In a world of ledgers, who holds the memory?
The market answered that question with a quarterly earnings beat. Block, the financial services company formerly known as Square, reported fourth-quarter results that pushed past consensus. Cash App carried the weight. Square carried the floor. Together, they delivered the kind of headline numbers that send analyst models into an upward scramble. Adjusted EBITDA came in ahead of expectations. Gross profit printed double-digit growth across both ecosystems. The stock reacted the way stocks react when the narrative is in the process of healing: with relief.
But I kept reading.
One sentence, buried deep in the shareholder letter, refused to leave my mind. The company said it expanded its use of artificial intelligence across software engineering. No drama. No press release flamboyance. Just a quiet operational note, the kind of line that earns a polite nod from a tech reporter and then vanishes into a summary deck.
I have spent twenty-six years watching protocols pretend to be people. I have audited governance contracts at three in the morning and watched a fourteen-million-dollar mistake get blocked by a single require statement. I have curated a carbon-neutral generative art exhibition on Tezos and walked the hills outside Boston whispering about digital souls. I know the difference between a codebase and a conscience.
Here is what I know about Block’s beat, and what the market missed:
The AI expansion is not a productivity story. It is a trust story. In a crypto winter, trust is the only asset with a meaningful drawdown.
Market analysts will annotate the quarter line by line. Gross profit from Cash App, gross profit from Square, bitcoin revenue, adjusted operating income, the slow recovery of consumer fintech after two years of contraction. All of that data is real, and all of it is secondary. The primary event is not the number. It is the quiet arrival of AI as the engineering steward of a platform that millions of people already treat as a bank.
We code the trust, but we must audit the soul.
That sentence has guided me since the 2017 audit era, when I used to read ICO whitepapers and feel the way an undertaker must feel at a party: everyone is having a wonderful time, and nobody wants to know that the body is already in the room. Today, the body is not a buggy smart contract. It is the assumption that more automation, generated and reviewed by large language models, will make financial infrastructure more trustworthy. That assumption is not just unproven. It is dangerous, because it is likely to be true in the short term and false in the long term.
Let me explain with the care the subject deserves.
Block is a company built from two products that share an unstated philosophy. Square gives small merchants the same payment rails that Visa and Mastercard once reserved for enterprise. Cash App moves money between individuals, offering direct deposit, a debit card, bitcoin buying, and, more recently, a taxable brokerage operation. Together, they form the most mainstream bridge between the legacy financial system and the crypto economy. Every QR code at a farmers market is a Square terminal in disguise. Every twenty-five-dollar bitcoin purchase made by a nineteen-year-old on a bus is a Cash App settlement arriving at a third-party custodial wallet somewhere in Nevada.
That last sentence deserves a pause.
Cash App is one of the largest retail bitcoin on-ramps in the United States. Its gross profit from bitcoin is notoriously thin. Block charges a small spread and passes the actual bitcoin through to a partner custody infrastructure. But user acquisition is enormous. For millions of people, Cash App is bitcoin. They do not use a decentralized exchange. They do not possess a hardware wallet. They do not know the difference between a hot wallet and a cold storage facility. They know only that the little green icon on their home screen holds their money, and that the company is audited by Deloitte, and that the government would never let something this big simply fail.
We have heard that phrase before.
In 2017, I declined a lucrative advisory seat with a prominent Ethereum-based DAO framework to perform an unpaid security review of its governance contracts. I found three critical reentrancy vulnerabilities in a codebase designed to automate collective decision-making on a blockchain. The vulnerabilities were not academic. One of them could have drained more than twelve million dollars from the project treasury. I spent three weeks in isolation, reading assembly-level output, tracing storage slots, reconstructing attack paths. I wrote the report in longhand first because I needed the arguments to be slower than the code. When the developers received the audit, one of them called me angry. He insisted that the smart contract language’s native transfer function protected against reentrancy. That misunderstanding would have cost his community everything.
I think about that angry developer now, every time a company announces an AI expansion across software engineering.
Because the new mistakes will be cleaner.
Proof is binary; meaning is fluid.
When I audited that DAO framework in 2017, the bugs were almost ugly. You could see the carelessness in the code. A missing checks-effects-interactions pattern. An external call before a state update. The signature of a developer who had learned Solidity from a blog post and never experienced the cold sweat of a drained balance. The patterns were visible to anyone with the patience to trace the execution path.
AI-generated code does not look like that. Large language models are trained on billions of lines of publicly available code, including the idioms of Ethereum, Solana, Bitcoin Script, and a dozen other execution environments. The code they produce is statistically idiomatic. It mirrors the shape of correct code. It passes linters. It compiles without warnings. It reads, on first inspection, like a competent engineer wrote it.
And then the edge case arrives.
The protocol is neutral, but the user is human.
An AI assistant asked to generate a vault contract might faithfully reproduce the standard withdrawal pattern. It might even include a reentrancy guard, because the guard appears in millions of training examples. But ask it to generate a vault that interacts with an external oracle, or a vault that performs a cross-chain callback, or a vault that must decide between two accounting treatments during a liquidation auction, and the statistical elegance begins to unravel. The guard is there. The ordering is not. The revert path is missing. The fee calculation rounds down when it should round up. The human reader sees what the model was trained to produce: a recognizable shape, a familiar silhouette of code.
I have started referring to this as the Velvet Fog Vulnerability. The code is not obviously wrong. It is softly wrong. It is wrong in a way that requires the reader to understand the intent behind the procedure, not just the syntax. It is wrong in a way that an AI reviewer, trained on the same corpora, will happily validate.
That is Block’s new AI reality. The shareholder letter says the company expanded its use of AI across software engineering. If you have engineering experience, you know what that means. Pull requests with AI-written descriptions. Test suites with AI-generated cases. Legacy code analysis with AI summarization. A company-wide culture shift toward accepting generated patches faster than a human review process can actually absorb. At scale, in a fintech company handling real money, real identities, and real enforcement actions, that shift is not an operational footnote. It is a change in the moral geometry of the codebase.
Now apply the same logic to Block’s bitcoin products. Cash App’s custodial architecture is a compliance-first design. The wallet is controlled by an institutional custodian. The addresses are subject to transaction monitoring. Funds can be frozen, clawed back, or seized in response to court orders. For most users, this is a feature. Block is a regulated financial services company; consumer protection is the brand promise. But for anyone who understands what bitcoin was invented to solve, this is a contradiction with the coin’s first principles.
I made the same argument about Circle and USDC in 2022, and I will make it here. Circle’s compliance-first strategy is what made USDC the preferred stablecoin for institutional investors. It is also what makes USDC a financial instrument of the state. Circle can freeze any address within twenty-four hours. That capability is the product, disguised as a risk control. When Cash App mirrors that model for bitcoin custody, it effectively transforms Bitcoin from a bearer asset into a claim on the solvency and goodwill of a corporate intermediary. The customer receives a promise. The ledger receives a shadow.
And the shadow is becoming increasingly AI-governed.
Every time a Cash App user sells a fraction of a bitcoin, the transaction passes through a decision layer. Is the user who they say they are? Does the counterparty address appear on a sanctions list? Is the transaction pattern consistent with money laundering, or with a normal person buying coffee with a legacy balance? Once, those decisions were human. Then they were rule-based. Now they are increasingly made, or at least pre-screened, by machine-learning models.
The efficiency gain is real. The moral leakage is invisible.
In 2020, I authored a whitepaper called Liquidity as Liberty. I argued that automated market makers could democratize financial access for the unbanked, providing liquidity without gatekeepers. That paper reached fifty thousand readers in its first month. It was born from a genuine conviction that decentralized protocols could outperform legacy intermediaries on every dimension that mattered: cost, accessibility, transparency, and resilience. Three core developers and I spent late nights refining the argument that financial sovereignty is a human right.
The bear market of 2022 tested that conviction. I watched centralized intermediaries collapse with the grace of a house of cards. High-profile exchanges, founded by people who claimed to embody decentralization, turned out to be nothing more than databases with a marketing budget. I withdrew from public discourse for six months. During that silence, I reached a painful conclusion: true decentralization requires not just distributed ledger technology, but robust governance models that anticipate the point-of-failure in every human organization.
Block is not a DAO. Block is a Delaware corporation with a founder-CEO and a board. Its bitcoin on-ramp is popular precisely because it is centralized. The trust is not cryptographic. The trust is legal, institutional, and cultural.
Which brings me back to AI.
When people ask me about the future of crypto, they usually expect me to talk about consensus algorithms, layer-2 rollups, or the latest DEX initiative. I am more interested in the engineer who reviews the code. When that engineer is an LLM, the accountability chain breaks. Not because the LLM is malicious, but because it is legible only to itself.
Consider the protocol’s governance. In the open-source world, a human who submits a malicious patch can be identified, confronted, and forked away from. A human who submits a subtly incorrect patch can be educated. An AI that generates subtly incorrect code at scale cannot be called to account. There is no commit to a memory, no emotional history, no reputation to lose. There is only a probability distribution over tokens, nudged by reinforcement learning into writing code that looks like other code.
People keep asking: who watches the watchers? The more urgent question is: who audits the auditor when the auditor was trained on the same flawed corpus as the code it is reviewing?
We are not moving money; we are moving belief.
Every earnings beat, especially in crypto-adjacent fintech, is a short-term affirmation of a long-term theology. The market sees Block’s better-than-expected results and concludes that mainstream adoption is working. Bitcoin revenue growing at a double-digit clip, Cash App gross profit expanding, Square’s merchant base stabilizing. But I parse the numbers differently. I see a centralized system growing thicker. I see more people holding bitcoin through a custodian that is increasingly leaning on AI for software engineering and transaction monitoring and, presumably, risk management. I see adherence to a model built on efficiency rather than sovereignty.
Do not misunderstand me. I am not calling Block a villain. I am calling Block a mirror. It reflects every tension that crypto has failed to resolve. We wanted self-sovereignty, but we built custodial apps that feel like banks. We wanted open protocols, but we built closed systems with compliance officers. We wanted reliable software, so we decided to let machines write and review the software, without asking the oldest question in philosophy: who is responsible when the machine fails?
And the machines will fail.
In 2021, I curated a digital exhibition of 150 generative art pieces on the Tezos blockchain. I chose Tezos because it was carbon-neutral, and because I was mid-way through a personal protest against the environmental devastation baked into proof-of-work infrastructure. I spent months building relationships with artists who shared my vision of sustainability and digital ownership. The exhibition attracted five thousand participants. It convinced me that ethical consumption was possible in crypto, and that aesthetic values could survive protocol migration.
But it also taught me something about memory. The art itself was immutable. The meaning was not. The same generative piece could be framed as a financial speculation, a cultural artifact, or an identity fragment, depending on the surrounding narratives. The protocol preserved the bytes. The community preserved the meaning. And when the community changed, the meaning changed with it.
That is what I fear about AI-generated software. The bytes will be preserved. The protocol will run. The numbers will balance. But meaning will be generated by a machine that does not understand mortality, loss, counterparty risk, or the difference between a rational liquidation strategy and a family emptying a college savings account.
Proof is binary. Meaning is fluid. AI excels at proof-shaped problems. It is catastrophically ignorant of meaning-shaped ones.
Now let me give the AI expansion the full respect of a contrarian argument. I am not a Luddite. I believe large language models will radically improve software engineering. They already do. In my own consortium work on decentralized identity frameworks, I have seen AI tools accelerate the tedious parts of protocol design: interface documentation, test scaffolding, formal specification drafting. The efficiency gain is not hypothetical. A well-designed LLM review assistant can catch the kind of trivial off-by-one error that once consumed an afternoon of human sleep.
The contrarian read of Block’s announcement is not that AI is wrong. It is that AI will make Block’s software better, more reliable, and more cost-effective. The market sees that as unambiguously positive. Shareholders should, on one level, celebrate. Lower engineering costs, faster feature shipping, higher gross margins, and a compounding moat against smaller fintechs without proprietary model infrastructure.
What the market fails to price is the centralization of cognitive authority.
When an organization adopts AI across software engineering, it does not merely delegate labor. It delegates judgment. The model decides what a normal function looks like. The model decides which patterns are idiomatic. The model decides which edge cases deserve attention, based on how often those edge cases appear in its training data. The result is a narrow but unrelenting homogenization of thought. During a liquidity crisis, the model-generated invariants will not imagine the unimagined. They will reproduce the anticipated. And in crypto, the unanticipated is the entire business model.
The 2017 DAO audit taught me that the bugs that matter are not the bugs you can search for. The critical reentrancy attack I identified did not rely on an obscure Solidity feature. It relied on a function that looked, to the human eye, like it was doing the right thing. The developer had read the documentation. The code compiled. The community reviewed it. But the internal accounting was slightly misordered, because the author assumed the transfer happened at the end.
Now imagine that assumption generated by a model, across a thousand contracts, at high speed. The model will not make the assumption because it is malicious. It will make the assumption because the assumption is statistically common. And because the code is produced in bulk, the review process will be more shallow, not deeper. The human reviewer will glance at the generated patch, see the idiomatic shape, and approve.
The shape is familiar. The soul is absent.
The protocol is neutral, but the user is human.
The user will still lose twelve million dollars. The only difference is that the loss will feel more modern.
Block’s position in the industry is unique because it has institutional credibility and crypto ambition. Jack Dorsey has always framed bitcoin not as a get-rich instrument but as the internet’s native currency. Block’s treasury holds a meaningful amount of bitcoin. The company builds mining rigs, explores decentralized identity, and has publicly discussed the dream of an open protocol for money. Every product decision walks a tightrope between the permissionless vision and the regulatory reality that, as a public company issuing quarterly guidance, it cannot afford to annoy.
The earnings beat should not be read as proof that the tightrope has become easier to cross. It should be read as proof that the tightrope is tighter than ever.
Cash App and Square drove better-than-expected results because they are familiar, regulated, and corporate. The user experience is smooth. The compliance burden is hidden. The bitcoin is bought and sold without the user leaving the app. All of this is exactly what you want from a consumer product, and the opposite of what you need from a financial governance layer.
The AI expansion takes that contradiction and puts it on rails. If the trust layer of the company is increasingly generated by a model, then the trust itself becomes a product of statistical approximation. You can no longer point to a human engineer who vouches for the withdrawal function. You can point to a model card, a training set, a benchmark score. None of those things can feel shame. None of them can be called into a congressional hearing and asked whether they understood the consequences of a faulty liquidation logic. None of them can recall the 2017 DAO and choose caution.
I was thirty-five in 2017 when I wrote that audit report in longhand. I am forty-two now. I spent the last decade building, auditing, and occasionally tearing apart decentralized systems. I have seen the ICO boom, the DeFi summer, the NFT painting on the wall, and the bear market that washed it all back into the sea. I have learned that the protocol is only as honest as the weakest assumption in its governance model.
AI is now the new assumption.
Block’s expansion of AI across software engineering is not a scandal. It is a signal. It tells us that the industry has accepted a model of development in which trust is manufactured by pattern recognition, and accountability is deferred to whatever entity owns the model. For Bitcoin, the last thing we need is more deferred accountability. We need the opposite. We need a system where the memory of past failures is embedded in the culture of the people who write the code.
I understand the temptation. I live in Boston, where startups ask me about AI daily. I see the leap in productivity. I see the appeal of generating a hundred test cases in seconds. I also remember the bored developer who once wrote a governance contract with a fatal reentrancy bug, because he was excited, because he was rushed, because he didn’t sleep the night before. AI never sleeps. AI never gets excited. AI never rushes. AI never loses focus. That is precisely what frightens me. AI never loses focus, which means it never stops pretending that the tricky cases are exactly like the average case.
In a world of ledgers, who holds the memory?
Block’s ledger is growing. Cash App users are buying bitcoin. Square merchants are processing payments. The AI is learning, and the software is shipping. But the memory of why we built this industry is still held by humans. We cannot outsource that memory to a model. We cannot let a generation of engineers grow up believing that code is trustworthy because it passed a machine review.
Code is not trustworthy. Code is a series of promises made by someone who did not truly know what they were promising, to someone who did not truly understand what they were being promised. The only way to keep those promises is to have humans who remember when promises were broken.
In the 2022 crash, I watched a centralized exchange that everyone called a pillar of the industry collapse into bankruptcy. The jokes were quiet. The remediation was loud. The lesson, I thought, had been learned. But now the industry is racing toward AI generation, AI review, AI compliance, and AI risk management. We are not repeating the mistakes. We are delegating the capacity to make mistakes. We are making a bet that statistical elegance is enough to prevent moral collapse.
Block’s better-than-expected results suggest the bet is, in the short term, profitable. Cash App grew. Square stabilized. The narrative healed. But I cannot shake the image of a future audit trail where every line of code was generated by a model, and every line of code was reviewed by a different model, and every human involved in the process is quoted as having “trusted the output.”
At that moment, the soul of the system will no longer be an auditor. It will be a statistician. And statistics, beautiful as they are, do not grieve the users who lost their money.
What do we do with this insight? How do we keep the benefits of AI without surrendering the ethical memory of the industry?
I think the answer lies in governance, not in gratitude. Organizations like Block must build explicit, disclosed boundaries around AI-generated code. Not a ban. A boundary. Every AI-generated patch should carry a visible provenance marker. Every security-critical smart contract should require a human audit that takes at least as long as it would have taken a human to write the contract from scratch. Every model used for risk scoring should be subject to an annual independent audit, not for performance, but for moral alignment. We do not need to stop the machines. We need to remember that the machines are not accountable to anyone.
The protocol is neutral, but the user is human. The user is the one who trusts the green icon. The user is the one who sends money to the wrong address. The user is the one who smiles at the merchant, scans the QR code, and never thinks about the ledger at all. The user is not a set of tokens. The user is a person, with a memory, and a family, and a fear of loss.
We code the trust, but we must audit the soul. The audit will fail if we delegate the code and then delegate the audit. We will end up with a system where no one remembers, no one grieves, and no one is responsible.
The quarter was good. Cash App and Square drove better-than-expected results. The company expanded its use of AI across software engineering. All true. But while the market celebrates the present, those of us who live in the past’s failures have a quieter obligation. We must hold the memory. We must ask the question that no language model will ever ask, because its training data does not include the feeling of watching a community lose everything: what does the user need to believe, and will the code still honor that belief when the average case stops being average?
I do not know the answer. I do know the question is more valuable than any quarterly beat.
Because in a world of ledgers, the memory is all we have. And if we hand that memory to a machine, the machine will keep it perfectly, and understand it not at all.
The market saw the headline. I saw the sentence. The sentence is the real story. The sentence is the beginning of the next audit, the next collapse, or, if we choose carefully, the next liberation.
We code the trust. We must audit the soul. Let us begin.
