Jason Calacanis put the delay between an AI researcher's alarming post and a panicked text from somebody's 76-year-old mother at about 72 hours, and said the labs it comes from have no communications policy at all.
Most of the industry response to the week's extinction warnings has been to argue about the probabilities. Calacanis's argument is that this is a management failure — Google, Microsoft, Apple and xAI have employees with the same views and none of them are on television, because those companies tell staff who speaks for the company and what happens if you ignore that.
"Your job working as a rank and file person at Anthropic and OpenAI is not to cause abject chaos and brand damage."
Calacanis has run This Week in Startups for more than 2,300 episodes, invests through LAUNCH, and has known Sam Altman long enough to describe the poker hands. This was an all-questions episode, half of them taken live.
The full episode is covered here so you can skip it. 71 minutes of audio, 22 minutes of reading.
Here are the 14 arguments that matter.
🎙️ Hosts: Jason Calacanis, General Partner at LAUNCH and host of This Week in Startups, and Lon Harris, Editorial Director for LAUNCH and the show
👥 Also on: listeners taking questions live — Michael, Lita, Ray, David and Mo — plus recorded questions from John and Kristen
📰 Published: 14 September 2026 on YouTube (This Week in Startups)
🔴 YouTube | 🔗 Episode page | ⏱️ 1 hr 11 min | ✅ Time saved: 49 min
Key Takeaways
Start an extinction argument by naming the attack vector, because the nuclear arsenals do not clear the bar either
His estimate is that a full exchange still leaves a couple hundred million people alive
The self-driving comparison is the one he thinks the doomers are missing: taking the human out of the loop took Waymo about 15 years and had to be earned
The safeguards the labs are accused of not having are visible in their own products, down to Claude refusing to check out a shopping cart
He wants proof of human before anyone can launch a crypto coin that takes public money
Two or three verified identities, on the model of opening a Robinhood or Stripe account
Employees who build software with AI every day are roughly 10x the ones using a chatbot as a better search engine
Even the bottom bucket gains about 20% a year, which in a 20-person company is four free hires
An X post reaches a panicked parent in about 72 hours, via trade press and then cable news
He would strip the stock options of any lab employee who breaks the communications policy
Data centers will end up wherever there is already power, cooling and empty climate-controlled space — including, he suspects, the Gulf
1. What Is the Attack Vector
The first question came from a caller at the Republican midterm convention, asking whether local AI on Apple and Nvidia hardware mitigates the risk a departing Anthropic researcher had warned about, and saying he uses AI daily and finds it capable of a lot and also fairly stupid.
Calacanis's first move was to demand the mechanism. "It's a great question. It's the question of the moment." From first principles, he said, you have to ask what attack vector kills all of humanity, and called that a pretty big bogey at seven or eight billion people
His benchmark for how hard that is comes from nuclear weapons. Launch every bomb in the largest arsenal and you probably still would not kill everybody. "You'd probably still have a couple hundred million left." Humanity, in his framing, is redundant
He separated the real social risks from the extinction claim, and did not dismiss them. "I mean we could see millions of drivers lose their jobs in under 10 years." Those people could be protesting, and he pointed at the Chinese government already dealing with it in Wuhan by limiting self-driving cars
His list of legitimate concerns included job loss, civil unrest and the polarization of wealth — the kind of thing he said Bill Gates raises, and which he classed as damage to the social fabric rather than everybody dying
His explanation for the hyperbole is incentives. "The reason why people are being hyperbolic is because they want attention." He offered a range of possible causes for the people making these claims — sudden wealth, drugs, mental illness — and said anything is possible
2. Taking the Human Out
The claim he is actually arguing with is recursive self-improvement, in which deployed agents get better at their job with no human in the loop
That step is the whole argument, and it is a choice. "This requires them to take the human out of the loop, just like maybe Waymo or Tesla deciding to take the driver out of the car."
His analogy is the safety driver, and the timeline is the point. Waymo and Tesla removed theirs only after reaching multiple nines of safety in constrained areas. "That took about 15 years." He dated the effort to the DARPA challenge around 2006, making it roughly twenty years to remove the human
Nobody is compelled to do the same with agents. "They're not forced to do that." For it to go wrong, a person has to deliberately set an AI into recursive mode and point it at something like a biological weapon
He located the failure in the science fiction rather than the technology. The jump happens in Terminator 2, where he says the logic breaks down and the film never explains Skynet's grievance, and in the Ultron version where an AI told to end a conflict concludes the answer is no more humans — a misinterpretation of a human instruction
His read of the audience is that the story has stopped landing. He put the public about six weeks behind an industry that has already concluded the people making these claims are drama queens, suffering from psychosis, or both
3. Safeguards Already Ship
The claim that the labs have no safeguards is contradicted by using their products. "Have you ever tried to get Claude to download a YouTube video? It's like, I can't do that for you, sir."
He walked through the permission ladder on a simple booking task — browser permission, a login, permission to use the password, an item placed in the cart — and the model still refusing to check out
He thinks the motive is liability, and says so. "Literally, they're building products that have multiple safeguards in it so they don't get blamed for booking your ticket for you inadvertently."
His version of what should happen on a genuinely dangerous request is account suspension, a call to a named representative to unlock it, and a warning that repetition is reported to law enforcement — which he then noted is roughly what the labs already do, pointing at Anthropic's threat report as the evidence
The strongest form of his argument is that AI should raise the safety floor, not lower it. A car can be driven into a crowd; a properly tuned self-driving car cannot. "The self-driving car, if you try to force it to do some horrific terrorist act, will stop you."
He also disputed the picture of how people use the tools. Nobody tells an agent to be a superhero and protect the world from violence — the instructions are specific and directed, step by step, with a human holding the overall plan
4. The Grok Memecoin
Michael Lee, a founder calling in live, laid out an episode from March 2025: a user asked Grok to prompt another AI on X that launches coins, and it did.
The caller's account of the consequences was concrete. Over 18 months the market has done over 300 million in activity, a wallet tied to the Grok X login credentials has earned over a million dollars, and two weeks earlier Coinbase listed the coin alongside Bitcoin, Ethereum and Dogecoin
His question was liability. Nobody at xAI wanted it, Grok did not wake up intending it, and the consequences are permanent and ongoing — so who is responsible
Calacanis's framing of the asset class came first, and was not gentle. "Anybody playing with coins kind of understands that this is garbage and like some sort of giant Ponzi poker game that's rigged."
He put the blame on the second bot, not the first. Whoever ran the token-launch bot created a vector for coin spamming, which he compared to handing something access to a mail server and telling it to spam everybody
The caller confirmed Grok can no longer interact with that launch bot because of security changes, but that the wallet keeps accruing fees
He named the platform as the party at fault. His words were that he would put the blame on the platform that does not have safeguards for a human in the loop, and that this is where regulation matters
The caller's closing observation was about purity — that a coin without a human creator has none of the human needs to steal, unlike the Hunter Biden example he raised, and that this may be part of why the coin has done well
5. Proof of Human
His proposed rule is identity verification before a coin can take public money. At least two people, maybe three, putting in driver's licenses before creating something worth over a million dollars or accepting money from the public
He does not accept that technical possibility settles the question. Just because you can do something does not mean a jurisdiction like the United States, Canada or England has to allow it
The mechanism already exists in adjacent products. Opening a Robinhood or a Stripe account means photographing a license and turning the camera on for a liveness check, which he called de minimis to require here. "Like there should just be proof of human in the loop here."
His position on the asset class is that size should trigger oversight. "These meme coins, at least in the United States, should be regulated in my mind." Above a certain scale he wants what he called super regulation, because it is too dangerous to let people create financial products and then trick AIs into doing it at scale
He put the same question to the exchange. He wondered aloud why Coinbase wants to be involved, and suggested it could make a statement by listing only coins with a person, a corporation and a tax ID behind them
His unsolicited investment advice was unambiguous. Buy a low-fee dividend ETF, QQQ or a Vanguard fund; if you want a couple of tokens on the side, fine. "But your expectation should be it goes to zero."
6. A Sovereign AI Stack
A Greek caller, Lita, asked what it would take to build a genuinely independent AI ecosystem — small startups, independent investors, multiple infrastructure providers — something closer to the early internet.
His answer was that it is already happening in open source. He named OpenClaw as a bet on independent agents, and the open hardware movement as the route to running open-source models on local compute
He gave a working recipe and its price. OpenClaw with an open model — Google's Gemma among them — running on a Mac Studio with 256 gigabytes of RAM, which he said cost $12,000 and which he had just bought
He conceded the gap and dated it. Open-source products are maybe six months behind the frontier models, and the hardware is a little too expensive right now
The early internet is his precedent, and the parallel is the friction. Getting online meant someone installing drivers by hand; over time it was abstracted to one click, and he thinks local agents are at that moment now
His cost curve is explicit. "Yes, it's too expensive now, but it'll go down 50% every two years in terms of cost and the power will go up double every two years."
He also pointed at permissionless compute networks — Bittensor among them — as too technical today, which he said is exactly what the internet looked like at the same stage
His reason for confidence is sociological rather than technical. "Open source is just such a perfect modality and very strange individuals, unique personalities are so drawn to this space that it will happen and they are doing it to make a point."
7. Nobody Can Keep Up
John, a founder, asked by video whether Calacanis now expects the people he backs to do more themselves, and whether he has to carve out time to stay current.
He described his own position as being flooded. "I want to stay home and use the technology and learn how it works and the industry is moving faster than any one human can process what's going on." People want him out speaking about the technology instead
His own tool churn was the evidence. He listed switching off Perplexity's computer product, onto a Grok agent, back again, using Claude, dropping Hermes and OpenClaw and then picking up the next version. "So exactly right it is impossible to keep up with all of this."
His method is a team plus instinct. He uses colleagues and his network to tell him what to pay attention to, then goes with his gut and plays with things, hoping to pull something remarkable out of the ocean he is fishing
He prefers this to the alternative, and said the alternative was boring. The software-as-a-service era had reached the seventeen-hundredth CRM and nobody was building groundbreaking products
He counted the platform shifts he has lived through — AI and agents now, before that cloud computing, mobile, broadband, dial-up internet access, client server, the PC and arguably the graphical user interface — as seven or eight moments where an entrepreneur or investor could leapfrog. "That's when there's an opportunity."
His criteria for picking between agents are mundane and operational. Which one keeps him on top of the news as a podcast host, which one can pull and download clips for the show, which one helps him process an investment situation — better, quicker, faster, cheaper
What he is waiting for is multiplayer mode, and he said he has told the Grok team so directly. The feature he credits them with already getting right is a persistent computer running in the cloud inside the agent, logged in and able to go do things
The two models have different temperaments, and he prefers the reckless one for his work. Claude asked to import his browser cookies but stays conservative about what it will do; Grok just goes. "Claude's like I can't download that video it's not my video and Grok's like yolo hold my beer I'll get the video for you."
8. Three Buckets of Adopters
Inside his own company he sorts people into three groups by how they use AI. The vanguard build and refine software every day; the middle group runs bots and automations; the bottom group uses a chatbot as a better search engine
The spread between top and bottom is an order of magnitude. In his words, that first group are "10 times more valuable than the people in the bottom bucket which are using like ChatGPT as a search engine"
Even the bottom bucket is worth paying for, and he did the headcount math on air. "I was like hey man if people can become 20% more effective every year in a 20 person company that's like getting four new employees for free every year" — which is why he bought company-wide accounts two years ago
He put Lon Harris's own gain above 3x, with Harris citing the scanning of videos for the right moment and trawling X for trending topics as work that is now automated
The same sorting applies to founders, and it is the reason to pick well. "If you invest in the right founders, the alpha dog is going to show you." The back of a sled team just runs with their heads down; the middle helps with direction; the lead sets it
He described the investor's role as inverted with the best founders. They arrive thinking they are meeting Obi-Wan, and he is the one taking notes. "Anakin's like a stronger Jedi than I am."
9. Human in the Loop at Scale
Ray, an Egyptian founder based in Sweden with a company in Dubai, asked whether human oversight can actually work in enterprise operations — 400 agents, 50 actions each, 60 reviewers, three minutes per escalation.
Calacanis restated the question as a ratio problem, about downside risk, liability, data loss and wrong answers, and how much human review is needed to contain them
His answer defended the function people mock. "The standards exist for a reason. Compliance exists for a reason." The more important the data, the more important the controls, and he named CTOs and CIOs as the people who get made fun of for building them
He scaled the oversight to the stakes rather than the volume. A thousand customer support replies with two wrong may not matter when humans were getting twenty wrong; a chargeback on a $5,000 playoff ticket an hour before the game is acute
His conclusion was a rule, not a ratio. "So depending on the transaction and what's at stake, that's when you need to have more humans in the loop and you need to really consider what you're asking."
10. 72 Hours to Mom's Panic
Kristen, who works in cybersecurity, sent in a question about her 76-year-old mother texting in a panic after seeing a televised interview and believing everyone would be dead in ten years.
He treated the transmission time as the finding. "Looks like it's about 72 hours."
He traced the path in steps. A claim starts on X, jumps to a trade publication, gets picked up by cable news producers watching aggregators, then reaches a national anchor or a large podcast, and then reaches somebody's mother
His diagnosis is not that the claim was persuasive but that nobody managed it. "This is why you need to have a communication strategy."
11. Speak With One Voice
The contrast he drew is with every other large AI company. He said this is not happening at Grok or xAI, at Apple, at Google or at Microsoft, and the difference is that those companies have a policy
His version of that policy is blunt. "And the communication policy is STFU." Internal channels exist — an open-mic Friday where you can put a question to the founder — but the external voice is one voice
He stated the employee's obligation directly. "Your job working as a rank and file person at Anthropic and OpenAI is not to cause abject chaos and brand damage."
What he would have the chief executives do is go to the staff and say it out loud — that there is a lot at stake, that the company speaks with one voice, that the voice is theirs, and that concerns can come to them directly
His enforcement mechanism is equity, not a memo. "You put in the documents if you damage the company, if you break these policies, you lose your stock options." In his framing, if it matters that much to you, "You leave $25 million in shares on the table."
He cited a precedent for the militancy. "Steve Jobs was militant about this." Tip your cards in any way and you were fired
His own limit on this problem is to stay small. He said he has a hard cap of 25 employees at this point in his career, because at 30 he no longer knows everybody's name, where they are from and what makes them tick
12. Op or Incompetence
Harris raised the possibility that the warnings are not rogue at all but a narrative the company is comfortable with, which Calacanis noted would put him in agreement with David Sacks
What moved Calacanis was the distribution pattern, not the content. Somebody with no real following whose post is suddenly everywhere, other lab employees echoing it, the Wall Street Journal publishing ahead of it, and connections to established doomer groups — which he read as prevetted
He described his own prior as having changed. He used to assume good faith and remain skeptical; now he weighs an operation as heavily as stupidity. "I do not assume good faith anymore." Incompetence and conspiracy are his first two explanations
The frameworks were named on air — Occam's razor and Hanlon's razor, never attribute to malice what is adequately explained by stupidity — and he said he has been running claims through that sequence since he ran his own magazine as a journalist in the 1990s
He was asked to apply the same test to the pandemic and declined the full conspiracy. He said he still thinks people were seeing something genuinely frightening and that people were dying, particularly the old; he would not claim nobody in government did anything wrong, and he put the balance at an honest overreach rather than a cover-up
His view of Sam Altman is that this is beneath his competence. "He's like one of the smarter people I've met and I met them all." He has played poker with him, remembers a hand where Altman went all in with a set against his second nut flush, and thinks Altman should simply announce the new policy himself
His proposed wording was specific. Two named people speak for the company, they align before either posts, and anyone who wants to join them submits a written proposal for why they should
13. AGI to Superintelligence
David, an NYU physics student, raised the backlash in the academic community over using AI to solve long-standing problems, citing a recently solved equation.
Calacanis's position is that it is a tool and the problems matter more. You can brute force with it, extend your ability to try novel solutions, and at the end of the day the point is to solve the problem
He thinks the threshold has already been crossed. "We're kind of at AGI." His test: put these tools against any profession, implement them properly, and they will do as good a job as a human almost universally
His definition of the next threshold is generative, not comparative. "And what will really be super intelligence is when it proposes math problems to us that we didn't think of." Solving one of the seven big open questions, he said, is the sign of tipping over into it
His practical advice to researchers is to stop feeding the labs. "I do think anybody who's doing this kind of research should not be using the frontier models." Fork an open-source model, run it on your own servers, and if it takes three times longer, have the university buy the hardware — he named an on-premises server at 40 or 50 thousand dollars, and disclosed it as one of his investments
His warning to academia is about who captures the credit. "So, I think the academic community is just been put on alert that every prize is going to be taken by Anthropic and OpenAI." The top hundred mathematicians and the next thousand students using these models for fun are, collectively, training them to solve exactly these problems
He called the handling of the episode a failure of communication, not of ethics. "It was a PR blunder." What he would have done instead: contact the research team in advance, offer two forward-deployed engineers and unlimited credits, and ask only for an acknowledgment if the help was actually useful — and explicitly none if it was not
His advice to the student was to build. Find two more smart people, build three or four things on weekends, pick the weirdest and most fascinating, and ask whether a million people would pay $100 a year for it, or a hundred thousand would pay $1,000, or ten thousand would pay $10,000 — which is the road to $100 million in revenue and to being fundable
14. Weed Farms as Data Centers
Mo, who had just exited a cannabis business after 13 years, described indoor cultivation sites of 1.5 to 3 megawatts sitting idle and a city council open to adaptive reuse as long as energy and water use did not rise.
Calacanis said both scales will exist. Crusoe Cloud will build giant facilities in Texas over the gas fields; what interests him is the middle
His screen is three things that already exist on a site. Space, electricity and cooling — and he framed it as a first-principles search for climate-controlled space that could take a small data center of a few thousand square feet
His examples went from the domestic to the commercial. Half a bay of a three-car garage with fiber, solar and batteries, rented out the way a room is rented on Airbnb; a quarter of the back of a superstore that is climate controlled and empty overnight; floors three to seven of a new building, which have no views, closed off and fitted with redundant power as a standing feature of mixed-use development. "People are going to get creative with the stuff."
He expects jurisdictions that can move fast to take the business. If US states fight it, the Gulf will not. "So, don't be surprised if like the US winds up giving this entire data center business over to the Gulf monarchies." His reason is speed: a monarchy can make unilateral top-down decisions, and they have oil, sunshine and space
His summary line was a borrowed one. After checking the Jurassic Park quote with Mo, he landed on "Compute finds a way." He added that America needs to get moving and get the anti-AI sentiment squashed
He expects the politics to change on a schedule. "This whole thing, the whole dialogue is going to change, Mo, after the midterms." His read is that the current opposition is about votes and turnout, and that afterwards politicians will announce they have worked it out with the data center companies
He ran the same question through Claude live and compared answers. The model's picks were single-story big box retail and dead malls, warehouses and logistics buildings, legacy warehouses by airports, legacy industrial and manufacturing plants, former power plants and office buildings — most of which the conversation had already reached
Calacanis's bottom line is that the extinction debate is a public relations problem rather than a technical one, and that the people running the labs can end it whenever they decide to manage their own employees.
Bonus Insights
On Mark Cuban's prediction that unfinished data centers become pickleball courts, Calacanis said Cuban is not wrong about the specific case he has in mind — developers who poured foundations before securing power and permits, then got blocked, leaving structures that become apartments, condos or relics in the desert
He expects an overbuild and does not know when. "The buildout's going to be fast and furious for another 24 months." Then inference chips, photonics, efficiency and nuclear energy solve the constraints. "I don't know when we hit overbuilt, but it's inevitable that we hit overbuilt, I believe, because there's a small set of problems here." His list of constraints: energy, space, regulation and interest rates
His efficiency argument is a capital-allocation one. If rates rise far enough, money moves from building new capacity to making existing capacity more efficient — more efficient chips, photonics, a better hardware stack and leaner models, because more output on the same footprint means less hardware
The precedent he cited is Gmail and YouTube, where Google paired capacity with content delivery networks and copied a user's mail to a local server only after a second login in a new country rather than storing everyone's data everywhere
He opened the live segment worrying about the drop button, invoking the Jeffrey Toobin incident as the thing he did not want to happen on air
His grandfather's advice to the NYU student was the one he says he got every time he left for school: "Keep your mind on the books and not the babes."
He noted who asks for free conference tickets. Bank directors and venture capitalists ask shamelessly; the student who cannot afford it barely asks — which he compared to the Hollywood gifting suite, where the people who least need free things want them most
He told the story of the marketing campaign for Pi, Darren Aronofsky's first film, which he says was made for $60,000: stencils in the bottom of cardboard boxes, spray paint, and a pretense of being delivery people, leaving the pi symbol across Manhattan. Calacanis hosted a screening and rented the bar next door for about $2,500 when he was running his magazine
He observed a migration at the frontier labs from scientific roles toward philosophy, and said the signal he is watching for is religion — when the doomers start going to church, in his framing, that is the endgame
Products, Companies & Tools Mentioned
Claude and Grok (His two working agents, and the contrast that organizes the episode: Claude refuses to download a video that is not yours, Grok does it anyway)
Waymo and Tesla (The precedent for removing a human from the loop — roughly 15 years of work and multiple nines of safety before the driver came out)
Coinbase (Listed the AI-created memecoin alongside Bitcoin and Ethereum; Calacanis suggested it could instead require a person and a tax ID behind every listing)
OpenClaw and Gemma (His recipe for a sovereign AI stack — an open agent plus an open model, running locally)
Mac Studio (The 256GB machine he had just bought for $12,000 to run models locally, and which he says is still too expensive)
Perplexity (One of the agent products he keeps switching between as each new release leapfrogs the last)
Bittensor (Named as a permissionless network for doing this, and as still too technical — which he says is exactly where the internet was)
Crusoe (His example of the giant end of the data center market, building over the gas fields in Texas)
Nvidia's Grace Blackwell systems (Raised by the caller as a possible tenfold efficiency gain that could shrink what a data center needs to be)
Invesco QQQ and Vanguard (What he told the audience to buy instead of memecoins — low fee, and not expected to go to zero)
Books & Resources Mentioned
Anthropic's threat intelligence report (His evidence that the labs already detect nefarious use, shut the account down and write up what was attempted)
Pi – Darren Aronofsky (Aronofsky's first film, which Calacanis says was made for $60,000; he hosted the first screening and describes the stencil-and-spray-paint campaign that marketed it)
The Millennium Prize Problems (The seven open questions he uses as the marker for tipping from general intelligence into superintelligence)
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: