Understanding AI · Updated June 2026

AGI, explained honestly

Everyone either hypes it or dismisses it. Here are both sides, taken seriously, with the real data — and the one result that cuts through the noise: in March 2026, every frontier model scored under 1% on ARC-AGI-3 while untrained humans scored 100%.

The honest answer: AGI has not been achieved by any rigorous measure of novel reasoning, but the goalposts have moved so far that "AGI is here" and "AGI is decades away" can both be said truthfully — because they mean different things. The disagreement is mostly about definitions and incentives.

The "it's here / coming soon" camp

The "impossible / decades away" camp

Name the conflicts of interest

This doesn't make anyone wrong — but it makes predictions less than neutral. Huang sells the chips. Altman raises the billions. Musk competes through xAI. The most confident claims tend to come from those with the most to gain from belief.

The data point that cuts through: ARC-AGI-3

On 24 March 2026, every frontier model — GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6, Grok-4.20 — scored below 1% on ARC-AGI-3. Untrained humans scored 100%. GPT-5.4 scored 0.26% at a cost of $5,000–$9,000 per task.

These are the same models that ace bar exams, score 70%+ on software-engineering benchmarks and write production code. Hand them a simple interactive puzzle with no instructions — the kind a 10-year-old masters — and they fail almost completely. Full detail on the ARC-AGI benchmark.

The definition problem nobody names

The models of March 2026 are incrementally better than those of 2024, not qualitatively different. If AGI was "5 years away" in 2024, how is it "already here" in 2026? Because it isn't the capabilities that changed drastically — it's the definition. Altman described achieving AGI in December 2025 as "spiritual rather than literal." An elegant way to say the goalposts moved.

What this means for your business

Ignore the AGI headlines for procurement. What matters is what today's models reliably do and reliably don't — which is stable, measurable and unaffected by the definitional war. That's the argument of our opinion piece, if it quacks, it's a duck, and the practical map in what AI can't do.

Curious whether scaling can ever get there? See can AI become intelligent, and the quick definitional guide AI vs AGI.