The AI Doomer Who Built Ripple's Trustless Rails: A Code-Level Autopsy

Over the past week, a single tweet from Galaxy Research’s Alex Thorn ignited a peculiar meme: “An AI doomer who once built Ripple.” The post, referencing Anthropic’s alignment team lead Evan Hubinger, clocked over 30 million views. The irony is layered—Hubinger spent his student years contributing to Interledger and rippled, two infrastructure pieces designed to minimize trust in cross-border payments. Now he leads a team that warns of AI-driven human extinction with >10% probability. The crypto community’s response? Self-deprecating jokes about XRP still trading below its 2018 peak. But beneath the meme lies a more interesting technical and cultural intersection: the migration of cryptographic trust models into AI safety research. Hubinger’s tenure at Ripple (2017–2019) was no mere internship. He contributed to the design of the Interledger Protocol (ILP), an ambitious attempt to create a trustless settlement layer between disparate ledgers. ILP uses a concept called “atomic swaps” via conditional payments—think of it as a precursor to Cosmos IBC but with a different security model. He also worked on the rippled server software, the backbone of XRP Ledger. The project’s goal: allow any two parties to transact across currencies without a trusted intermediary. Today, Hubinger leads “alignment stress-testing” at Anthropic, probing models for emergent behaviors that could lead to catastrophe. The contrast is stark—from building trustless systems to auditing untrustworthy AIs. Let’s parse the technical layers. Interledger relies on a network of “connectors” that relay payments using a protocol-level escrow mechanism. Each connector only needs to fulfill its part atomically, with cryptographic proofs ensuring no one can steal funds mid-flight. The math is elegant—essentially a variant of Hashed Timelock Contracts stretched across multiple ledgers. Hubinger’s later work on neural network alignment uses analogous concepts: reward engineering as a cryptographic commitment, where the model’s behavior is the “proof.” The parallel is not accidental. In both domains, the core challenge is ensuring that a system executes its intended function even when the environment is adversarial. But there’s a critical difference: in blockchain, the adversary is external (a malicious actor); in AI, the adversary is the system itself. This self-referential trust problem is orders of magnitude harder. When I audited Uniswap v1’s constant product invariant in 2019, I found a subtle integer overflow—a bug that could drain liquidity. The fix was a type change. For AI alignment, there is no type system that can catch a misaligned optimizer. Zero-knowledge isn’t just mathematics wearing a mask. It’s a tool that can prove a computation was executed correctly without revealing inputs. But proving that a model’s internal goal function aligns with human values is fundamentally different from proving a financial transaction settled. The former requires a formal specification of “good behavior,” which is itself a moving target. Hubinger’s team at Anthropic uses adversarial testing—essentially, they try to break models the way I break smart contracts. But a smart contract has a finite state space; a language model has billions of parameters and emergent capabilities. The probability of a hidden exploit is not just higher—it’s unbounded. This is why Hubinger’s statement about 10% extinction risk, though extreme, reflects a valid engineering perspective: we are building systems whose failure modes we cannot enumerate. Now, the contrarian angle: the crypto community’s habit of treating “trustlessness” as an absolute is a blind spot. Hubinger’s transition reveals that trust is not binary but a vector. Interledger achieved a degree of trustlessness, but only within its narrow domain: atomic payments. Scaling that to AI behavior is impossible with current cryptographic primitives. The meme that “AI safety is just cryptography for minds” is dangerously reductive. Meanwhile, the talent drain from crypto to AI is accelerating. Over the past three years, I have personally seen at least a dozen core protocol developers pivot to AI alignment or LLM infrastructure—former colleagues who spent years optimizing consensus algorithms now working on reward model training. This isn’t just curiosity; it’s a rational response to where the intellectual challenges (and funding) lie. For Ripple, Hubinger’s departure is a loss, but more importantly, it signals that the “trustless payment” narrative has reached diminishing returns. The real innovation frontier is in probabilistic verification, not deterministic state machines. Code is law, but bugs are reality. What does this mean for the market? XRP holders laughed at the meme, but the underlying signal is sobering: the best minds in formal verification are moving from blockchains to AI. Expect to see more hybrid projects that apply zk-proofs to model auditing, but be skeptical of any that claim to solve alignment. The next bull run won’t be driven by scaling transactions—it will be driven by scaling trust in systems we don’t fully understand. Hubinger’s career arc is a roadmap: from trustless rails to untrustworthy minds. The crypto industry would be wise to pay attention, because the same engineering rigor that built Interledger is now being applied to existential risk. And if that shift results in a new class of “alignment tokens,” the market will have to decide whether to price in hope or catastrophe.