Doug Colkitt - Will AI Crack Crypto, Bunker Mode, and the Y2K Comparison

Summary

• Doug Colkitt sees AI’s mathematical leap as a regime change in uncertainty, not evidence that crypto will break within months. Models progressed from SAT and Olympiad-style questions to Erdős problems, a claimed Navier–Stokes result, and OpenAI’s release of 400 open problems across multiple fields. The concern is that “all bets are off the table” when nobody knows how quickly long-held hardness assumptions might fall.
• Justin Drake’s “bunker mode” addresses a real cryptographic asymmetry, but Doug considers the prescription extreme today. Elliptic-curve and lattice systems may contain exploitable structure, while hashing is comparatively “brute force and dumb.” Drake therefore proposed keeping public keys hidden behind hashes and moving assets to a fresh address after every transaction.
• A signature breakthrough would threaten far more than token balances. Doug says forged signatures could enable website and message impersonation, disrupt banking apps, and let attackers spoof validators so proof-of-stake chains become reorgable at will. His stark framing: “The internet as we know it is kind of broken.”
• Doug’s base case is a gradual migration, closer to Y2K remediation than an overnight extinction event. Hash-based signatures already offer a fallback, although they are cumbersome and potentially too large even for a Solana transaction. The decisive variable is speed: years permit coordinated upgrades; weeks might not.
• He is skeptical that OpenAI has secretly cracked discrete logarithms or elliptic curves because AI’s headline math results remain narrow and incremental. Models are “very, very good at checking a box,” often crossing a theoretical barrier by a tiny amount without immediate practical utility. He considers a sudden, operational cryptographic break possible but “pretty unlikely.”
• AI may strengthen crypto security before it undermines cryptography. Formal verification can turn program requirements into theorems and prove that code follows them, attacking the smart-contract bugs Doug considers today’s dominant risk. He thinks this could produce “quote unquote ‘perfect security’” and points to formally verified work in Zcash and his own Lean-based EVM AMM.
• The actionable defense is ordinary operational security, not constant wallet migration. Doug recommends holding most funds in a two-of-three multisig—Squads on Solana or Gnosis Safe on Ethereum—using two hardware wallets with different stacks plus a trusted mobile wallet. “You’re more likely to fuck up and lose your coins” attempting bunker mode than to have AI crack them today.

Deep dive

1. AI crossed from contest math toward open research

• Doug traces the progression from SAT questions to AMC and International Math Olympiad problems, then to challenging Erdős problems. He says they were difficult, though not necessarily especially important.

• The “shot across the bow,” in his telling, was AI solving Navier–Stokes, one of seven $1 million Millennium Prize problems, only one of which had previously been solved. He stresses that the result was technically significant even though its construction was narrow and had no immediate practical application.

• OpenAI then dumped 400 open math problems at once. More important than the count, Doug says, was its breadth: models that had been uneven across subfields suddenly “seem good at, like, everything now.”

• He attributes the shift mainly to chain-of-thought reasoning and reinforcement learning on verifiable rewards, not one magical discovery. Base models remain “autocomplete…super, super souped up,” but can now write out reasoning, check their work, and hill-climb on problems with verifiable answers: “data plus scale equals intelligence.”

2. Cryptography’s assumed hardness is suddenly repriced

• Modern cryptography depends on operations believed to be difficult, even though their difficulty is not always proven. The prior pace of mathematical progress made hard problems seem unlikely to break soon; discrete logarithms are one example. AI has widened the possible variance over the next year or two.

• Doug thinks Justin Drake oversold “months, not years,” yet accepts the underlying warning. Elliptic-curve and lattice signatures have mathematical complexity that might conceal “some weird way to reverse it”; hashing is less complex and comparatively brute-force.

• Drake’s proposed bunker mode exploits that distinction: Ethereum addresses hash public keys, but spending reveals the key. His remedy is to move everything into an unused address, then repeat after every transaction; Doug calls that “pretty extreme” and doubts it is warranted today.

3. Broken signatures would become an internet-wide crisis

• Doug broadens the blast radius beyond crypto: Zoom conversations, authenticated messages, websites, and banking apps all rely on signatures to establish identity. If attackers could forge them, “anyone can fake being someone else,” potentially forcing people back to in-person banking until systems migrated.

• Proof-of-stake chains face a second-order failure even if individual balances remain protected. Forged validator signatures could permit arbitrary reorganizations, meaning “the chain’s not gonna run.” Doug says Bitcoin would be the exception—the chain that keeps running—because proof of work depends on hashing rather than signatures.

• Hash-based signatures provide an escape route, and Doug thinks a lot of military systems have moved to them while NIST pushes them. Their drawback is operational: they are “really big” and “a pain in the ass”—so large, he says, that one signature would not fit inside a Solana transaction.

4. The Y2K analogy makes timing the decisive variable

• Doug remembers 85%–90% of people expecting Y2K to be fine while roughly 10% hoarded food. The calm outcome did not prove the threat imaginary: “People spent a lot of money to fix all those Y2K bugs.”

• His analogous crypto scenario turns on fast versus slow. A breakthrough unfolding over years gives networks time to adopt hash signatures; publication one month before exploitation leaves little runway. A week-scale leap looks “very unlikely,” months remain doubtful, and years are plausible enough to prepare for.

• Secret OpenAI breakthroughs cannot be ruled out, but Doug sees no supporting pattern. Existing results typically cross a boundary “by a very, very small amount”; cracking elliptic curves with immediate practical consequences would represent a qualitatively different capability.

• Early warning could come from lower-security “canaries”: perhaps $10,000 locked behind a 96-bit construction rather than 256 bits. An attacker improving gradually should harvest that low-hanging fruit before stronger systems—Doug’s signal to say, “Oh, shit, I gotta go into bunker mode.”

5. AI’s nearer-term payoff is better code and practical math

• Smart-contract errors remain the risk Doug actually loses sleep over after writing code associated with millions of dollars. Formal verification lets developers encode required behavior as theorems and prove the implementation follows them, making feasible work that was previously too tedious and lengthy for humans.

• He cites Zcash’s new shielded pools and his own EVM AMM written purely in Lean as early examples. Shielded Zcash may also be relatively defensive in a doom scenario because an observer does not see an exposed signature, but a Merkle-tree-based shield—though Doug concedes consensus could still break.

• Doug tested AI’s research usefulness on integer multiplication after people mocked OpenAI’s improvement as cosmically impractical. He reports OpenAI’s form as roughly 2^1.82m − 1.82 and his revised form as 2^1.34m—a huge reduction that remains astronomical. The transcript also contains a separate “1 over 2^82” phrase, so the normalized figures are unclear; others reportedly reached 2^16, though he said he needed to check.

• Absent one of those canaries being broken, his threat model remains mundane: North Korean hacking, zero-days on computers, and wallet compromise. His own two-of-three multisig setup diversifies hardware and software failure modes, while Baranos AI—his early project—is exploring AI adjudication of prediction-market outcomes.