True randomness is critical. Why?

The one-paragraph version

On October 17 a hacker took 2.94 trillion QU out of 1,404 Qubic wallets in one hour and forty-four minutes. The chain kept running the whole time. What broke wasn't the protocol, it was the wallet sitting next to it: when making a seed it asked the browser for a random number, and on Apple devices the browser handed over a coin with only four billion sides. A laptop counts to four billion before lunch. So the hacker counted. Here's the interesting part: CfB has been saying since March that randomness is a matter of life and death for AI.

Below is my attempt to understand why.

Part 1. Ok so what is randomness, actually

How do you think about randomness?

We think of it as noise. Something you sprinkle in so a game is fair and a lottery is a lottery. It's the other way around. Randomness is the only part of a system that cannot be predicted from the system itself. Every rule, every weight, every line of code can be read, copied and run forward by anyone who has it. A truly random bit is the one thing an observer can't derive. It's the system's only private thought.

One word matters here: truly. Because a computer can't do randomness. At all. A processor is a machine that does exactly what it was told, and does it exactly the same way the second time. When a program "takes a random number", what it really takes is a formula: a starting number, the seed, and it spins it through arithmetic to get a long, pretty, unpredictable-looking sequence. Looking. Know the seed and you know the whole sequence till the end of time. This is called pseudorandomness, and von Neumann said it best back in 1951: anyone trying to get random numbers out of arithmetic is "in a state of sin". He was joking. He was also right.

So where do you get the real thing? Only from the world, never from a formula. Thermal noise in a transistor. An atom decaying. The quantum "this way or that way" that physics forbids you to predict, in principle. All of these share one property: no observer, even with complete information about the system, can tell what comes next. Not "hard to compute". There is nothing to compute.

And here's the trap that makes all of this so easy to miss: you can't tell from the output. A sequence from a formula and a sequence from an atom look the same. They pass the same tests. Equally "random" to the eye. The only difference is whether somewhere there exists a number that lets you replay the whole thing from the start. For a formula, there always is. That's why a weak generator never gives itself away. A wallet runs on it for years, looks fine, nobody notices anything, until someone finds the seed.

Take real randomness out of a system and the system doesn't break. It becomes transparent to whoever holds the seed. It runs. It answers. It looks alive. It just has no secrets left, and not a single choice that is its own.

A wallet seed is a 55-letter phrase that is supposed to exist nowhere else in the universe. With a weak generator, that phrase sits in a short list that someone else can print out. The wallet "works". The signature "verifies". The key just belongs to two people. That's the whole October hack in one paragraph, and notice, it's not about Qubic. It's about any system that once asked somebody for a coin and never asked where the coin came from.

Part 2. A thought experiment. Try swapping the wallet for an AI

Today's AI, a large language model, is fully deterministic inside. Same function, same input, same answer. The only place where there is any randomness at all is the very last step, when one word gets picked from a list of likely ones. And it gets picked by an ordinary pseudorandom generator with a seed set by whoever owns the server. Ask a model for a random number from 1 to 10 and it'll say 7 way more often than it should. It isn't flipping a coin. It's remembering what a coin flip looks like. That's how it's built.

Now picture not a chat but an agent. One that trades, negotiates, guards a network, makes its own keys, breeds its own successors. Same question we asked the wallet: what if its randomness can be read or nudged? I count four things, each worse than the last:

1. It can't keep a secret. Every key, password, nonce, one-time code comes out of the same coin. Readable coin = nothing it protects is protected. Not "weakly protected", not protected at all, because all of cryptography rests on one assumption: the number you picked, nobody else could have picked. The wallets in October are this in miniature: 1,404 "unique" seeds that turned out to be a page from a short list. An agent with a generator like that will hand its keys to anyone who thinks to look, and never notice.

2. It can't play. Against another agent your only defense is being unpredictable: bluffing off-schedule, probing where no pattern says to probe. Game theory calls it a mixed strategy, and it only works if the other side can't see your dice. Poker with open cards isn't poker. A negotiation where your walk-away point is known isn't a negotiation. A trading bot whose "random" intervals can be computed is a bot that gets front-run on every trade. An agent with known dice doesn't lose sometimes. It loses every time, quietly, to whoever looked. And the nasty part: it won't feel it. Its stats will just look like bad luck.

3. It can be steered, and nobody will see it. This is the main one and the least obvious. If you control the source of randomness you don't need to touch the AI's code. The code is open, audited, everyone's read it, and that means nothing. Tilt the coin a little. Slightly more heads than tails. The AI keeps "freely" choosing exactly what you wanted, forever, and no audit will ever find it, because there's nothing in the code to find. Looks autonomous. Isn't. That's the difference between hacking and owning: a hack leaves traces, a tilted coin doesn't. And trusting the chip vendor, the cloud, or the operator is still trusting somebody. Real randomness is the kind where that somebody doesn't exist at all.

4. It can't evolve. Aigarth is selection: spawn variants, keep what works, repeat. That's how everything alive works, and it works because mutations are nobody's design. Nature doesn't choose who mutates. Tilt the coin and you no longer have a wild population searching for intelligence, you have a breeder picking winners. Sometimes it even looks better: faster, smoother, more predictable. But a breeder can only breed what he can imagine. Evolution breeds what nobody imagined. If the goal is to find a mind that never existed before, a breeder is the worst possible tool. Whatever grows under a tilted coin is the breeder's creature, not the network's.

Add up the four. An AI with riggable randomness has no secrets, no edge, no autonomy and no real evolution. It can be as smart as you like, changes nothing. A very fast puppet.

Part 3. Why the emphasis, why now, why Qubic

Every big AI today runs on ordinary servers with ordinary pseudorandom generators, seeded by whoever owns the box. Nobody worries, because today's models don't hold keys, don't trade against each other, don't evolve on their own. The day they do, "who holds the coin?" turns into "who holds the AI?", and most of the industry will find out it never asked.

Qubic's core, by the way, doesn't ask anyone for a coin. When I asked where it gets its randomness, CfB answered me: "we use RDRAND". That's the hardware generator inside the CPU, fed by thermal noise in the silicon. Physics. Which is why the chain held in October and what broke was the app next to it. It's not yet "randomness nobody can rig", you still have to trust Intel, but it's the right side of the road.

CfB's vision is that all of this becomes obvious within a decade. And that a network of hundreds of independent machines, where no single party controls the final roll, is where an unriggable coin can actually live. The wallet hack doesn't prove that bet. It loudly demonstrates the first half of it: a system with weak randomness is a system that belongs to someone else. We just watched one belong to someone else for an hour and forty-four minutes.

Conclusion

Seems like we're closer now.

Bookmark this tweet.