hawkes.aiKey ideas
- The one idea — The model samples from the distribution rather than always taking the top — because always-greedy text is measurably repetitive and dead.
- How it is shown — The weighted wheel spinning; greedy's degeneration loop shown against sampled liveliness.
- The trap to avoid — "The model chose that word" — the sampler chose; the model provided the ballot. Blame (and control) belong to the right machine.
- What it sets up — The roll's dials.
Force an AI to always pick its most likely word, and the text curdles — bland, then circular, phrases chasing their own tails. The randomness is the cure.
The one idea
The model samples from the distribution rather than always taking the top — because always-greedy text is measurably repetitive and dead.
Here's the moment this series opened with, finally built from parts: the machine does not pick its best word. It rolls weighted dice across all of them. That choice is stranger and smarter than it looks. Because the obvious design is sitting right there: always take the favorite. Greedy decoding: deterministic, simple, sounds like rigor. Why would you ever inject randomness into a precision machine? The answer was discovered empirically, and it's one of the field's great surprises. Run greedy for a paragraph and watch it curdle. The text turns bland, then circular — phrases repeating, sentences chasing their own tails. It's a measured, replicated result: always taking the top word drives generation into repetition loops, because the most probable continuation of slightly-repetitive text is more repetition. The favorite, chosen every time, is a trap that feeds itself.
How it works — the demo
The weighted wheel spinning; greedy's degeneration loop shown against sampled liveliness.
Now restore the wheel. Mostly it lands on the favorite — the weights see to that. But sometimes a rival wins, and that small surprise knocks the text out of its rut into fresh phrasing, new angles, actual prose. The randomness in your AI's replies isn't sloppiness — it's a deliberate cure for a measured disease. The dice are load-bearing. Keep the org chart straight, because it's about to matter commercially. The tower forms the opinion — a hundred floors of it. The head and softmax price that opinion into a market. And the sampler — this little wheel, maybe five lines of code — makes the actual choice. The most consequential word-picker in the machine is also its very smallest part. So the trap: "the model chose that weird word." Usually — the sampler chose it.
The trap to avoid
"The model chose that word" — the sampler chose; the model provided the ballot. Blame (and control) belong to the right machine.
Why it matters — and what’s next
The roll's dials.
The model provided a ballot; the wheel drew from it, and sometimes draws a legitimate long-shot. Teams burn weeks blaming models for behavior that a sampler setting caused — and ship fixes by turning a dial, not by switching models. Diagnose the right machine. And here's what makes this practical: the wheel has dials. One famous knob that reshapes the whole market before the spin. Three gate-levers that decide which wedges are even allowed on the wheel. Every settings panel in every AI product you've used is these four controls wearing product names. Next episode: the famous one. One slider that turns the machine from robot to poet — and the mechanics behind it are two lines of math. Temperature.
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 2:45 of narration
Here's the moment this series opened with, finally built from parts: the machine does not pick its best word. It rolls weighted dice across all of them. That choice is stranger and smarter than it looks.
Because the obvious design is sitting right there: always take the favorite. Greedy decoding: deterministic, simple, sounds like rigor. Why would you ever inject randomness into a precision machine? The answer was discovered empirically, and it's one of the field's great surprises.
Run greedy for a paragraph and watch it curdle. The text turns bland, then circular — phrases repeating, sentences chasing their own tails. It's a measured, replicated result: always taking the top word drives generation into repetition loops, because the most probable continuation of slightly-repetitive text is more repetition. The favorite, chosen every time, is a trap that feeds itself.
Now restore the wheel. Mostly it lands on the favorite — the weights see to that. But sometimes a rival wins, and that small surprise knocks the text out of its rut into fresh phrasing, new angles, actual prose. The randomness in your AI's replies isn't sloppiness — it's a deliberate cure for a measured disease. The dice are load-bearing.
Keep the org chart straight, because it's about to matter commercially. The tower forms the opinion — a hundred floors of it. The head and softmax price that opinion into a market. And the sampler — this little wheel, maybe five lines of code — makes the actual choice. The most consequential word-picker in the machine is also its very smallest part.
So the trap: "the model chose that weird word." Usually — the sampler chose it. The model provided a ballot; the wheel drew from it, and sometimes draws a legitimate long-shot. Teams burn weeks blaming models for behavior that a sampler setting caused — and ship fixes by turning a dial, not by switching models. Diagnose the right machine.
And here's what makes this practical: the wheel has dials. One famous knob that reshapes the whole market before the spin. Three gate-levers that decide which wedges are even allowed on the wheel. Every settings panel in every AI product you've used is these four controls wearing product names.
Next episode: the famous one. One slider that turns the machine from robot to poet — and the mechanics behind it are two lines of math. Temperature.