Read it once. Get back something you can actually use.

In technical terms

The core is a scaling-law extrapolation wedded to an intelligence-explosion thesis. Progress is counted in "orders of magnitude" (OOMs) across three compounding inputs: raw compute, algorithmic efficiency, and "unhobbling" (tools, agency, scaffolding). It projects another GPT-2-to-GPT-4-sized jump by ~2027.

• Effective compute: OOMs from both spend and efficiency gains, not hardware alone.
• Unhobbling: turning a base model into a planning, tool-using agent, argued to be worth multiple OOMs of apparent capability.
• Recursive acceleration: automated AI research compresses a human decade into roughly a year once models do the research themselves.

Formal summary

Thesis: trend-line extrapolation of compute, algorithms and unhobbling implies AGI is plausible by ~2027 and a fast superintelligence transition shortly after, which makes this a national-security problem before a product one.

Key points

• Progress is measured in compounding OOMs, not single breakthroughs.
• The decisive phase is an automated intelligence explosion, not the first AGI itself.
• Security of model weights and a state-led effort are framed as the real bottlenecks.

Remember: the argument is an extrapolation, not a proof; its force is in the trend and the lack of attention paid to it.

For animation

Character: Leo, a lone researcher at a desk in a dim San Francisco apartment.

• Visual: Leo draws a small dot on a graph; it climbs. Narration: "It started where everyone could see it."
• Visual: the line crosses a row of human silhouettes, schoolkid then student then expert. Narration: "Each year, one more level cleared."
• Visual: the line keeps going past the last silhouette into the dark. Narration: "The line doesn't stop where we do."
• Visual: the room fills with mirrored Leos, each drawing the next curve faster. Narration: "And then it starts drawing itself."
• Visual: pull back. The apartment is one lit window in a vast skyline. Narration: "The only question left is who's watching."

For a graphic

Through-line: a single rising staircase read left-to-right, where each step is an order of magnitude and the climb never plateaus.

The eye starts bottom-left at "GPT-2 · 2019", climbs steps labelled with the three inputs (compute, algorithms, unhobbling), passes a dotted "human expert" line near top-centre, and continues off the top-right into a glow marked "automated research." A thin second line peels off at the top and loops back down, the recursive acceleration, landing the whole argument as one connected image rather than a bullet list.