hawkes.aiKey ideas
- The one idea — Generation is sampling from a distribution — invention welcomed is "creativity," invention where facts were required is "hallucination." Same machine state.
- How it is shown — Two prompts, identical machine hum: a real tower (facts exist) vs an invented tower (facts don't) — the bars rank candidates either way.
- The trap to avoid — Demanding "just stop hallucinating" without changing decoding or grounding is demanding it stop generating.
- What it sets up — Systems can be built that force 'I don't know.
Everyone demands AI 'just stop hallucinating.' Here's the problem: there's no lying module to remove. The machinery that invents fake cases is the machinery that writes the poems.
The one idea
Generation is sampling from a distribution — invention welcomed is "creativity," invention where facts were required is "hallucination." Same machine state.
The poem that amazed you and the court case it invented came out of the same mechanism. Not similar mechanisms. The same one. Remember the machine: score every possible next chunk, roll weighted dice, commit, repeat. Every output — sublime or absurd — is a walk through those rolls. Point it at a poem, a brand name, a plot twist — and those mid-height bars are the whole value. A likely-but-not-certain path through language is novelty that still hangs together. Nobody checks a poem against a database. We call this creativity, and we're right to love it. Now point the same machine at a legal citation, a medical dose, a biography.
How it works — the demo
Two prompts, identical machine hum: a real tower (facts exist) vs an invented tower (facts don't) — the bars rank candidates either way.
The dice keep doing their one job: producing plausible continuations. A plausible-sounding case name. A plausible dosage. When the task needed truth and the roll produced likelihood, we call it hallucination. The machine's internal hum is identical. Here's the uncomfortable center: nowhere in that loop is there a step that asks "is this true?" Likely and true overlap often — that's why it's useful — but the machine only ever measured likely. Watch it feel no difference. "Tell me about the Eiffel Tower" — bars rise, words flow. "Tell me about the Glorbon Tower" — a thing I just invented — bars rise, words flow. Height, architect, opening year.
The trap to avoid
Demanding "just stop hallucinating" without changing decoding or grounding is demanding it stop generating.
Why it matters — and what’s next
Systems can be built that force 'I don't know.
It doesn't know it crossed a line, because from inside, there is no line. So the trap: telling it — or your vendor — to "just stop hallucinating." Unchanged, that demand means "stop generating." The real fixes change the machine's situation: ground it in retrieved documents, check its outputs, or — later in this series — gate it so it must earn the right to answer. There's even a dial — you'll meet it properly later as "temperature." Turn it up: bolder rolls, fresher phrasing, more invention. The same turn raises both the poetry and the fabrication. One dial, two names, depending on whether you wanted facts. So retire the question "why does it lie?" and load the better one: what could force truth to outrank likelihood? Machinery for exactly that exists — gates that make a model refuse rather than guess — and this series builds up to it, piece by piece. Same mechanism. Your job is choosing where its rolls are welcome.
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
The poem that amazed you and the court case it invented came out of the same mechanism. Not similar mechanisms. The same one.
Remember the machine: score every possible next chunk, roll weighted dice, commit, repeat. Every output — sublime or absurd — is a walk through those rolls.
Point it at a poem, a brand name, a plot twist — and those mid-height bars are the whole value. A likely-but-not-certain path through language is novelty that still hangs together. Nobody checks a poem against a database. We call this creativity, and we're right to love it.
Now point the same machine at a legal citation, a medical dose, a biography. The dice keep doing their one job: producing plausible continuations. A plausible-sounding case name. A plausible dosage. When the task needed truth and the roll produced likelihood, we call it hallucination. The machine's internal hum is identical.
Here's the uncomfortable center: nowhere in that loop is there a step that asks "is this true?" Likely and true overlap often — that's why it's useful — but the machine only ever measured likely.
Watch it feel no difference. "Tell me about the Eiffel Tower" — bars rise, words flow. "Tell me about the Glorbon Tower" — a thing I just invented — bars rise, words flow. Height, architect, opening year. It doesn't know it crossed a line, because from inside, there is no line.
So the trap: telling it — or your vendor — to "just stop hallucinating." Unchanged, that demand means "stop generating." The real fixes change the machine's situation: ground it in retrieved documents, check its outputs, or — later in this series — gate it so it must earn the right to answer.
There's even a dial — you'll meet it properly later as "temperature." Turn it up: bolder rolls, fresher phrasing, more invention. The same turn raises both the poetry and the fabrication. One dial, two names, depending on whether you wanted facts.
So retire the question "why does it lie?" and load the better one: what could force truth to outrank likelihood? Machinery for exactly that exists — gates that make a model refuse rather than guess — and this series builds up to it, piece by piece. Same mechanism. Your job is choosing where its rolls are welcome.