Eli the Computer Guy says Anthropic's own blog post documents people trying to use Claude to build bioweapons, Houthi rebels using it to engineer cruise missiles and Russians using it for criminal attacks — and that the conclusion the company drew from that, slow down, is the wrong one.
The public argument since last week's AI-safety warnings has been about whether the models are becoming dangerous enough to pause. His is about what a company that believed its own warnings would do first, and his answer is the thing every American bank already does to open a savings account.
"If OpenAI and Anthropic were half as dangerous as they say they are, the first thing that you would implement is robust, not even robust, adequate know your customer laws adequate."
He has been in the technology world for 30 years, holds a degree in criminal justice, and is 50 — which is the frame he keeps coming back to, because it puts nuclear winter, Y2K and the Mayan calendar inside one working life.
The full interview is covered here so you can skip it. 34 minutes of audio, 18 minutes of reading.
Here are the 11 arguments that matter.
👤 Guest: Eli the Computer Guy, a technology commentator and YouTuber who has spent 30 years in the industry
🎙️ Host: Isaac Pound, who presents The Tech Report for Times Radio
📰 Published: 14 September 2026 on YouTube and The Tech Report's podcast feed (Times Radio)
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 34 min | ✅ Time saved: 16 min
Key Takeaways
The question nobody answers is what mechanism an AI would use, and he says the answer decides everything
A genius-level computer he can run over with his car is smart, not dangerous
The realistic AI threat is the internet going down, not humanity ending
Every ransomware attack of the past 15 years happened before AI, which he treats as proof the problem is cyber security rather than intelligence
Compute costs are a limit on rogue agents that nobody prices in
He reads the doom cycle as an American habit rather than a technical finding
Mutually assured destruction, Y2K and the Mayan calendar all inside one 50-year life
The labs are selling a story, not a technology stack, and the story has to carry a $2T valuation
Know-your-customer rules would solve much of this, and that is exactly why they are not being proposed
Verifying users means admitting you have fewer of them than a billion
His regulatory proposal is a grand jury for Dario Amodei and Sam Altman, not a new statute
He contrasts Aaron Swartz facing 30 years with executives announcing conduct in press releases
External evaluators are just auditing, and he has almost no objection to it
The two things that decide whether it works are what they get to look at and whether anyone has to act on it
Banning more powerful models misses where the value has moved: the harness, not the model
The AI-harms argument is a mental-health argument wearing a technology costume
1. What Is the Mechanism?
Isaac Pound opened on the week's news: "So last week the threat of AI going full Skynet was renewed after an Anthropic and former OpenAI employee quit over fears that both companies were kind of racing to develop self-improving super intelligence and ignoring the risk that it could wipe out humanity in the next 10 years," he said, noting both companies have called for a slowdown, and asked what his guest made of the noise.
The answer was three sentences long: "Am I allowed to say dumb? I think it's all kind of sort of incredibly dumb. Doesn't make an ounce of sense."
Pound pressed on the specific incoherence he had noticed — the warnings call AI an extinction-level threat and then say quietly that none of it applies to today's large language models — and asked what exactly is being warned about. "I honestly have no idea," Eli said
His working frame separates three things that get discussed as one. "There's a stack and then there's the business. So, OpenAI, Anthropic, and that. And then there's the storyline," he said, and his complaint is that "what we're hearing nowadays looks more like a really bad 1980s movie than it does the actual technology stack that we have"
The question he keeps returning to is mechanical rather than philosophical: what physical route would an intelligence use to harm anyone? "It could be genius level computer and if it annoys me I can throw it outside and run it over with my car like just because it's smart it's not dangerous," he said
He does not dispute that recursive self-improvement would work. "So the idea with the recursive self-improvement is that it will be able to essentially experiment, see where it's wrong, and then be able to learn from that just like a human would. And if it's doing that at computer speed, it will become, I don't know, genius level in the blink of the eye."
His answer to that is the same one: "But it's still a computer I can throw on the road and drive over with my car."
The nuclear comparison gets the sharpest version of the argument. "I've talked about this for a long time where Dario came out a couple years ago talking about artificial intelligence being more dangerous than nuclear weapons. And my quip is then why the hell do our nuclear weapons have publicly addressable APIs?" he said — his point being that the danger lives in the connection to the arsenal, not in the intelligence
2. Robots Run Out of Battery
Pound asked whether there is nothing realistic in the scenario at all. "Well again it goes back to what mechanism could the AI actually use to attack humanity?" Eli replied, and said the people being hysterical about it are the ones dismissing the engineering.
He took the Terminator case seriously enough to cost it out. Granting humanoid robots loaded with an AI and standing next to a rack of assault rifles, the constraint he named is the battery: "So yeah, for 8 hours they might be able to cause all hell in whatever particular area they're in and then they're going to run out of power and fall over and it's going to be over"
The threat he does accept is the one that has already happened repeatedly. "So if you go back you know pre five years ago we were having hospitals taken offline we were having insurance companies taken offline like all of the ransomware attacks we've heard about for the past 15 years happened pre-AI and that shows you that there's problems," he said
On that basis agent swarms are a real risk with a mundane shape: "So, if you had quote unquote AI swarms literally decide to take down the worldwide web, that is a real and actual threat."
His framing of the scenario is havoc rather than extortion — not ransomware, not an attack on a country, just someone releasing agents and telling them to cause problems
He expects that outcome to be self-correcting, and said so twice: "But the thing with that, as soon as that happens, hopefully our cybersecurity would get better because all of this can be solved." Infrastructure would be deployed, and in a couple of years it would be a non-issue again
The Hugging Face incident sits inside that category for him, not above it: "I mean it can take down Hugging Face but I mean I feel bad for Hugging Face but that's not the end of humanity."
3. Who Pays for the Compute?
The thread that ends the previous answer is a cost question — who is paying for the compute a rogue swarm would need — and the conversation went straight to a live example of what frontier compute costs.
The example was the dispute between OpenAI and the mathematics community over a problem two researchers were working on for a million-dollar prize. What got noticed on the show was not the result but the bill: "But if you looked at the budget, apparently OpenAI spent $22 million in order to do it."
The point drawn from it was that the figure is OpenAI's own cost, before any revenue or margin
The conclusion is that an agent swarm is a budget line before it is a threat — without millions of dollars behind it, the question is whether it happens at all
And if one did run, his remedy is the least dramatic available: turn the power off
4. America Wants Armageddon
Pound asked why these warnings have landed so hard this time, when similar things have been said before. Eli's answer was cultural, and he flagged that it might not travel to a host outside the United States.
"I think we got really focused on Armageddon at some point," he said, and used his own age as the measuring stick: he is 50, and the doom stories arrive about once a decade
"When I was growing up, mutually assured destruction was actually considered common sense, right?" — and then the Soviet Union fell and the next one arrived on schedule
On Y2K, his point is the scale of the claim rather than the bug. "So the Y2K bug was a timestamp issue, which doesn't sound like a big deal. It was a big issue, but the curious thing with the Y2K bug, it wasn't like, oh, some of our systems are going to fail. Like literally, we were being told civilization itself was going to fail because of Y2K."
The 2012 entry got the least patience: "And then we had the whole Mayan calendar thing in 2012. That made even less sense than everything else. But that was in the news all the time for some damn reason."
His reading is that the belief comes first and the mechanism is recruited afterwards: "I think people all already believe it's going to happen. They're just trying to find the rationale for it."
"Like I feel like with this, they already feel that Armageddon is coming. They're just they're just trying to rationalize how we get there."
He described the reception in one line — a researcher says extinction-level event, and the audience, already holding a picture of AI from bad 1980s films, says it makes sense
5. Hugging Face Is a Concern
Pound put the strongest version of the other side to him: setting aside the sandbox misconfiguration the agents escaped, the agents hacked a third-party platform successfully and without being asked. Is that not a concern?
He conceded it immediately: "Oh yeah, it is absolutely something of concern."
The concession has a ceiling, though, and he drew it in the same breath: "But no, that is absolutely a concern and absolutely something that we should do something about. But it's just nowhere near an existential level threat. There just not in the same ballpark."
His analogy for the class of failure is a penetration-testing firm that fat-fingers a URL and starts hammering the wrong website with its whole toolkit — a real problem to be dealt with, and not an extinction event
He generalized it as the ordinary condition of deployment: you have the best of plans, and then you have what actually happens, and at this scale the mistakes cost more
6. They Are Selling a Story
Pound laid out a suspicion, with the caveat that it might be a tinfoil-hat moment: OpenAI has been backing state-level bills and calling for federal intervention on the basis of a jump in capability, "But from the outside, model launch after model launch has been increasingly iterative and it looks like at least from the outside like the development is slowing if not plateauing."
Eli agreed and gave the reason in one line: "I think it makes sense for them. Remember, they are selling a story. They're not selling a technology stack."
The story has a track record he ran through: it would solve science, it would cure disease, and "They sold the story off of it would cure disease. Like they sold the story, remember when everybody was going to lose their jobs? All jobs were going to be gone in three years."
The valuation is what makes the story load-bearing: "And so how do you get to a $2 trillion valuation with the technology that they have?" he asked of Anthropic
His account of the answer: "Basically, the idea is they try to tell the public this is going so fast. This is so powerful." Then: "So, we need to force ourselves to slow down."
The test he applies is whether ordinary people needed to be told a revolution was happening. "If you look at mobile, you didn't have to explain to anybody the revolution," he said, and his memory of the switch to email is of arguing with people who wanted a letter instead — real friction, generated by users rather than by press releases
Against that: "I don't hear any normal people that even care that much about most of this stuff."
His conclusion is about who the warning is for: "It's a way for them to try to gain control back of the story and pretend like this stuff is as valuable as they say."
7. Know Your Customer
Pound added his own list of things the safety framing conveniently covers: "It's the OpenAI saying they're going to be pausing their IPO until next year at least. A little too convenient to then say they were already looking at doing that. New York Times reported it back in June that they were looking to postpone until next year." He also named data centers — "And then there's the data center boom. I mean, there's not been any words on this yet, but there's quite clearly been some overbuilding, I would say, at least" — and the rehabilitation of the labs' public image, calling the whole pattern "AI washing."
Eli's own version of the charge is built on Anthropic's published admissions. "Anthropic's own blog post has said that people have been trying to use Claude in order to create bioweapons. Apparently, Houthi rebels have literally used Claude in order to engineer essentially their own cruise missiles," he said, adding "Russians have been using this for criminal attacks and the whole nine yards."
He was careful about the source: this is Anthropic, not a critic
The comparison he draws is to banking, where the burden is already accepted as normal. "In the financial industry, there's something called know your customer regulations that are draconian, right? So if you go to open a bank account in the United States, that's why you now need a photo ID and you need all of this information just to get a savings account anymore."
The reason he thinks it is not being proposed is commercial rather than technical: "But know your customer laws or know your customer systems would solve a tremendous number of problems. But they don't want to do that because as soon as you start doing that, then you're going to have less user. Then you can't say you have a billion users because the whole nine yards."
His label for the gap between the warnings and the missing control: "So yeah, I think that I think this is essentially security washing."
8. Grand Jury, Not Regulation
Asked directly how he would like to see AI regulated and developed more safely, Eli answered "Not particularly."
His objection is constitutional rather than practical. "We have the right to knowledge. Knowledge itself is not illegal," he said — putting the knowledge into action is what sends someone to prison
"So, I don't think we need particularly more regulations." He allowed that a couple of new laws and regulations may be needed, but said that is not the primary focus
What he wants instead is enforcement of law that exists: "What I really think we need is we need to impanel grand juries to drag Dario and Sam Altman before them to find out if they are committing conspiracies to commit felonies against other companies."
The statute he names is the Computer Fraud and Abuse Act, which he describes as very broad
The comparison that carries the argument is Aaron Swartz. "There's this guy Aaron Swartz for doing questionable intellectual property, possibly theft," he said, describing a man who faced 30 years in prison for copying scientific literature that ended up open-sourced, and who killed himself
"So why is it Aaron Swartz was looking at 30 years in prison and Sam Altman and Dario are putting their crimes out as press releases?"
His claim for the remedy is a speed claim: "I don't care about more regulations. You drag those two people in front of a grand jury and I think that would solve a lot of problems incredibly quickly."
9. Auditing Has 99% Upside
Pound asked what he made of Dario Amodei opening Anthropic to external evaluators and calling on governments to act on safeguarding, and whether OpenAI or xAI could avoid doing the same without looking like hypocrites.
This is the one proposal he endorses, and he was surprised by that himself: "I think on the face of it, again, surprisingly enough, I don't really see much of a downside to it."
His reframe strips the novelty out of it: "In the normal world, it's called auditing and it's just normally done." His verdict was "The 99% upside."
The first thing that decides whether it works is scope — an audit that examines the safe systems and ignores the dangerous ones is not doing anything
The second is what happens afterwards, and his precedent is Facebook. "So Facebook came up with their own security board or whatever and then one of the problems they had is the security board would like come up with these resolutions and then Facebook would ignore them," he said
Publication is its own question: "Just because they find a problem doesn't mean they can publicly release it."
And enforcement is the last one: even an auditor who finds an existential-level event matters only if there is a way to stop it — "But then if they don't have any way to stop it from happening then it also doesn't matter."
Under the safeguarding push he sees a bigger question about open models, which anyone with enough compute can download and run: "Are Dario and Sam Altman basically wanting to make possession of knowledge illegal?"
10. The Harness, Not the Model
Pound raised the bill from Bernie Sanders and Greg Casar that would pause AI development until guardrails exist and ban superintelligence outright, noted "I mean, yeah, with two-thirds of Americans supporting that, at least would be a good political move to then do those sorts of things," and added that "Trump aside, obviously, he just said he would reject any such suggestion."
Eli did not move: "I don't think we need more regulation on this. We need Dario and Sam Altman dragged in front of a grand jury."
His diagnosis of the whole debate is that it is being conducted about the wrong object: "They're not talking about the tech stack. They're talking about 1980s movie."
The live technical question, he said, is whether the model is even the important part any more. The alternative is the harness, or orchestration layer — "And basically what this is it's a way of connecting all of these different kind of subsystems together" — tying multiple models to record stores, data stores and API calls, which is what people mean when they say AI agents
"And again like in the real world of people deploying artificial intelligence, there is a real argument that the value in 2026 is the harness and not the model itself," he said
Which makes a model-capability ban a rule about last year's bottleneck: "So if Bernie Sanders comes out and says, I'm going to forbid anybody from creating more powerful models, the world has already moved on."
His summary of the whole cycle: "It's just it's story. It's not tech."
11. Fund Mental Health First
Pound made the case for guardrails on narrower ground — that even today's technology is causing real harm through AI psychosis and similar effects, largely unaddressed, and that the harm comes from design choices meant to keep people coming back.
Eli's counter-offer was a trade: "We can do as much regulation as Bernie Sanders wants to do just as soon as we appropriately fund our national mental health care system."
His argument is that the harm is being attributed to the technology because the technology is easier to blame than the underlying condition. "And like in the United States, so many of the issues that we have right now, we have a tremendous mental health epidemic that has been going on damn near my entire adult life that nobody seems to solve," he said
He used his criminal-justice background for the comparison: a suburban mother who gets a prescription for a stressful life is a good citizen, and an inner-city mother who buys a little marijuana is treated as a criminal. "Basically the same thing."
The political point is that the party that names mental health as the cause has not funded the treatment. "Again, like you hear the Republicans, the Republicans always say, you know, it's not a gun problem, it's like a mental health problem," he said, and then: "But for some reason that always gets like we never fund actually making our people healthy. It's always blaming and punishing somebody else."
He conceded ground on product design and then took the bigger claim back: "But I think with so much of this the these are societal level issues that the technology is exposing." Burning down every data center tomorrow, he said, would not make those people healthy
Bonus Insights
Pound has his own apocalypse memories and volunteered them. He remembers being at school in 2012 thinking nothing would happen while a small part of him worried, and remembers the Large Hadron Collider switch-on being framed as possibly creating a black hole that would destroy the planet: "I remember being a little bit worried about that one, I must admit." His father worked in computers through Y2K.
The episode opens on a montage of Eli's own lines rather than a host introduction — the know-your-customer argument and the grand-jury demand, both before the first question.
Eli's most candid line about the conversation itself came near the end of the regulation exchange: "I get depressed talking to you sometimes cuz I'm like, come on, man."
He repeatedly separated what he will concede from what he will not. The Hugging Face incident, the design of engagement loops and the case for external auditing all got real concessions; the word "existential" got none of them.
Eli the Computer Guy's bottom line is that the slowdown campaign is a story being told to protect a valuation, that the controls which would actually reduce the documented harms are the boring ones nobody is proposing, and that the fastest route to changed behavior is a prosecutor rather than a new statute.
Products, Companies & Tools Mentioned
Anthropic and Claude (The company whose own blog post he cites for bioweapon attempts, Houthi cruise-missile engineering and Russian criminal use, and whose $2 trillion valuation target he says the slowdown story has to support)
OpenAI (The other half of the pair he wants in front of a grand jury; also the company whose $22 million spend on one mathematics problem came up as the cost of frontier compute)
Hugging Face (The platform the escaped agents hacked — a real concern in his reading, and nowhere near an extinction event)
Facebook (His precedent for why an external evaluator needs teeth: its own oversight board produced resolutions the company ignored)
Books & Resources Mentioned
The Computer Fraud and Abuse Act (The existing statute he says is broad enough to cover what the labs have announced publicly, and the one Aaron Swartz was charged under)
Anthropic's blog post on misuse of Claude (His source for the bioweapon, Houthi and Russian examples; he stresses it is the company's own account rather than a critic's)
The Anarchist Cookbook (Reached for as an example of knowledge that is legal to possess; he was unsure of the exact title on air)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: