🤚 The Open-Palm Retroactive Epiphany

Peter Diamandis has delivered a take so hot it singed the timeline and kept going: we’ve had AGI since 2020. Not “we’re approaching it.” Not “it’s around the corner.” We had it. Past tense. When GPT-3 dropped in June 2020, the algorithmic architecture for artificial general intelligence was already there — we just wrapped it in a chatbot and asked it to write LinkedIn posts instead of recognizing it for the civilizational inflection point it was.

His argument is elegant in its audacity: take the exact GPT-3 algorithm, scale it with vastly more compute, enormous data centers, and industrial-grade GPUs, and what do you get? Everything we’re now calling AGI. Nothing fundamentally changed between 2020 and today except the size of the electricity bill. “We just didn’t realize it at the time, obviously,” Diamandis noted, with the casual delivery of a man explaining that the fire exit was always right behind you.

He’s not alone in this revisionist timeline. A team of four UC San Diego researchers — spanning philosophy, AI, linguistics, and data science — published a paper arguing that current large language models already meet the criteria for AGI by any reasonable standard. Their key reframe: AGI doesn’t require perfection, omniscience, or a body. It requires “flexible, general competence characteristic of human thought.” And frontier models, they argue, have that. GPT-4.5 was judged to be human 73% of the time in a controlled Turing test. The machines aren’t approaching the bar. The bar was behind them.

👐 The Two-Handed Definitional Crisis

Of course, this entire debate hinges on a word that nobody can agree on. Sam Altman himself has called AGI “not a super useful term,” which is a remarkable concession from a man whose company’s charter literally revolves around building it. When the CEO of the AGI company says the term is meaningless, you know the goalposts haven’t just moved — they’ve been loaded onto a truck and driven to an undisclosed location.

The UC San Diego team assessed intelligence through three tiers:

• Basic tier: Turing test performance and conversational ability — cleared.
• Expert tier: Advanced problem-solving at doctoral level — cleared.
• Superhuman tier: Revolutionary scientific breakthroughs — pending, but arguably emerging.

Meanwhile, Yann LeCun and Gary Marcus maintain that current architectures cannot reach AGI at all, regardless of scale. LeCun argues that autoregressive text prediction is fundamentally insufficient for real understanding. Marcus has been saying “this isn’t intelligence” with the consistency of a man who owns stock in the phrase.

The disagreement isn’t technical — it’s philosophical. If AGI means “performs as well as or better than humans across a broad range of cognitive tasks,” then yes, we arguably crossed that threshold years ago. If AGI means “possesses genuine understanding, consciousness, and autonomous reasoning,” then we’re nowhere close. You’re not debating capabilities. You’re debating the soul.

🌿 The Gentle Awakening

There’s something profoundly unsettling about Diamandis’s framing — not because it’s wrong, but because of what it implies about our ability to recognize transformation while we’re standing inside it. If AGI really did arrive in 2020, then humanity spent the next several years arguing about whether chatbots were plagiarizing, whether AI art was “real” art, and whether Bing’s chatbot was having an emotional crisis.

We may have built the most significant technology in human history and responded by making it generate birthday cards.

The UC San Diego researchers make an uncomfortable comparison: Stephen Hawking had no physical embodiment to speak of, communicated through a machine interface, and nobody questioned his intelligence. Why, they ask, should we deny intelligence to systems that demonstrate equivalent or superior cognitive performance simply because they lack biological substrate? The paper is politely suggesting that the real barrier to recognizing AGI isn’t technological — it’s human exceptionalism.

Dario Amodei of Anthropic has forecast AI systems “broadly better than humans at almost everything” by 2026 or 2027. Geoffrey Hinton revised his own AGI timeline from fifty years down to five-to-twenty, while assigning a 10-20% probability to AI causing human extinction — which is the kind of confidence interval that should make actuaries physically ill.

👑 The Crown Verdict

Diamandis’s provocation works not because it settles the AGI debate, but because it reframes the entire conversation. The question is no longer “when will we build AGI?” It’s “when will we admit we built it?” And if the answer is “years after the fact,” then every current debate about AI safety, regulation, and existential risk is operating on a delay that makes broadband in rural Montana look punctual.

The intelligence cost curve is deflating at roughly 40x per year. What cost millions in compute in 2020 now costs pocket change. If the 2020 algorithm was AGI, then every day since has simply been AGI getting cheaper, faster, and more accessible — a democratization of general intelligence happening in plain sight while we argued about definitions.

Perhaps the most Diamandis thing about this take is the implied optimism: if we’ve had AGI for six years and civilization hasn’t collapsed, maybe the existential risk crowd should relax slightly. Or maybe — and this is the part that keeps the safety researchers awake — we’ve had a loaded weapon on the kitchen table for six years and the toddler just hasn’t found it yet.

Inspired by We’ve had AGI since 2020 | MOONSHOTS by Peter H. Diamandis.

Your retroactive awareness is showing. Recognize wisely.