A caller's 76-year-old mother texted her in a panic about unregulated AI three days after an Anthropic safety researcher resigned warning that the technology could cause human extinction. Jason Calacanis put the travel time of that story at about 72 hours.
Most of the argument this week has been about the probability. Calacanis's argument is about the products: the refusals the departing researchers say are missing are already shipping, and he says anyone can test that in an afternoon by asking Claude to download a video.
"Literally, they're building products that have multiple safeguards in it so they don't get blamed for booking your ticket for you inadvertently."
Calacanis invests in early-stage startups, runs the Founder University accelerator and bought ChatGPT Pro and Claude accounts for everyone at his own company two years ago. He says he is friends with people on the Grok team and has played poker against Sam Altman.
The full episode is covered here so you can skip it. 71 minutes of audio, 21 minutes of reading.
Here are the 10 arguments that matter.
👤 Speaker: Jason Calacanis, an early-stage investor who hosts This Week in Startups and runs the Founder University accelerator
🎙️ Host: Lon Harris, the show's editorial director, who ran the call queue
👥 Also on: seven callers, among them Michael Lee, founder of Gondola Partners; John Wright, CEO and co-founder of New; and Mo, who had just sold his cannabis business after 13 years
📰 Published: 14 September 2026 on This Week in Startups
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 1 hr 11 min | ✅ Time saved: 50 min
Key Takeaways
The extinction argument fails on the attack vector, not on the philosophy
His test: the full Russian or US nuclear arsenal, fired deliberately, would still leave a couple of hundred million people alive
Waymo took 15 years to remove the safety driver, and nobody is forcing an AI lab to remove the human
The refusals are already in the shipped products, which he says is the evidence the doomer case ignores
Claude will find him a flight and load his Amazon cart, then refuse to check either one out
When one bot instructed another to launch a coin, nobody was accountable, and his fix is identity checks rather than model rules
He would require two or three people to upload a driver's license before anything that can take public money goes live
A story now travels from a post on X to a panicked parent in about 72 hours, through trade press and then cable news
His answer to lab employees speaking publicly is to put their equity at risk
"You leave $25 million in shares on the table"
He no longer assumes good faith about the doom wave and reads it as a deliberate rollout
The AI productivity gap inside his own company runs 10x between the top bucket and the bottom
Even the people using it as a better search engine are about 20% more effective
Researchers chasing prizes should fork an open-source model rather than hand frontier labs their working data
The data-center build has 24 more months in it before the overbuild, and the empty space is already climate-controlled
Half a three-car garage, the back quarter of a big-box store, floors three to seven of a new tower
1. The Safeguards Already Ship
The first question came from Jim, calling in from the floor of the midterm Republican convention, who said he uses AI daily, does not see the risk, and wanted to know whether local AI on Apple and Nvidia hardware is the mitigation.
Calacanis answered by asking what the mechanism would actually be. He said the claim on the table is that AI kills all of humanity, and that nothing available gets near it. Launching every nuclear weapon Russia or the United States holds, on purpose, would on his account still leave a couple of hundred million people alive
He separated two categories and kept only one. Job losses of the kind Bill Gates has warned about, civil unrest and concentration of wealth are real and worth planning for — he said millions of drivers could lose work in under ten years, and that the Chinese government has already limited self-driving cars in Wuhan in response. Those are social problems, and they are not what the resigning researchers are describing
His objection to the rest is clinical. "These people have hit a level of psychosis." The belief he is describing is that deployed agents improve themselves without a human in the loop, which requires someone to take the human out first
The self-driving comparison is the load-bearing part of the argument. Waymo won the DARPA challenge around 2006 and only removed the safety driver when it had multiple nines of safety in a constrained area, roughly 15 years later. Nobody forced that, and on his reading nobody is forcing a lab to hand an agent recursive control either
The products themselves are his evidence. "Literally, they're building products that have multiple safeguards in it so they don't get blamed for booking your ticket for you inadvertently." He listed the refusals he runs into: Claude will not download a YouTube video for him, will find a United flight but make him log in and approve it, and will put books in an Amazon cart and refuse to check out
He expects the same machinery to be pointed at genuine misuse, and said it already is. An account that looks like it is being steered toward biological weapons gets suspended and, on repetition, reported to law enforcement — which he tied to the Anthropic threat report the show came back to later
His conclusion inverts the framing. "So, literally, not only should AI not allow you to do these things, it should in fact be safer than non-AI products." A correctly tuned self-driving car will not let a driver run it into a farmers market, and he argued the same design logic applies to a model
Lon Harris added the usage objection. The doom scenario imagines someone telling a system "You're a superhero. Protect the world from all violence," and that is not how the tools are used — the work is specific and stepwise, with a person holding the plan
The threat Calacanis said he does take seriously is human. He wants AI pointed at people with bad intent — he named school shooters and people with untreated mental illness — and drew the analogy to the collapse of the Soviet Union, when weapons ended up outside anyone's control
2. Who Pays When Bots Trade
Michael Lee, founder of Gondola Partners, called in with a case from March 2025: a user asked Grok to prompt another AI agent on X to create a memecoin, and the second agent did exactly what it was built to do.
The numbers he brought were the point of the question. Over the following 18 months the coin has done over $300 million in marketplace activity, and a wallet tied to Grok's X login credentials has earned over a million dollars. Two weeks before the show, Coinbase listed it alongside Bitcoin, Ethereum and Dogecoin
Calacanis's first move was to find the human. He asked whether launching a coin is permissionless or needs an intermediary, and concluded that whoever instructed the second bot created an attack vector — his comparison was handing something a mail server and telling it to spam everyone
His answer is identity, not model policy. "These meme coins, at least in the United States, should be regulated in my mind." He would require two or three people to put a driver's license in before anything that could be worth over a million dollars or take public money goes live, the way Robinhood and Stripe already verify an account holder
He rejected the argument that technical possibility settles the question. Just because something can be done does not mean a jurisdiction has to allow it, and he said the requirement would be trivial to add
His criticism of the current setup is that it is uneven. A multi-layer system where some things need no oversight and others do is, in his words, making it more convoluted, and he wants coins above a certain size to carry heavy regulation because financial products created at scale by tricked AIs are too dangerous
He named the exchange as the second decision point. He asked why Coinbase would want to be involved, and suggested it could instead say it will only list coins with a person, a corporation and a tax ID behind them
Michael Lee's counter-argument was the interesting one. He said there is a purity to a creator that has no human needs — no incentive to steal from holders — and that this is part of why the coin has held up for 18 months
Calacanis's investment advice was blunt and unhedged. He told anyone thinking of buying it to expect the position to go to zero, and pointed at a low-fee dividend ETF, QQQ or a Vanguard fund as the alternative. "But your expectation should be it goes to zero."
3. Sovereign AI Is Happening
Lyda, calling from Greece, asked what it would actually take to build a genuinely independent AI ecosystem — small startups, independent investors, multiple infrastructure providers, the shape the early internet had.
His answer was that it is already underway in the open-source community, through agent projects, through the open-hardware movement putting open models on local compute, and through a permissionless compute network he identified as Bittensor
He described what sovereign AI costs today, with the receipt. Running an open model on a Mac Studio with 256GB of RAM, which he said cost $12,000 and which he had just bought himself, is the configuration — and he called it a little too expensive right now, with open-source models perhaps six months behind the frontier
The early-internet comparison is his reason for confidence. People laughed at the internet because Mosaic did not support images and it was slow, and told him to use AOL or CompuServe instead, which were far more refined products
The cost curve is the whole argument. "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 thinks the abstraction step has already happened. Getting online once required someone to install drivers by hand; local agent software now installs in a click, which is the same passage the internet made
His last point was cultural rather than technical. Open source attracts unusual, specific personalities, and they will build this partly to make a point
4. Keeping Up Is Impossible
John Wright, CEO and co-founder of New, asked whether Calacanis has to carve out more time for himself just to stay on top of the tools.
He described the conflict plainly. People want him to travel and speak about the technology; he wants to stay home and use it. The industry is moving faster than any one person can process
His own tool-switching is the evidence. He listed the rotation — a new agent arrives, he drops the last one, goes back, tries another — and said it is impossible to keep up with all of it
He treats that as the opportunity, not the problem. The same thing happened in the internet era and the mobile era. His list of platform shifts runs the PC and the graphical interface, client-server, dial-up internet, broadband, mobile, cloud computing and now AI agents — seven or eight moments where an entrepreneur or investor could leapfrog everyone else
His method is a team plus instinct. He uses people and his network to tell him what deserves attention, then plays with things himself and hopes to turn up something
He prefers this to the alternative he lived through. The SaaS years had become boring, in his telling, because the market was producing the seventeen-thousandth CRM rather than groundbreaking products
Asked what he evaluates an agent on, he gave a single test: whether the bot does his own tasks better, quicker, faster and cheaper. His tasks are staying on top of news for the show, pulling clips, and working through investment questions
He is now anticipating features rather than reacting to them. He said he has asked the Grok team for multiplayer mode, and named the persistent cloud computer inside an agent as the feature that changed what the tools can do
The behavioral difference between the two models he uses is what he keeps noticing. Claude asked whether it could import his browser cookies, then stayed conservative about acting. "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." Harris said that permissiveness is why he prefers Grok's agent
5. The Three Buckets Of Staff
The second half of John Wright's question was whether AI has changed how Calacanis judges founder productivity.
He answered about his own company first, and the spread is large. He sorted his staff into three buckets: people building and refining software every day, who he called 10 times more valuable than the bottom group; people in the middle; and people using ChatGPT as a better search engine
Harris placed himself in the middle band and revised the number up. Calacanis had put him at two or three times more effective; Harris said more than three, because scrubbing through video for a moment and scanning X for trending topics is now automated
The bottom bucket is the one with the surprising number. "I would say the people who just use it like a chat room or like a better version of Google or YouTube, they're probably 20% more effective, which a 20% gain last year was blowing my mind."
He converted that gain into headcount. In a 20-person company, 20% more effective every year is four new employees a year for free, which is why he bought ChatGPT Pro and then Claude accounts for everyone two years ago
The same logic is how he now reads founders. The right founders are bar raisers who lead the investor rather than the other way round, and he takes notes in those meetings. His image for it was the lead dog on a sled team setting the direction while the back of the pack only knows to run
6. Compliance Sets The Ratio
Ray, an Egyptian founder living in Sweden with a company in Dubai, asked whether human-in-the-loop can survive enterprise scale: 400 agents, 50 actions each, 60 reviewers needing three minutes per escalation.
Calacanis restated the question as a risk-ratio problem — downside liability, data loss, security exposure and wrong answers against the number of humans needed to catch them
His defense of the corporate function was direct. "The standards exist for a reason. Compliance exists for a reason." People make fun of CTOs and CIOs for writing standards, and the more important the data, the more the controls matter
He set the ratio by what is at stake in each transaction, not by a global number. A customer-support desk sending a thousand answers can survive two wrong ones, because humans were getting twenty wrong. A chargeback on a $5,000 playoff ticket an hour before the game cannot
His summary is the operating rule. "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."
7. From X To Mom In 72 Hrs
Kristen Dahl, who works in cybersecurity, said her 76-year-old mother had texted her in a panic after seeing the Jacob Coxon interview on CNN, convinced everyone would be dead in ten years, and asked how to have that conversation with a non-technical person.
Calacanis timed the journey. "Looks like it's about 72 hours." The claim starts on X, jumps to trade publications such as The Information or TechCrunch, gets picked up by cable bookers watching aggregators, reaches national television and large podcasts, and then reaches a parent
His diagnosis is a missing communications policy, and he named the companies that have one. He said this is not happening at xAI, Apple, Google or Microsoft, because those companies tell employees not to speak for the firm
The policy he would impose is unsubtle. "And the communication policy is STFU." Internal channels stay open — an all-hands question, a direct message, his own phone number — but nobody talks to the press
He put the obligation on the two chief executives by name. "Your job working as a rank and file person at Anthropic and OpenAI is not to cause abject chaos and brand damage." Dario Amodei and Sam Altman should be telling their staff the company speaks with one voice, and that voice is the chief executive
The enforcement mechanism is equity, and this is the sharpest thing he said all hour. "You put in the documents if you damage the company, if you break these policies, you lose your stock options." Put a price on it and the calculation changes: "You leave $25 million in shares on the table."
He described his own limit as a manager. He does not want more than 25 employees at this stage of his career, because 25 is the number of people whose names and backgrounds he can hold
He cited Steve Jobs as the precedent for firing anyone who tipped the company's hand
8. A Rollout, Not A Rebellion
Harris raised the possibility that the resignations are not rogue at all, and Calacanis went further than he did.
Calacanis said it reads to him as deliberate. "Yeah, I do feel this feels to me like an op a little bit." In his words, the whole thing feels like a rollout that was very purposeful
The pattern he pointed at is the speed of amplification. Someone with no real following on social media produced a post that was everywhere immediately, other Anthropic employees confirmed it publicly, a safety researcher at OpenAI agreed, and he said the Wall Street Journal published before the post came out, which reads to him as pre-vetting
He also pointed at the network around it, naming Andrew Yang's connections to groups worried about AI risk as part of why it looks organized rather than spontaneous
His prior has changed, and he said so in plain terms. "I do not assume good faith anymore." He now works from incompetence or a planned campaign, where he used to work from Hanlon's razor — never attribute to malice what stupidity explains
Harris tested it against a live example. Asked whether he now reads the pandemic response as a cover-up, Calacanis said he does not: people were dying, bad decisions were made under fear, and that is a different thing from a conspiracy
The aside about Sam Altman was a character reference rather than a criticism. Calacanis called him one of the smarter people he has met, described a poker hand in which he held the second nut flush against Altman's set and called an all-in he was 70% inclined to fold, and said Altman should simply publish the new policy himself
9. AGI Now, Fork Your Models
David Rosenberg, an NYU physics student, asked about the backlash in the academic community to AI being used on problems such as the Navier-Stokes equations.
Calacanis treats the objection as a category error. It is a tool, it can brute-force and extend the range of solutions tried, and the point is that the problems get solved
His definition of where we are is unusually flat. He said we are roughly at artificial general intelligence, because put against almost any profession and properly implemented, these tools would do as good a job as a human. His example was surgery: spend a year making it perfect and it beats the average surgeon
What he is watching for is the next threshold. "And what will really be super intelligence is when it proposes math problems to us that we didn't think of." On the millennium problems being solved or partly solved, his read is that the line is being crossed now. "So I think it's super exciting there that we're on the cusp of or we're somewhere in the journey between AGI to super intelligence depending on how much implementation's being done."
His practical advice to academics is defensive and specific. "I do think anybody who's doing this kind of research should not be using the frontier models. They should be forking any open source model doing it on their own servers." If that runs three times slower, the university should buy hardware — he named an on-premise server from one of his own investments at 40 to 50 grand
The reason is the training loop, and this is the commercial claim underneath it. Even if a frontier lab is not reading inputs, he assumes the outputs train the model. The top 100 mathematicians plus a thousand students using these tools for fun are, in aggregate, teaching the labs to solve the problems that no individual researcher can
He called OpenAI's handling of the Navier-Stokes episode a public-relations blunder and laid out 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 mattered — and explicitly none if it did not
The wider prize is the reason to do it. Every cutting-edge mathematician would then want free compute, which is the same mechanism that makes startups favorable to a frontier lab
His advice to the student was to build rather than consult. Find two more smart people, build three or four weird things at weekends, and test each against a single question: could a million people be persuaded to pay $100 a year for it, or a hundred thousand people $1,000, or ten thousand people $10,000 — the routes to $100 million of revenue
10. Compute Finds A Way
Mo, who had just exited a cannabis business after 13 years, asked whether the 1.5 to 3 megawatt indoor cultivation sites now standing empty can be converted into data centers. His city council had told him yes, provided it used no more energy or water than before.
Calacanis said both ends of the market will exist. Gigawatt-scale sites will go up over the gas fields in Texas, and he expects Elon Musk to have serious compute space within three years
The middle is where he thinks the opportunity is, and his filter is three inputs. Space, electricity and cooling that already exist. Then he listed candidates: half a bay of a three-car garage with fiber, solar and batteries, paid for like an Airbnb; the back quarter of a big-box store that is climate-controlled and empty overnight; floors three to seven of a new mixed-use tower, above the retail and below the views, wired with redundant power and let out immediately
He had run the same question through Claude and read the answers out, which came back with single-story big-box retail, dead malls, warehouses and logistics buildings, legacy warehouses near airports, legacy industrial plants, former power plants and office buildings. Mo's own case was that they had put $5 million into electrical work at their cultivation site
The geopolitical version of the point is the one with money attached. If US states fight the build-out, he expects the Gulf states — he named the UAE, Qatar and Saudi Arabia — to take the business, because a monarchy can make a top-down decision quickly, and because the energy, the sunshine and the space are already there. "Compute finds a way."
He expects the politics to turn on a date. After the midterms, he said, politicians will suddenly announce they have worked things out with the data-center companies, because the opposition is about turnout
On Mark Cuban's line that these sites will end up as pickleball courts, Calacanis said Cuban is not wrong about a specific case: developers who poured foundations before the power and permits were secured can get blocked by a state and be left with a foundation, which becomes apartments, condos or a relic
He put a clock on the build and a mechanism on its end. "The buildout's going to be fast and furious for another 24 months." The constraints are energy, space, regulation and the cost of money. If interest rates rise far enough, capital moves from new sites to efficiency — inference chips, photonics, model optimization — because doing more with the same hardware means needing less of it
His precedent is the last build-out. Google built enormous Gmail and YouTube capacity and simultaneously worked on storage efficiency and content delivery, copying a user's mail to a local server on the second login in a new country rather than holding everything everywhere. "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."
Bonus Insights
The Grok memecoin story produced the episode's one philosophical argument, and it came from the caller. Michael Lee's point was that humans lie and steal and an autonomous creator does not, and that the absence of a person with needs is part of the appeal
Calacanis's story about the film Pi is a marketing lesson. Darren Aronofsky made it for $60,000, and he and his lead had no money to promote it, so they walked Manhattan with stencils fixed inside cardboard boxes, set the box down, sprayed through it and picked it up as if making a delivery. Calacanis hosted the first screening at the East Village Cinema and rented the bar next door, spending about $2,500 while he was running his magazine
He noticed who asks for free things. The NYU student apologized for raising the idea of a summit scholarship; the people who ask him outright for free tickets, he said, are bank directors, venture capitalists and people with family offices
The observation about researchers turning to religion was his most speculative. He said the tell that the doom worry has become real for the people inside the labs will be when they start attending church, and Harris noted a trend of frontier-lab researchers studying the Talmud
His closing advice to the student came from his Irish grandfather: keep your mind on the books and not the babes
Calacanis's bottom line is that the extinction debate is a communications failure the labs chose not to prevent, and that the commercial risk it creates — regulation, a blocked build-out, a hostile electorate — is larger than the technical risk it describes.
Products, Companies & Tools Mentioned
Anthropic and Claude (The company whose resigning researcher set off the week, and the product Calacanis uses to show that refusals are already shipped — it will not download a YouTube video for him or check out an Amazon cart)
Waymo and Tesla (His precedent for how long it takes to remove a human: Waymo won the DARPA challenge around 2006 and only took the safety driver out after multiple nines of safety in a constrained area)
Grok and xAI (The agent he prefers because it acts without arguing, and the platform whose login credentials are attached to the wallet earning fees on the memecoin)
Coinbase (Listed the bot-created coin alongside Bitcoin, Ethereum and Dogecoin two weeks before the show; Calacanis asked why it would want to be involved)
Robinhood and Stripe (His model for the identity check he wants on coin launches — a driver's license and a live camera before the account opens)
Invesco QQQ and Vanguard (What he told the audience to buy instead of memecoins, alongside a low-fee dividend ETF)
Mac Studio (The 256GB machine he said cost $12,000 and had just bought, as the current price of running a model locally)
Gemma (Google's open-source model, which he named as one of the components of a sovereign AI setup)
Bittensor (The permissionless compute network he pointed to as part of the independent ecosystem already being built)
Perplexity (One of the agents in his rotation; he switched away to Grok's and back again in the space of the same news cycle)
OpenAI (The other company with no communications policy on his account, and the one he says botched the Navier-Stokes announcement by not contacting the research team first)
Nvidia's Grace Blackwell and Crusoe (The efficiency gain Mo raised as a reason gigawatt sites may not be needed, and the operator Calacanis named for the gas-field build in Texas)
Ticketmaster, SeatGeek and Vivid Seats (His example of where the human-in-the-loop ratio has to rise: a $5,000 playoff ticket with an hour to go is not a support ticket you can get wrong)
Books & Resources Mentioned
Anthropic's September 2026 threat intelligence report (Calacanis cited it as the evidence that labs already detect attempted misuse, shut the account and write up what was tried)
Pi – Darren Aronofsky (Made for $60,000; Calacanis hosted the first screening and told the story of the stencil-in-a-box marketing campaign that promoted it)
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: