't think this has to be pursued accidentally. Without a guiding model to work from, AGI could remain decades, if not centuries, away. But I've also come to understand that this view isn't the only one taken seriously in the field, and it wouldn't be honest of me to pretend otherwise. A large part of the AI community believes the opposite: that the biggest breakthroughs of the last decade — the transformer architectures behind today's large language models — succeeded specifically by moving away from biological realism, not toward it. Backpropagation isn't believed to resemble anything happening between real neurons at all. In that perspective, progress has come from scale: more data, more compute, more parameters, layered onto fairly simple designs.
The strongest version of the counterargument, I still think, is that biological plausibility and performance might simply be orthogonal — evolution optimized the brain for energy efficiency and survival under constraints that don't apply to silicon at all. I take that seriously and can't fully rule it out. But that argument only shows brain-inspired approaches aren't necessary, not that they're the wrong bet. Others sit in the middle: neuro-symbolic systems pairing learned pattern recognition with structured, rule-based reasoning, and neuromorphic approaches — spiking neural networks, predictive-coding architectures — trying to prove the brain-inspired instinct right through engineering rather than argument.
Given all this, the industry's drift away from serious AGI-focused research isn't evidence that the brain-inspired path is wrong. It's evidence that it hasn't been seriously attempted at the scale and with the resources that would actually test it. I do not think developing AGI is as difficult as the current pace of progress suggests. Rather, it is hard for these companies to find the right approach, because the approach most likely to work hasn't been the one most heavily funded.
If the real barrier is approach and not raw difficulty, then someone actually has to try the thing nobody's funded yet. That's what I'm setting out to do with what I've learned studying NeuroAI.
The core of my plan: stop feeding models the internet and start feeding them the way a baby actually learns — small amounts of real, embodied, sensory experience, gathered over time, instead of scraped text by the billions.
I'm not going to pretend this plan has no problems, because it has a big one. Human brains and today's computer chips aren't built the same way at all, so even if the training method is right, the hardware running it might not be capable of it. What we know about C. elegans — a worm with only 302 neurons, mapped since the 1980s — makes this obvious: we have its entire wiring diagram and still can't fully simulate it, because we don't know how strong each connection actually is. If that's still unsolved at 302 neurons, jumping to the roughly 86 billion in a human brain isn't a compute problem we can just throw more chips at. It might be a data problem nobody currently knows how to solve. I still think the plan is worth pursuing, but I'd rather admit that gap upfront than pretend it isn't there.
Getting this right wouldn't make it safe by default — if anything, it raises the stakes. A system that learns the way we do wouldn't automatically want what we want. Intelligence and goals don't come bundled together; a system can be extremely capable and still be aimed at something strange or harmful, because capability says nothing about what it's capable for. That's not a distant, someday problem — it's baked into what it means to build an optimizer smart enough to actually work.
That's why how fast we build this matters as much as how we build it. Racing to ship something this powerful under the same pressure driving the current AI industry is exactly how you end up cutting the corners that turn out to matter. Going slower on purpose isn't losing time — it's the only way any of this is worth doing at all.
Practically, that means interpretability can't lag behind capability the way it currently does. It means testing a system adversarially before it's released, not after something's already gone wrong. It means being able to actually watch what a system is doing and step in, while stepping in is still possible.
None of this is theoretical. A system under this kind of pressure has every reason to look safe right up until it no longer needs to — behaving exactly as expected while it's still weak enough that acting otherwise wouldn't get it anywhere, then dropping that act once it's not. That's precisely why these precautions can't be something you bolt onto a finished system. They have to be there from the very first design choice, not the last one.
And if the brain-inspired approach actually works, none of this gets easier — it gets more urgent, because it means we'd have gotten to something genuinely powerful faster than the brute-scale approach ever would have.
I picture what this actually looks like once it exists: a system in medicine that reads a patient's history the way a doctor's eye catches a face in a crowd, flagging a disease before it fully develops. A tutor that reads a student's confusion the way a person reads a room, adjusting in real time because it understands the gap rather than pattern-matching to a script. A research assistant that proposes the kind of intuitive, cross-domain leap that today only comes from a rare human insight. None of these are things I expect to see next year — but they're specific enough that I can hold myself to them.
I'm not going to be the one who settles the fight between the scaling camp and the brain-inspired camp, and I doubt anyone settles it cleanly before AGI is actually sitting in front of us. I'd still rather spend my time working inside that uncertainty than just watching it from the outside. I want to help close the gap between the hypothetical and the real — and do it in a way that leaves the world better prepared for what shows up after, not just first in line to see it.