Hook
An Anthropic safety researcher publishes a claim: AI has a >10% probability of “killing all humanity.” No methodology. No simulation logs. No hash of the model’s internal red-team report. The bytecode lies; the transaction log does not. Here, the log is empty.
Jacob Coxon’s resignation letter on X screamed “They are racing toward self-evolving superintelligence, gambling with our lives.” Two employees left within hours. The market reacted with silence — no volatility spike, no price dislocation. But volatility is noise; structural flaws are signal. The structural flaw here is not the AI itself, but the complete absence of a verifiable risk model.
I have spent 24 years watching crypto projects make identical claims: “Our protocol is secure.” “Our audited smart contract has 0% risk.” Then the transaction log reveals a reentrancy bug that drained $50 million. In crypto, we demand reproducibility. In AI safety, we accept a blog post with a single number and a moral panic. That is a failure of forensic integrity.
Context
Anthropic is the poster child of “safe AI.” Founded by former OpenAI employees, it raised billions on a thesis that Constitutional AI and red-teaming can bound existential risks. Its internal safety culture was once considered a moat. Now two employees walk out, citing an unquantified >10% existential catastrophe probability. The public narrative: “Even the safety researchers are terrified.”
But as a data detective, I cannot verify that number. The claim lacks any of the following: - A causal model linking model capability to extinction. - A Bayesian update based on observed training failures. - A Monte Carlo simulation with reproducible seeds. - An open-source audit trail of internal stress tests.
Compare this to how I evaluate a DeFi protocol. When Compound tweaks its interest rate model, I pull 50,000 historical transactions, simulate liquidation cascades under 12 volatility scenarios, and publish the raw data. The AI industry, by contrast, operates on faith. Faith in the researcher’s intuition. Faith in the lab’s closed-door red-teaming. Faith in a probability that cannot be peer-reviewed.
Pressure tests expose what calm markets hide. The calm market here is the silence from Anthropic’s leadership. They have not released a rebuttal, a counter-probability, or a third-party audit of their risk framework. That silence speaks louder than any tweet.
Core
Let me apply the same forensic rigor I used in 2021 when I detected wash-trading patterns inflating CryptoPunk floor prices by 15%. I traced wallet clusters, time-stamped transactions, and identified 12 addresses that were cyclical self-trading. The evidence was reproducible. Anyone could run the same query on-chain.
Now examine Coxon’s >10% claim. Deconstruct it into components that could, in theory, be verified:
- Model capability cutoff: What specific capability (e.g., autonomous replication, persuasive manipulation) triggers the existential scenario? The article does not specify. Without a defined trigger, the probability is meaningless — it is a floating signifier.
- Base rate reference: Existential risk from AI is a one-off event. There is no historical frequency to calibrate against. Any probability is a subjective prior, not an empirical frequency. Coxon’s 10% is a guess dressed as a statistic.
- Causal pathway: “Self-evolving superintelligence” is a vague concept. Does it require recursive self-improvement? Agency? Access to critical infrastructure? The pathway matters because it determines what safety measures exist. Constitutional AI, for example, is designed to bound reward hacking — but if the pathway is through exploitation of software vulnerabilities, then constitutional AI is irrelevant. Coxon provides no pathway. The bytecode lies; the transaction log does not. But there is no log.
During my DeFi stress testing in 2020, I modeled liquidation risks by running 100,000 Monte Carlo simulations with varying collateral ratios and price impact curves. I published the source code. Aave and Compound could replicate my results. If they disagreed, they could show their own data. That is how science works.
The AI safety community, however, operates like a cult of internal intuition. Researchers propose probabilities based on “feel” or “informal reasoning.” This is not good enough for a technology that could, by their own admission, kill everyone. Reproducibility is the only currency of truth. Coxon’s claim is bankrupt.
Let me propose a minimum standard for any future existential risk claim: - Publish the full set of assumptions in a machine-readable format (JSON, CSV). - Provide a sensitivity analysis: how does the probability change if you vary key parameters? - Include a log of all internal red-team exercises that informed the estimate, with timestamps and wallet signatures.
Absent this, the claim is noise. And I have spent 24 years learning that noise is not signal. Silence in the logs speaks louder than tweets.
Contrarian
The contrarian angle is not that AI existential risk is low. It is that the current method of communication — a blog post, a resignation, a tweet — is structurally identical to how crypto scams operate. The scammer says: “Trust me, my token will 100x.” The AI safety researcher says: “Trust me, my internal model says 10% extinction.” Both rely on authority and narrative, not on verifiable data.
I was part of the 2017 Solidity audit wave. I found integer overflow bugs in three ICO projects that prevented $2 million in losses. The developers did not mean to steal money. They made honest mistakes. But they also refused to publish their audit logs. Sound familiar?
Anthropic and OpenAI are racing. That is a fact. But the race is not just about model capability. It is a race to define the narrative of risk. Whoever sets the probability threshold controls the conversation. Coxon’s >10% is a political move, not a scientific one. It pressures regulators. It attracts media. It gives moral cover to resignations. But it provides zero actionable information for anyone who wants to actually mitigate the risk.
Data does not dream; it only records. The record here is empty.
The real structural flaw is the lack of an independent, transparent risk audit framework for frontier AI labs. The crypto industry learned this lesson after Mt. Gox, after Luna, after FTX. We now have on-chain attestations, proof-of-reserves, and real-time risk dashboards. The AI industry has none of this.
If Coxon truly wanted to reduce existential risk, he would have published the methodology. He would have opened the black box. He did not. That is not an accident. It is a symptom of a culture that prizes moral signaling over technical transparency. Trust the hash, verify the execution path. The path is invisible.
Takeaway
Next week, watch for one signal: does Anthropic or OpenAI publish a formal risk framework with open-source simulation code? If yes, the >10% claim becomes a data point. If no, it remains a ghost in the machine.
I will be monitoring the transaction logs of their public repositories. If no code appears, I will treat the claim as unverified noise. And I will continue to apply the same principle I have used since 2017: The bytecode lies; the transaction log does not. Coxon’s log is empty. So is his probability.