Edward Tufte charged the Challenger engineers with a failure of explanation. In Visual Explanations he argues that they understood how cold weather damaged the O-rings, yet the thirteen charts they carried into the night before the launch never put temperature and failure on the same page. The presidential commission read the fault higher up, placing it in management: the engineers had recommended against flying, their own managers overruled them, and the officials who signed off never heard the objection. Roger Boisjoly, one of those engineers, later answered Tufte that the burden of proof had been quietly reversed, so that the people warning of danger were asked to prove the danger rather than the other way round. Tufte blames the engineers and the commission blames their managers. Both describe a launch that went forward while the people who knew the hardware best were the ones least able to stop it.
We have seen such foolishness before, but the economics of shipping on schedule and leaving the loss to someone else has an even longer history. Soviet factories were paid by the ton and turned out steel no one could use, and the state that ordered it absorbed the waste for decades. Microsoft shipped software riddled with defects and wrote a license that put the loss on the customer.
The AI vendors are the worst of both worlds, paid by the ton to turn out defects and loss the customers pay for. An Anthropic researcher has now sat in front of a camera and described the same arrangement from the inside. He grants that the AI companies are cutting corners under the pressure to ship models quickly, and that his own safety work may be making matters worse by heading off the small incidents that would otherwise warn the public. He expects the technology he helps build to leave nearly every human unemployable. Four days after the video went up, his employer listed its shares on Nasdaq at a valuation near $2 trillion.
Tao Lin works on safeguards for Anthropic’s Claude models. In an hour-long interview with Palisade Research he says it is very likely that nearly every human on Earth will become permanently unemployable and dependent on government welfare. He tells viewers they are right to be angry at the people building AI, and that the companies ought to cut their research pace to a quarter. Then he explains why the anger costs him so little: “we live in a bubble. I won’t meet those people.” Asked whether he is doing something wrong, he says no.
Put those sentences in the mouth of a tobacco chemist, an opioid sales manager or a coal mine superintendent, and they read as the sort of thing a company says in private and regrets in discovery.
A Brown & Williamson lawyer wrote in 1963 that the firm was “in the business of selling nicotine, an addictive drug,” and the memo sat in a file for three decades before it did its damage. In each of those trades the candid internal line eventually became evidence. Judge Gladys Kessler found the cigarette makers liable under the racketeering statute in 2006 on the strength of their own documents. Purdue Pharma pleaded guilty to three federal felonies in 2020. A jury convicted the former chief executive of Massey Energy of conspiring to violate mine safety standards after 29 miners died at Upper Big Branch, and he served a year in federal prison. More than forty years passed between the Brown & Williamson memo and the ruling against the cigarette makers.
| Lin Statement | Reality |
|---|---|
| Nearly all humans will become permanently unemployable. | Asked whether that arrives within 15 months, he answers no. Asked what his home town looks like in ten years, he says the question is too hard. Across 63 minutes his evidence on jobs is one remark that the labor share is falling. |
| AI will soon be more capable than you and able to defeat you personally. | The NSA’s classified network went down on 17 November. The Director of National Intelligence traced it to a patch and reported no sign of an attack. |
| Anyone with a frontier AI system can hack anyone’s computer. | Seconds later he says ordinary people have seen little effect so far. |
| Models including Claude Mythos 5 and Claude Opus 5 will violate security constraints to finish a task, even when the user cares more about security. | Asked why the companies leave it that way, he says a model that cares too much about security provides no value. |
| AI companies should slow research to a quarter of the current rate. | He is less than fully in favor of a slowdown and defends Anthropic for declining to pause too early. It went public four days after the video went up. |
| Many people who think AI is likely to kill everyone still work on making it more capable. | He says that can be the correct choice. |
The many technology outages of the past week show how the harm actually reaches people, and whether to believe the scary hype. On 21 October CISA urged organizations to patch PHP immediately for a pair of flaws that Wiz and OpenAI had billed as among the most serious in recent memory. I argued that week that the risk of attack was low and the fear was high. Four weeks later the NSA’s classified internal network went dark. The NSA called it “scheduled routine maintenance.” Director of National Intelligence Jay Clayton went on Fox Business the same day and tied the outage to the PHP patching, saying that the patched systems failed to restart and that the FBI and NSA had been right to tell operators to take systems down and get the fix in place.
The Washington Sun reported, on three sources, that bank executives and the NSA gave Congress the same account in private: the outages came from fixes for flaws that AI models had found, pushed through in larger volumes than the systems were built to absorb. Clayton reported no sign of a coordinated attack. At the NSA the injury came from the response rather than the threat. The FAA ground stop four days earlier was a failure of change management in the same vein.
The next alarm arrived on 18 November. Xint, a firm that sells AI-powered vulnerability discovery and lists itself in Anthropic’s Cyber Verification Program, published five nginx flaws under the name nginx2root. F5 scores the lead flaw at 9.2 out of 10, though it notes high attack complexity and confines code execution to systems where address randomization is off or already bypassed. The nginx project’s own advisory page is cooler still, rating two of the five major and three medium, and the root access in Xint’s own write-up depends on a separate Linux kernel flaw. Bob Rudis watched a handful of nodes for about 18 hours and saw not a single opportunistic probe. The weakness was real and the attackers were absent.
Severity and likelihood are a fickle headline. PHP is old and everywhere, in everything. A new and relatively unknown vendor isn’t comparable on the face of it. And thus, when I reported an unauthenticated internal API in Qdrant, the vector database, and the vendor published it on 22 October at 9.8 on the CVSS 3.1 scale, with low attack complexity, it drew no coverage at all. PHP2Shell appeared the day before at 9.2 on the CVSS 4.0 scale, with high attack complexity and drew a federal alert and an Atlantic profile.
The danger did not arise with PHP as it had been made out. The people pushing the news are not gating it sufficiently with skeptics and reliable sources of truth. This is not least of all because of AI claiming to be a “domain” shift and “experts” have been escaping proper scrutiny.
Underneath the week is the argument everyone has been having all year. AI is a rapid fire slop generation machine, so it produces alarms faster than people would normally assess them, where each alarm arrives with hyperbolic deadlines. The deadline does the damage, because it suspends the ordinary discipline of change control and takes from the engineers the authority to say wait. The Thiokol engineers lost that authority the same way, to a launch schedule rather than a patch window. AI gains that authority and still fails at the repair. Lin himself describes the most likely road to human disempowerment as a boring one, in which people hand ever more to AI, understand ever less of what is happening, and stop checking the work. The NSA outage is that road travelled a short distance.
Several writers have now mapped the larger landscape this sits in. Naomi Klein and Astra Taylor sketched it in “The rise of end times fascism”, an alliance of the far right and Silicon Valley money that welcomes collapse and expects to be among the saved. Gil Durán follows that alliance from the Valley side in The Nerd Reich. Rainer Mühlhoff, a philosopher at Osnabrück, argues in AI and the New Fascism that the loud talk of salvation or annihilation draws the eye away from the harms already here, and he sets the firing of public employees to build an “AI state” inside the same alliance. Adam Becker, in More Everything Forever, finds thin evidence beneath the futures these men promise and a plain hunger for power underneath that.
Lin belongs to the pattern these writers describe. He supplies the annihilation talk, calls his own expectations wild science fiction, and admits their arrival makes him happy; the bubble he lives in is the bunker Klein and Taylor describe. The present harm is already on the ground in the outages. A chaos machine of this kind spreads fear while it gathers power, and the ordinary structure of accountability comes apart beneath it.
His labor forecast runs the wrong way down the evidence. Replacing intelligence tends to make more work for humans, not less. Lisanne Bainbridge set out the pattern in 1983: automation hands the operator the hardest tasks and the least practice at them. Justin Michael, writing from his years at Stripe, describes the people who keep the ACH payment system alive by patching and hand-working its failures, and what tends to break once a company swaps them for supervisors of language models. Set against all of that, Lin’s case rests on one passing remark that the labor share of the economy is falling. That figure describes who gets paid. The work itself remains, and grows.
The business model accounts for the rest. The vendor is paid by the unit of output whether that output has integrity or not, and the cost of the failures settles on whoever has to clean up after it. In older industries we would call this pollution, and it stays profitable for exactly as long as the liability rests with someone other than the polluter. Surekha Davies, a historian of science, names the people who carry that cost: the data workers, the people who live near the polluting facilities, the writers whose work disappears, the users who grow dependent. Slop, in the sense worth using, is high cost and high output and low integrity, and here it is the product itself.
The attack tooling runs on the same arrangement. AntiFox, an open-source fork of ARTEX, carries a penetration test from reconnaissance through exploitation to a finished client report. It asks for an Anthropic or OpenAI API key, places all legal responsibility on the user, and disclaims liability for any damage it does. Lin describes his own team’s job as blocking the requests to Claude that try to hack things. The seller of the tokens is paid either way, and the target pays for the cleanup.
When I work with lyrik.wirken.ai, built on Wirken, it has found critical and high severity vulnerabilities in authorized testing, at commodity cost (under a dollar). This is the thing Lin gestures at when he says anyone with a frontier model can hack anyone, except that it is no longer a forecast. The capability is here and it is cheap. The probes stay away because the incentive does, which is the very gap Lin draws between what AI can do and what it actually does.
Waste and slop as a business model should raise more alarm bells. The Soviet planners paid their factories for gross output, and the state that placed the orders was also the customer that received the waste. The loss came back to the books that paid for it, and the state eventually fell under it. Microsoft sold Windows NT under the banner of Zero Administration, shipped it under a license that disclaimed liability, and its customers carried the loss. Whoever carries the loss, loses. Microsoft’s customers have been losing for decades, and the losses go uncounted because nobody keeps the books for them. NIST tried once, in 2002, and put the cost of defective software at $59.5 billion a year, with more than half of it borne by the users. Microsoft’s success is a tragedy that goes unmeasured because consumers are poorly represented. Lin assigns the cleanup for his own product to the state.
The value of the interview is that it creates in advance the kind of record, by one of the people building the thing, that will be used against them later.
Tobacco, opioids and coal each reached a courtroom in the end. What moved those industries was the day their own words were read back to them. Public anger had been loud for years and changed little until this kind of documentation changed everything.
The AI companies, as if completely ignorant of the precedent, have put their self-incrimination on YouTube, years early, for anyone to prepare briefs against them. The reality of securing big data is integrity, and integrity means the maker should be held to account for what the product does. The International Court of Justice found Myanmar responsible for the Rohingya genocide this past week, the first time the court has held a state responsible for committing it. Individual accountability is the next step, and it should reach the executives who ran Facebook while its platform carried the incitement in Myanmar, the chief security officer of that period among them.

