hawkes.aiKey ideas
- The one idea — Vector arithmetic captures relationships — the analogy equation actually computes, honestly caveated as approximate and at its best on famous examples.
- How it is shown — king − man + woman resolving to queen; Paris − France + Japan → Tokyo; the misses shown too.
- The trap to avoid — Both lazy verdicts — "just statistics" and "it understands" — miss what's really there: learned geometry precise enough for algebra, fuzzy enough to surprise.
- What it sets up — Layers push these points around — that's where thinking lives" (Act 02) · "what does 'near' even mean?
Take the numbers for king, subtract man, add woman, and the nearest word to where you land is queen. Nobody programmed that. It fell out of the map.
The one idea
Vector arithmetic captures relationships — the analogy equation actually computes, honestly caveated as approximate and at its best on famous examples.
You can literally do math on words. Watch. Take the coordinates for "king." Subtract the coordinates for "man." Add the coordinates for "woman." You land on a new location — a point where no token lives. Now ask the map: what's the nearest word to here? The nearest word… is "queen." Nobody programmed that. No rule anywhere says royalty minus male plus female equals queen. It fell out of the map — out of billions of nudges that placed words where their company demanded. Why does it work? Because "king minus man" isn't noise — it's a direction. Roughly: royalty, with maleness stripped out. And the same arrow that carries man to woman carries king to queen. The relationship is the geometry.
How it works — the demo
king − man + woman resolving to queen; Paris − France + Japan → Tokyo; the misses shown too.
The same trick, everywhere. Paris minus France plus Japan: Tokyo. Walking minus walk plus swim: swimming. Capital-of is an arrow. Tense is an arrow. Plural is an arrow. Now the honest part. It's approximate. The famous examples are the cherry-picked best; plenty of analogies miss or land merely-near. And in modern models the point moves with context — "bank" by the river sits somewhere else than "bank" holding your money. The clean arithmetic lives at the map's foundations, not its surface. The trap — both lazy verdicts.
The trap to avoid
Both lazy verdicts — "just statistics" and "it understands" — miss what's really there: learned geometry precise enough for algebra, fuzzy enough to surprise.
Why it matters — and what’s next
Layers push these points around — that's where thinking lives" (Act 02) · "what does 'near' even mean?
"It's just statistics" waves away structure precise enough to do algebra on. "It understands" inflates geometry into a mind. What's actually there is stranger than either: a learned geometry of meaning. Precise enough for arithmetic. Fuzzy enough to keep surprising the people who built it. Hold the picture: every word a point, every relation an arrow. Soon, the machine act — where layers grab these points and push them around each other. That's where thinking lives. But one honest question first: with four thousand coordinates… what does "near" even mean? Measuring distance in meaning-space is its own three minutes. Next.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
You can literally do math on words. Watch.
Take the coordinates for "king." Subtract the coordinates for "man." Add the coordinates for "woman." You land on a new location — a point where no token lives. Now ask the map: what's the nearest word to here?
The nearest word… is "queen." Nobody programmed that. No rule anywhere says royalty minus male plus female equals queen. It fell out of the map — out of billions of nudges that placed words where their company demanded.
Why does it work? Because "king minus man" isn't noise — it's a direction. Roughly: royalty, with maleness stripped out. And the same arrow that carries man to woman carries king to queen. The relationship is the geometry.
The same trick, everywhere. Paris minus France plus Japan: Tokyo. Walking minus walk plus swim: swimming. Capital-of is an arrow. Tense is an arrow. Plural is an arrow.
Now the honest part. It's approximate. The famous examples are the cherry-picked best; plenty of analogies miss or land merely-near. And in modern models the point moves with context — "bank" by the river sits somewhere else than "bank" holding your money. The clean arithmetic lives at the map's foundations, not its surface.
The trap — both lazy verdicts. "It's just statistics" waves away structure precise enough to do algebra on. "It understands" inflates geometry into a mind. What's actually there is stranger than either: a learned geometry of meaning. Precise enough for arithmetic. Fuzzy enough to keep surprising the people who built it.
Hold the picture: every word a point, every relation an arrow. Soon, the machine act — where layers grab these points and push them around each other. That's where thinking lives. But one honest question first: with four thousand coordinates… what does "near" even mean? Measuring distance in meaning-space is its own three minutes. Next.