The Truth About AI, Quantum & Crypto's Encryption
Summary
The panel’s base case is that current AI math gains do not meaningfully weaken the cryptography securing major blockchains. ZK Dragon says OpenAI’s roughly 770 mathematical results came nowhere near factoring or the elliptic-curve discrete-log problem; his estimate is that an equivalent deployment of all Bitcoin’s specialized hardware would still require “literally a trillion times more computation.” His threshold for concern is actual progress on those core problems: “We’ve seen literally no progress on the big problems.”
Quantum remains the more credible long-term threat discussed, but its first high-value target would probably be encrypted state communications rather than cryptocurrency. Mert argues that a capable state would prioritize intercepted military, diplomatic, and other encrypted communications over blockchain signatures. The panel says a state would have limited chances before disclosure forced broader upgrades. Under an extreme hypothetical of a million-fold algorithmic improvement, ZK Dragon still estimates roughly 10 GW running for a million years with Bitcoin-class ASICs; GPUs would be more than 1,000 times slower.
Protocol adaptability—not “bunker mode”—is the investable defense against an uncertain threat horizon. Mert argues Zcash benefits from narrow scope, a paranoid cryptographic culture, and fewer programmable surfaces; Illia says NEAR separated accounts from cryptographic keys so it can add or disable signature schemes without changing addresses. Solana’s history of surviving outages and major rewrites is offered as evidence that temporary performance costs are preferable to an existential failure: “Either you go to zero or you survive.”
Even a disclosed cryptographic breakthrough would not instantly release every coin to an attacker. In the panel’s one-month-warning scenario, chains could freeze elliptic-curve operations while exchanges, cross-chain networks, and trading venues shut their boundaries; seed-phrase holders could later prove control through zero-knowledge recovery. Early Bitcoin balances predating seed phrases—including Satoshi-era coins—might remain frozen and unspendable, while MPC users present another exception.
AI-enabled smart-contract exploitation is the present risk, and the panel rates it vastly above foundational cryptography failing. Mert puts the relative risk of losing money to a contract hack at “probably a billion times higher—maybe a trillion times higher,” while Illia describes using AI to make formal verification 100–200x cheaper, formalize NEAR’s virtual machine, and turn contract guarantees into composable properties. The practical goal is not merely verified source code, but enforceable claims such as “if a user deposits money, they can withdraw it.”
Mert’s objection to Justin Drake’s framing is about proportionality, not preparedness. He grants that core developers should eliminate surprises, but calls “bunker mode” inappropriate because AI can strengthen systems as well as attack them, and because volatile charts create a “panic flywheel” on Crypto Twitter. His investor framing is low entropy: demand credible contingency plans, but recognize that acting on an extraordinarily remote scenario may lose more money than the scenario itself.
The panel’s durable end state is layered cryptographic redundancy rather than a proof that nothing can ever break. ZK Dragon imagines 256-bit security with different cryptographic assumptions arranged in a two-of-three or similar multi-layer scheme, allowing one broken primitive to be rotated without crisis; Illia adds continuously upgradeable software and formal verification. Post-panel, the hosts conclude that this favors systems designed to evolve—especially Zcash and NEAR—while creating a longer-run monetary question for a more rigid Bitcoin: “The basic principle of money is that I can hold it there and store value there.”
Deep dive
1. AI’s math breakthrough is not yet a cryptographic breakthrough
Thread Guy opens with two feared attack paths: quantum computation and AI-accelerated “manual decryption.” His Gaara analogy captures the anxiety—crypto’s protective sand stays flat while “more powerful attacks, more powerful opponents” accelerate toward assets worth trillions.
ZK Dragon replaces that framing with the actual hard problems: RSA depends on factoring a huge number into its prime components, while much blockchain cryptography rests on elliptic-curve discrete logarithms. Cryptography turns problems nobody knows how to solve efficiently into signatures and other useful systems.
The approximately 770 mathematical results publicized by OpenAI did not approach the analytical core of either problem. Cryptographers, intelligence agencies, and financially motivated attackers have examined these assumptions for years; ZK Dragon’s inference is that repeated failure is evidence of hardness, not proof of invulnerability.
Illia agrees on incentives: many showcased math problems may have received one researcher’s career, while breaking widely used cryptography attracts “every agency in the world.” AI scales applied intelligence faster than prior tools, but it is entering a field where enormous human and financial effort already exists.
2. A trillion-fold gap absorbs impressive-sounding improvements
ZK Dragon’s anchor estimate is that recreating Bitcoin’s global specialized-hardware footprint for an elliptic-curve attack would still leave roughly a trillion-fold computational gap under known methods. A 10x algorithmic gain merely turns that into 100 billion times Bitcoin-scale computation.
Mert translates a plausible 15% speedup into an engineering response: increase the parameter size—perhaps from n bits to n + 24 bits—and update affected systems. Even Matthew Green’s warning that public-key cryptography might be over was followed, Mert notes, by a less-amplified explanation of such adaptation.
ZK Dragon expects RSA to show a thousand-fold acceleration before elliptic curves do, affecting older browser TLS handshakes rather than blockchain assets. Asked how quickly AI is reducing the elliptic-curve gap now, his answer is categorical: “On the software side, literally zero.”
3. Famous math problems and cryptographic assumptions face different incentives
ZK Dragon complicates the incentive argument with academia’s treatment of Millennium Problems: professors may discourage PhDs and postdocs from spending careers on them because failure destroys tenure prospects. “If one day you succeed, okay, you’re a hero,” but researchers may be ridiculed throughout the attempt.
The disagreement matters because spectacular AI results on neglected problems do not map cleanly onto assumptions continuously tested by cryptographers. Mert’s trader-oriented version is simple: compare the payoff from a niche theorem with the payoff from reading state secrets, faking certificates, or stealing billions.
Thread Guy asks whether breaking one primitive should immediately discredit every other primitive. ZK Dragon rejects that inference: the field concentrates on roughly “one to five problems,” and success against an unused, lightly studied construction says little about the heavily attacked assumptions securing deployed systems.
4. Zcash, Solana, and NEAR embody three kinds of resilience
Mert’s case for Zcash begins with limited scope: it is a store-of-value network rather than a broadly programmable platform. He opposes adding collateralized stablecoins because every new function expands risk; a focused system staffed by “hyper-paranoid programmers” can prepare for edge cases years ahead.
Solana is harder to generalize about, but Mert points to its 2022–23 outages and subsequent major rewrite. If an emergency upgrade reduced performance by 10% for six months, he would accept the trade: survival permits later optimization, while an unaddressed extreme risk can take the asset to zero.
Illia says NEAR was designed in 2018 to separate human-readable accounts such as
illia.nearfrom their cryptographic keys. It supported ECDSA and Ed25519 from the beginning and added lithium, a network-based post-quantum cryptography option, this year, after standardization and implementation matured beyond cryptocurrency.A high-value cold account can use four or five distinct cryptographic keys and require them all to sign together, spreading implementation, mathematical, and quantum risk. Illia says disabling an old primitive is a three-line change and adding one across the protocol is about 200 lines. He separately cites Solana as a protocol updated roughly every six weeks.
5. An emergency freeze could preserve most funds—but not every balance
Tulip King stress-tests the worst disclosed scenario: OpenAI announces a quadrillion-fold elliptic-curve improvement and promises publication in 30 days. Networks could “pull the emergency lever,” disable elliptic-curve features, and freeze affected balances before attackers assembled sufficient hardware.
Seed-phrase users could later recover through a zero-knowledge proof showing knowledge of the seed phrase and HD path corresponding to a public key. Tulip King says most holders could emerge after months of inconvenience, though MPC accounts would require different handling.
Satoshi-era Bitcoin is the sharp exception: those coins can be frozen, but probably cannot use this recovery because they predate seed phrases. The protection might therefore leave historically early balances unspendable rather than handing them to an attacker.
Tulip King locates the immediate response at system boundaries. Exchanges, cross-chain networks, and trading platforms would halt first; chains could later implement or roll back to the point at which the decision was adopted because the urgent danger lies in reconciling on-chain state with other chains and off-chain venues.
6. Formal verification turns AI from attacker into defensive infrastructure
Illia redirects attention from speculative math to cyberattacks already amplified by AI: “This is a much, much more practical problem that actually affects users every day.” The challenge is securing applications built atop a sound base layer, not merely preserving the layer’s cryptographic assumptions.
A roughly 10-day swarm of agents might represent $10 million of compute, which ordinary teams will not spend. Illia says his team reduced many verification costs by 100–200x; a DeepSeek-based harness ranks first on two benchmarks while costing about 200x less than GPT- or Claude-based runs.
The next step is a fully formalized NEAR virtual machine. Proving a contract property is insufficient if the runtime does not implement the modeled behavior exactly; formalizing the underlying stack lets developers prove operational guarantees such as depositors retaining a withdrawal path.
Illia distinguishes this from Move’s compiler-enforced object properties, including token supply and prevention of double spending. Once applications introduce shared state—his example is Uniswap’s LP positions and pooled tokens—developers need broader logic proving behavior across every possible deposit, withdrawal, and trade.
7. “Bunker mode” mistakes preparation for panic
Mert gives Justin Drake the charitable case: perhaps he sees accelerating breakthroughs and wants the industry ready. His criticism is that a field leader has “a responsibility to explain both sides,” including AI’s ability to improve verification and help build defenses.
Zcash and other teams already have contingency plans, while engineers across the industry have strong incentives to keep the systems alive. Mert calls the probability of quantum breaking Bitcoin next year “practically zero,” yet still supports preparation because investors want low entropy and long-term guarantees.
What he rejects is the leap to “bunker mode.” Markets were already volatile before Drake’s post, and Crypto Twitter let price charts validate an alarming story, creating a “panic flywheel”; Mert argues that selling or abandoning blockchain is much more likely to inflict loss than the posited break.
8. Nation-state economics shape how a real break would unfold
Mert says quantum’s realistic first target is encrypted communication, not a transparent blockchain. A state would intercept traffic now and decrypt it once capable; access to an adversary’s military or diplomatic communications dominates the value of spending scarce early capacity on blockchain signatures.
The likely first actors are the United States or China. If OpenAI or another private company found the method, ZK Dragon says it would have to report it to the government, comparing independent possession to a company inventing nuclear weapons and predicting intervention or nationalization if it withheld the result.
An algorithm is only half the attack. ZK Dragon describes taking over TSMC for a year, supplying more than 10 GW, and operating infrastructure too large to hide; GPUs are more than 1,000 times slower than purpose-built ASICs. Even assuming a million-fold algorithmic improvement, his rough Bitcoin-hardware scenario still runs for a million years.
His cautionary analogue is 2013’s Logjam episode: Snowden’s disclosure preceded recognition that weak SSH parameters let a government spend billions rather than the normal quadrillions-of-years baseline. Discovery took about two years, but ZK Dragon argues that formal verification means there will never be a comparable accidental gap.
9. The practical end state is redundancy, rotation, and continuous learning
Thread Guy asks whether AI can deliver “perfect” cryptography. ZK Dragon says genuine finality would resemble proving P is not equal to NP and showing that the chosen problem cannot be solved faster—an outcome so strong that practitioners do not realistically expect it.
His attainable version is 256-bit security plus independent layers: use different cryptographic assumptions in a two-of-three or similar arrangement, so one failure causes “no stress” and can be replaced before another breaks. Security becomes an update process, not a claim that one primitive lasts forever.
Illia’s anticipated AI step-change is real-time learning beyond months-long pretraining runs or million-token contexts. He says LLM progress has been slower than he expected since experimenting with live coding in 2017, yet the panel is already proving zero-knowledge evidence with 128-bit post-quantum security.
Tulip King closes with a retail hierarchy of risks: a smart-contract hack is “probably a billion times higher—maybe a trillion times higher” than foundational cryptography failing. An unidentified closing speaker offers the more mundane worry list: supply chains, routers, and SIM-card provenance.