The weird thing about the AGI conversation for the last couple of years is that the people most confident about when it’s arriving are way more confident than the people who actually study how thinking works. The can keeps getting kicked down the road and it’s worth asking why.

Dario Amodei at Anthropic has said flat out he expects AI “broadly better than all humans at almost all things” by 2026 or 2027. It’s an actual line in a policy filing to the U.S. government back in March 2025. Elon Musk keeps moving his date too, from “smarter than any single human by end of 2025” to now saying it could land by the end of 2026.

I couldn’t help but wonder… if our AI still hallucinates, still loses the thread of a conversation, still can’t remember what we told it five minutes ago, was AGI ever really around the corner, or were we just around the corner from disappointment?

Meanwhile, ask a neuroscientist “how does the brain actually produce a reasoned thought” or “how does reasoning work” and it all gets a lot more vague. We know where in the brain stuff happens; the prefrontal cortex is responsible for planning, the hippocampus is for memory, etc., but knowing the neighborhood isn’t the same as knowing the mechanism. Today, nobody can trace an actual thought, like solving a new logic puzzle, down to “these neurons fired in this order and that’s what a thought is.” It’s just not a thing right now. How we reason and come up with what we do is just not understood. It’s kind of like gravity that way, we know it exists, we can observe it and even predict it, but nobody’s actually explained why it’s there.

Then there’s consciousness, which is another input into reasoning. The two are a bit tangled. There are basically two camps here, and MIT’s Open Encyclopedia of Cognitive Science defines them well. On one hand, intelligence is just about running the right computations regardless of who or what is running them. In other words, it doesn’t matter whether it’s a computer or a human as long as it’s coded correctly. That’s “computational functionalism” – Wikipedia’s computational theory of mind page explains this further. The other side says consciousness requires biology to arise at all, and without it, you don’t get real reasoning either. NOEMA has an article that makes this case. These two sides don’t agree, and there’s no experiment sitting around that settles it. Even if AI consciousness turns out to be possible, we simply don’t have a way to test for it yet. Read the two linked pages, they’re chock-full of good stuff.

So why does anyone believe AGI is close, given all that? A simple argument is that we don’t need biology to do many things living things do. The most basic case is that we have planes that work just as well as birds, without flapping their wings.

On the other hand, AI isn’t really reasoning right now. Gary Marcus points out that current AI systems are still just pattern matching machines at the end of the day. They’re really good at predicting likely outputs based on statistics, which obviously isn’t reasoning, despite what your favourite LLM says it is doing. In fact, the models are just outputting intermediate steps that are then being taken into account when deciding the next token, nothing fancy. He’s made a version of this case more than once, basically arguing the “AGI is right around the corner” claims keep moving the goalposts.

Additionally, the people forecasting AGI without a financial stake in it are way less confident than the CEOs. Metaculus, a predictions platform, has it at roughly 25% by 2029 and 50% by 2033. That’s later and way softer than what most public figures are saying out loud. Even Demis Hassabis at DeepMind, who’s about as credible as it gets in this space, hedges to something like a coin flip by the end of the decade, and he specifically says creative reasoning and scientific discovery are way harder than the stuff AI’s been crushing lately, like coding and math.

Honestly, I think a decent chunk of this is just people using the word “AGI” to mean totally different things and it’s pretty subjective what their interpretations are. These interpretations include:

1. “automates most jobs that involve thinking” or
2. “beats humans on basically every test” or
3. “actually understands things the way a person does”

Those are three pretty different interpretations so of course people can look at the same AI system and land in totally different places.

All this philosophical back-and-forth is fun, but what does it actually look like day to day, right now? Here’s what is true today in software development: LLMs are helping people at least in the tech space solve problems faster. At the same time, in the tech space, they’ve also let people dive into projects they could only dream about before. Finally, it’s created a lot of code, which isn’t always good, so it’s created a much bigger need for processes such as code reviews and QA.

What we won’t be seeing any time soon is autonomous AI systems creating new and interesting systems purely on their own to try and solve complex problems that exist today. Without someone behind them guiding them, they’re still, at the end of the day, just calculating the next most likely token.