Should the recipe be public?

Question 1

Coca-Cola keeps its formula in a literal vault in Atlanta. A century of mystique, billions in profit, and no law says they owe anyone the recipe.

Accepted Answer

Trade secrets are as old as trade. Nobody pickets the vault. Hold that instinct — we're about to raise the stakes on what's inside it. (85% of Americans answer "Fair play" — floor est.)

Question 2

A lab invents a life-saving drug. Society's deal: a 20-year monopoly to profit — but they must publish exactly how it's made, and when time's up, anyone may manufacture it as a generic.

Accepted Answer

That's the patent bargain: disclosure in exchange for monopoly. Secrecy forever or publication now aren't the only options — there's a clock. Keep the clock in your pocket. (72% of Americans answer "Good deal" — floor est.)

Question 3

In 2019, OpenAI built a text generator it declared too dangerous to fully release — fake news, impersonation, chaos. It came out nine months later anyway. Civilization held. Compared to that 'too dangerous' model, today's freely downloadable models are bigger by—

Accepted Answer

Roughly 1,000x. Every generation of AI has had a model someone called too dangerous to release, and so far every one of them is now a free download with a tutorial on YouTube. Either the danger was oversold — or the fuse is just long. (50% of Americans answer "More than 100x" — floor est.)

Question 4

A master locksmith writes a book showing exactly how every common lock is picked. Burglars can read it. So can every lock-maker, security guard, and homeowner in the world.

Accepted Answer

This fight already happened — in 1853. Locksmiths called publication 'aiding thieves'; the publishers won, and modern computer security runs on their logic: the flaw the good guys can't see is the one the bad guys are already using. (52% of Americans answer "Publish" — floor est.)

Question 5

January 2025: a Chinese lab gives away a model nearly as good as America's best, free for anyone to download, reportedly built for under $6 million. That Monday, America's biggest AI chipmaker's stock drops. How much value vanished in one day?

Accepted Answer

Almost $600 billion — the largest one-day loss of any company in stock market history. A free recipe did that. Whatever you believe about open weights, the market believes they're real. (55% of Americans answer "More" — floor est.)

Question 6

A new AI learns not by reading the internet, but by asking an older, smarter AI millions of questions and studying its answers — a technique called distillation. OpenAI, which trained on everyone's writing without asking, complained that DeepSeek may have done exactly this to them.

Accepted Answer

The art student from Market 01 is back — except this time she's the one being studied. In July 2026, fifty American AI companies signed a letter defending distillation as a legitimate way for machines to learn from machines. If reading everyone's work to build a mind is fair, it's surprisingly hard to explain why that mind's own answers are sacred. (35% of Americans answer "Theft" — floor est.)

Question 7

Researchers stripped the safety guardrails off a downloadable model in about five minutes, for fifty cents of computer time. Once the weights are public, every safeguard is a suggestion. The counterpoint: a RAND red-team found chatbots gave would-be bioweapon builders nothing they couldn't already find on the open internet.

Accepted Answer

This is the strongest card the vault side holds: open weights are forever, and un-releasing is impossible. 47% of voters oppose open-sourcing powerful models; 23% support it. The rest are where you might be right now — split between the locksmith and the burglar. (47% of Americans answer "Lock it down" — AIPI/YouGov, 2023)

Question 8

Powerful AI exists either way. The real question is custody. Option one: its full recipe is controlled by a handful of corporate boards you'll never sit on. Option two: it's downloadable by five billion people — including the very worst of them.

Accepted Answer

When Meta open-sourced Llama 3, opposers outnumbered supporters roughly 4 to 1 — but only 31% of Americans say they trust AI companies to develop AI responsibly either. America distrusts both custodians. It just distrusts them differently. (51% of Americans answer "The boards" — AIPI/YouGov, 2024)

Question 9

Plot twist: America's best models are mostly locked in vaults. China's best models are free downloads — and they're now the most-downloaded AI models on Earth, running in startups from Jakarta to São Paulo.

Accepted Answer

The best open models now trail the frontier by only ~3–4 months, and Chinese models account for nearly half of top open-model downloads. The vault buys a shorter head start every year — and the CEO of the most famous vault admitted in 2025 they'd been 'on the wrong side of history' and shipped an open model. (40% of Americans answer "Vault anyway" — floor est.)

Question 10

July 2026: Nvidia's CEO posts an open letter — his first post ever — asking Washington not to restrict open models. Within days, fifty-plus companies sign it, including sworn rivals OpenAI, Google, and Meta. Their boldest claim: openness is the SAFER path, because a locked-up model can fail in ways nobody outside the vault can see. The two biggest names missing: Amazon — and Anthropic, the lab loudest about AI safety.

Accepted Answer

Both readings are true at once. The letter's best card: open-source software already proved that many eyes find flaws faster than one vault does. The skeptic's best card: every signer profits if you believe them — starting with the chipmaker who wrote it, who sells the shovels either way. Google's AI chief signed while noting its free Gemma models have been downloaded 300 million times. Its crown-jewel Gemini models remain, of course, in the vault. (45% of Americans answer "They're right" — floor est.)

Question 11

Last question. Remember the drug-patent clock? Apply it to AI: labs may keep frontier models fully closed for, say, five years of profit and safety testing — then the weights must be published, like a patent expiring into generics.

Accepted Answer

You just designed AI policy with a 200-year-old tool. Almost nobody in this fight is purely vault or purely torrent — the real argument is over the length of the clock. Let's see where you closed. (58% of Americans answer "Take the deal" — floor est.)

Is keeping your best recipe secret fair play?