I've been in AI for a few months, and one thing hasn't left my head since: AGI doesn't exist yet. It's one of two parts of AI — the other being superintelligence — that's still just hypothetical. That's what pulled me in. It's why I decided I wanted to help build it into what it should actually be. Once it's real, I don't think it becomes its own separate industry. Like electricity or the internet, it gets folded into every sector instead, changing each one over time instead of all at once.
When I started, I figured I'd spend most of my time on narrow, task-specific systems — models that translate language, recommend products, catch fraud. What I didn't expect was how fast I'd get stuck on the gap between what those systems can do and what a human brain does without even trying. A kid sees an unfamiliar animal once and remembers it for life. A chess novice watches a few games and starts predicting outcomes with real skill. The systems I was building needed thousands, sometimes millions, of examples to get anywhere close. That gap is what pulled me from just studying AI toward wanting to actually push it somewhere more general. It's not a side interest anymore — it's the direction I want my work to go.
How to actually get there, I still haven't figured out, and thinking it through has taught me more than anything else so far. My instinct is that AGI gets built by studying the brain more closely — its architecture, its efficiency, how it learns from so little. That's not just a gut feeling. It comes from a real asymmetry nobody's explained away: the brain solves problems most AI still can't, using a fraction of the data and a fraction of the power, and no one's shown that's just a side effect of scale instead of something about the architecture itself. I think this has to be done on purpose — without a real guiding model, AGI could stay out of reach for decades, maybe longer.
But that's not the only view people take seriously, and it wouldn't be honest to pretend it is. A lot of the field believes the opposite — that the biggest breakthroughs of the last decade, the transformers behind today's LLMs, worked specifically because they moved away from biological realism, not toward it. Backpropagation isn't thought to resemble anything real neurons actually do. In that view, progress has basically come from scale — more data, more compute, more parameters, stacked onto pretty simple designs. Some people sit in the middle — neuro-symbolic systems mixing learned pattern recognition with structured, rule-based reasoning. And some are building hardware and models straight from neuroscience — neuromorphic chips, spiking neural networks, predictive-coding architectures — basically trying to prove my instinct right through engineering instead of arguing about it.
The strongest counterargument is that biological plausibility and performance might just have nothing to do with each other — evolution optimized the brain for energy and survival, under constraints silicon doesn't have, so there's no real reason something built totally differently couldn't beat it anyway. I take that seriously. I can't fully rule it out. But that only shows brain-inspired approaches aren't necessary — not that they're the wrong bet. Given how much of the gap left between narrow AI and general intelligence still looks like a sample-efficiency and energy-efficiency problem, one system already solving both at once feels like something worth digging into, not a coincidence to shrug off. I don't think anyone actually knows which path wins, and at this point I've stopped expecting a clean answer. But out of all of them, brain-inspired is the one I keep coming back to — it's the only path that treats the proof already sitting in every human skull as something actually worth reverse-engineering.
What gets me most is the brain's pattern recognition specifically. It spots an animal at a glance, reads someone's expression in a fraction of a second, predicts how a situation's about to unfold almost instantly — all from barely any input, running on about as much power as a dim light bulb. It's honestly disappointing that today's AI still can't match that kind of efficiency, even when it beats humans on narrow benchmarks. Building something that actually outperforms the brain in general — actually making AGI real — means figuring out how it does this with so little data and so little energy, then building that into how we design AI, instead of just scaling past the problem.
To make that real: picture an AGI system in medicine that reads a patient's history the way a doctor's eye catches a face in a crowd — noticing a pattern spread across years of routine visits and catching a disease before it fully shows up, not because it trained on that exact case, but because it generalized the way a brain does. Or a tutor that reads a student's confusion the way a person reads a room, adjusting on the fly instead of running a fixed script, because it actually gets where the gap is instead of just pattern-matching. Neither of those is happening next year. But they're specific enough that I can hold myself to them, instead of just saying "AI will change everything" and leaving it at that.
Right now, AGI is still just a hypothetical word — defined more by what it needs to do than by anything that actually exists. That gap, between picturing it and actually building it, is what disappoints me most. It's also what drives me. I don't think I'll settle the fight between the scaling camp and the brain-inspired camp on my own, and I doubt anyone settles it cleanly before AGI actually shows up. But I'd rather spend my time inside that uncertainty, working toward an answer, than watch it from the sidelines. I want to help close the gap between the hypothetical and the real — and see AGI go from a term in a paper to something that actually makes everyday life better for the people living it.