Artificial General Intelligence — a machine that can think across any domain the way a person can — used to be a question of “if.” Sometime around 2023 the conversation flipped to “when,” and the loudest voices started measuring the answer in months. Three years later we have a cleaner way to check that optimism against evidence. Spoiler: the gap between the hype and the scoreboard is enormous.
The 2026 reality check nobody puts on a keynote slide
The single most useful number in the whole AGI debate right now comes from the ARC-AGI benchmark, created by researcher François Chollet. Unlike most AI tests, it doesn’t reward memorization — it measures skill-acquisition efficiency, i.e. how well a system learns a brand-new task it was never trained on. That is much closer to what we actually mean by “general” intelligence.
On the 2026 version of the test, ARC-AGI V3, the frontier models — the same ones marketed as nearly-AGI — scored about 0.3%, and they burned roughly $5,000 to $9,000 of compute per task to get there. Humans score 100% on the same puzzles, usually in a few seconds, for the price of a cup of coffee and some patience. That is not a rounding error. That is a different category of thing.
Defining the elusive AGI: what does “general” really mean?
Part of why AGI feels perpetually “almost here” is that nobody agrees on the finish line. One popular industry definition is “a system that can automate the majority of economically valuable work.” It’s convenient — and it’s also a goalpost shaped like a business plan rather than a description of intelligence. The ARC Prize team prefers a tighter framing: a system that matches the learning efficiency of a human. Move the definition and you move the headline, which is exactly why timelines are so slippery.
The timeline debate: optimism vs. skepticism
The optimists point at a real, dizzying curve: models that went from barely writing a paragraph to passing bar exams and shipping working code. The skeptics point at the same curve and ask a sharper question — better at what? The honest answer is that progress has been spectacular on tasks you can grade automatically (code, math, multiple-choice) and far shakier on the open-ended, never-seen-before problems that define general intelligence. We didn’t necessarily build smarter machines; we built much better-trained ones. (If you enjoy watching the hype meet the hard edges, our AI apocalypse tier list sorts the panic from the plausible.)
The limits of language models: why bigger isn’t necessarily smarter
Chollet’s core thesis is uncomfortable for the “just add more GPUs” crowd: large language models didn’t get smarter, they got better-trained on verifiable domains. Scale up the data and the parameters, and the model gets sharper at things that look like its training. Hand it a genuinely novel, non-verifiable problem and the apparent intelligence quietly evaporates. Researchers have a name for this lopsided profile — “jagged intelligence”: superhuman in narrow valleys, startlingly clumsy on the ridges in between. It’s why the same model can draft a legal argument and then fail a puzzle a seven-year-old solves on the first try. (Even the people running these labs trip over the definition — see Jensen Huang’s own AGI argument quietly proving the opposite of his point.)
Beyond language: the promising paths that aren’t just “more scale”
If scaling alone won’t get us there — and ARC-AGI strongly suggests it won’t — the interesting research is happening elsewhere: program synthesis and “test-time” learning (systems that reason out a new procedure on the spot instead of recalling one), neuro-symbolic hybrids that bolt logical structure onto neural pattern-matching, and agents that build and test their own world-models. None of these are finished. All of them are more honest about the actual bottleneck: generalization, not memorization.
A gradual ascent, not a sudden leap
The cinematic version of AGI is a single midnight where the machine “wakes up.” The likelier version is boring and slow: a long climb where capabilities arrive unevenly, each one useful, none of them the whole thing. Treating every model release as the last step before the summit is how you end up disappointed on a two-year cycle. The benchmark scores are the altimeter, and right now the altimeter says we’re still in the foothills.
Why this matters even if you never build a model
The hype isn’t harmless. It moves trillions in investment, shapes policy, and quietly sets expectations for tools millions of people now use daily. Knowing that today’s AI is a brilliant, jagged, narrow technology — not a baby superintelligence — changes how you use it: as a fast, fallible assistant to double-check, not an oracle to obey. (Want it on your own machine, no hype required? Here’s how to run AI locally.)
The honest bottom line
AGI may well be coming. But “coming” is doing a lot of work in that sentence. The 0.3%-vs-100% gap on ARC-AGI is the clearest evidence we have that scaling, on its own, is not the road to general intelligence — and that the people quoting you a date are guessing. Awareness of the current limits, and curiosity about the new architectures trying to fix them, is what turns AGI from a marketing horizon into an actual research program.
Sources & further reading
- ARC Prize — What is ARC-AGI?
- ARC Prize — 2025–2026 results and analysis
🐾 Curiosity looks good on you. Explore the Goodies, or find our illustrated books on Amazon.
