Legalize MATH

Magic Internet Math is back. After the summer that set the Bitcoin world on fire — the Cold Card entropy disaster — Brian and Rob Hamilton (AnchorWatch) reunite for a two-hour deep dive on the event Rob spent three sleepless weeks at the center of, and what it taught them about AI, entropy, censorship, and self-custody.

The spine of it: when funds started moving off Cold Cards, Rob loaded the firmware into the LLMs. Claude and Codex hedged and downgraded; Kimi K3 — a Chinese open-weight model — instantly printed the vulnerability. That contrast becomes the episode's thesis: the machines see more than they let on, and the open models will say what the censored ones won't. From there it opens into the math of entropy (why you roll your own dice), a love letter to libsecp256k1, and why sovereignty requires actually knowing the math — because they can censor the models, but they can't ban the math.

In this episode:

• The Cold Card entropy disaster — MK3 (~20–32 bits) vs MK4 (~70 bits), "cannon fodder for the next wave," and Rob's pinned-tweet triage guide for single-sig and multisig holders

• Using LLMs to find the vulnerability — Claude and Codex censored and hedging, Kimi K3 printing it instantly: "the machine sees more than it lets on"

• Open-weight Chinese models as the workhorse — LLMs as "slot machines" that spit out vulnerabilities; 100 scans at $100 beating one $10k elite scan

• Why you roll your own dice: y² = x³ + 7, 2²⁵⁶ points, equal probability, and XOR (addition mod 2) to combine entropy the machine can't fake

• "Cryptography is a weapon; entropy is the bullet" — brain wallets, hash functions, and why crypto books assume good entropy

• Bitcoin Core's four entropy sources, Alex Waltz's call-trace deep dive, ralo mcfluid, and why "Satoshi's keys are the AD-IQ proof" the RNG is sound

• Custody done right — River and owning your stack, Prime Trust, "only the paranoid survive," and "the soundest sleepers are the first to get wrecked"

• FROST and Frostsnap (Lloyd & Nick), nonces, and the two places you actually need randomness

• Agent-orchestration reality — lazy downstream agents, "any ambiguity in delegation will be used against you," front-loading the architecture, and interviewing the CTO to grill yourself on the gaps