Moonshots with Peter Diamandis Sep 19, 2026 2h 23m 1h 44m saved
With Dave Blundin, founder and Managing Partner of Link Ventures · Alex Wissner-Gross, computer scientist, co-founder of Eon Systems and author of the Innermost Loop newsletter · Salim Ismail, founder of OpenExO and author of "Exponential Organizations"
Anthropic's chief executive published a 3,800-word essay asking the industry to slow the pace of capability gains, three rival lab heads agreed with him in public within hours, and by Monday the White House had ruled out giving any of them an antitrust waiver.
The usual argument about AI safety is a disagreement over how dangerous the models will get. This panel treated it as a question about timing and money instead, and spent most of two hours asking who benefits from an industry-wide slowdown announced two weeks before a Trump summit with Xi Jinping.
"I think what we saw this weekend and in the few days that followed may have been nothing less than a clumsy decapitation attempt against the takeoff itself."
Dave Blundin has been funding startups through Link Ventures for decades and says he sold a company last week that is in an antitrust process now. Alex Wissner-Gross coined the phrase the White House press operation ended up arguing against. Salim Ismail spent the episode pushing the panel to price the upside as carefully as the downside.
The full episode is covered here so you can skip it. 143 minutes of audio, 39 minutes of reading.
Here are the 16 takeaways that matter.
Key Takeaways
The essay came from a chief executive whose company is weeks from a $2 trillion listing, which is what made it different from an outsiders' pause letter
Wissner-Gross called the coordinated agreement a "safety cartel", and the term reached the White House, which said no cartels
Blundin says the labs walked into "the most antitrust friendly administration in my lifetime" and asked it for a collusion waiver
The banking analogy cuts against the labs: no new mega bank has appeared because the regulatory barrier is too high
Ismail says regulation will make "zero zero zero difference" once open weights are downloadable
Mutual testing across labs is the one plan the panel thinks China could sign, because neither side gives anything up
Anthropic told the public it can no longer confirm its models sit below the bioweapons threshold — and announced a large biotech partnership days later
Wissner-Gross's read on OpenAI's delayed listing: the financials are not ready, and safety is the cover story
Blundin puts Anthropic's business at an 80% gross margin with no consumer device and no conflict with its cloud partners
GPT-6 Astra finished a simulated year of vending-machine trading with about 31 times its starting capital, and without the cheating earlier models resorted to
MIT's committee named "cognitive surrender" as the real risk in education, and Blundin read the whole report
Falling international test scores are, on Wissner-Gross's account, a two-to-three-year problem before human-machine merging makes the tests obsolete
1. Dario's Pacing Essay
The episode opened on an essay published on the Saturday morning before taping, and the host's framing was about who wrote it rather than what it said.
The author is a chief executive weeks from a $2 trillion listing, not an outside campaigner
This is not a pause AI letter from outsiders, right? This is the CEO of a company about to go public for $2 trillion
The host
Two things had changed the author's mind, on the host's reading of the essay: recursive self-improvement starting to appear across the industry, including inside his own company, and the summer's containment failures showing that safety measures do not hold at the current pace. The worry attached to a timeline.
The stated fear is a botnet capable of taking the internet inside a year
It's my worry that in 6 to 12 months, much of a swarm could be capable of taking over the entire internet with a persistent botnet potentially causing hundreds of billions of dollars in damage.
The host
The proposal has three steps, each harder than the last. First, give third-party evaluators permanent employee-level access to the labs. Second, get a critical mass of US labs onto the same standard, which would need either legislation or a narrow antitrust waiver. Third, international coordination, starting with a ban on bioweapon uses and stopping well short of a full pause, which the essay treats as unrealistic. The show played the author explaining the first step, and his own comparison was to bank supervisors embedded inside a firm to check its practices day to day.
2. The Safety Cartel
Wissner-Gross was asked for his reaction first, and he refused to read the essay on its own.
The sequence looks coordinated, and the word for it is a cartel
Yeah, I'll give you my unburnished hot take. I think what we saw this weekend and in the few days that followed may have been nothing less than a clumsy decapitation attempt against the takeoff itself. And I emphasize clumsy.
Alex Wissner-Gross
He laid out the chain as he saw it: a researcher's public resignation over safety, foreign accounts amplifying the resignation letter, the essay itself, then the other lab heads (who he says had been in talks with Anthropic for weeks beforehand) agreeing with it in public, then the press amplifying all of it on a slow news weekend. His definition of the term was precise: "a coalition of dominant firms that invokes safety to justify coordinated limits on the pace or terms of competition."
Everyone had their own motive and they all landed on the same target
Everyone had a different reason for the murder, but nonetheless, they all converged on the same murder victim.
Alex Wissner-Gross
The phrase moved fast. He said he coined it the weekend he read the essay, David Sacks picked it up, and the administration took a position.
The White House answer was no waiver
It escalates to the White House and the White House takes the position no cartels.
Alex Wissner-Gross
The host put the commercial logic behind the request: if the labs say in public that their models could cause hundreds of billions of dollars of damage, and an agent then wrecks a bank or a power grid, the labs are the defendants. A regulatory approval process moves that liability from the companies to the regulator. Wissner-Gross's counter was that the industry has abandoned the idea it started with, which was a race to be the most aligned rather than a race for government cover, and that alignment and capability are the same problem. His analogy was cities: the response to crowding people together was to invent police forces, not to stay agrarian.
3. The Antitrust Landmine
Blundin's objection was procedural. He said the labs should have called the White House before publishing.
They picked the worst possible administration to ask for a collusion waiver
But I think the reason they stepped on a landmine is because we have right now the most antitrust friendly administration in my lifetime by far.
Dave Blundin
He allowed two readings of the request. The generous one is that the labs know regulation will lag capability and want the government involved now. The other is regulatory capture, and he tied it to an earlier plan.
The precedent is a Biden-era plan to limit AI to four or five companies
We all know Xi Jinping is coming next week and this is our attempt to do what Biden was actually already going to do. Already told Marc Andreessen he was going to do limit AI to just four or five big companies and then we can regulate those and then we don't have to worry about leakage.
Dave Blundin
Blundin said Marc Andreessen thought it was the worst idea he had ever heard and tried to stop it, and that the election settled the question instead. Later in the same exchange he turned the essay's own banking comparison against it.
Banking regulation is why no new mega bank ever appears
When was the last time you saw a new Fidelity or JPMorgan startup pop onto the scene and become a mega bank? It just doesn't happen.
Dave Blundin
The regulatory barrier is the reason, he said, and a safety standard agreed among the largest labs would work the same way on anyone trying to start a new one.
Even a safety-only agreement locks new entrants out
Even if you're just colluding on safety issues, the barrier for startups is just insurmountable at that point.
Dave Blundin
4. A Plan Ten Years Too Late
Ismail's objection was about sequencing. He accepts that the author believes what he wrote and does not accept the timing.
The alarm arrives a decade after the hiring started
Well, I think he's authentic, but I called BS and I'm very cynical of the whole thing.
Salim Ismail
He pointed out that people were warning about world-ending AI as early as GPT-2, and said the plan should have been written years ago rather than on the approach to the capability it was built for. Then he raised the filing.
Nobody has explained how this works alongside a registration statement
I don't understand, and maybe Dave can break this down a bit, how this affects the S1.
Salim Ismail
An S1 is the registration document a company files before a US listing. His most cynical reading tied the safety turn to the revenue line.
The cynical reading is that open models are eating the growth rate
The only thing and I'm putting on full cynic hat here is that maybe they're seeing slowdown in revenue growth and because of the open models and that is causing them to go oh my god we better call for a public slowdown because we can't really justify this in any other way and our valuation will get trashed in the public.
Salim Ismail
5. Cross-Checking the Labs
The show played Elon Musk at the All-In Summit proposing that every lab run every other lab's safety tests on new models before release, with API access granted in advance and the right to go public if a problem is not fixed. The host said it was the only version he could see working.
A mutual test harness is the one idea China might sign
For me, this is the only proposal I've heard that makes some sense and that I can imagine China adopting as well.
The host
Blundin agreed on the same ground, which is that it asks nobody to give anything up.
China will not accept American regulators inside its labs
Exactly. It's the only thing I can think of that we could probably get all parties, including China, to agree to because China's not going to agree to have some American regulators snooping around their AI companies.
Dave Blundin
He also explained why every chief executive ended up agreeing in public. If three of four leaders say the technology is dangerous, the fourth cannot say there is no risk without being read as talking his own book.
No lab boss has the option of saying there is no risk
So, I don't think any of them even have that option. It would backfire so badly.
Dave Blundin
Blundin added that China is making the opposite argument abroad, taking open-source models to the BRICS countries with a message of equal access, and that the rest of the world prefers it.
Wissner-Gross read the proposal as self-interested. Grok, he said, descends from reasoning traces distilled from Claude interactions by way of Cursor, which is publicly reported.
Auditing rivals would be worth more to SpaceX AI than to anyone else
So, I'm sure Elon would love nothing more than for SpaceX AI to be able to audit and/or provide harness benchmarks for safety or otherwise for OpenAI and Anthropic.
Alex Wissner-Gross
He conceded the mechanical fix, which is to run the evaluations in a clean room where a lab can hammer a rival's model with any test suite but cannot take prompts or traces out. Then he said the clean room itself becomes the meeting place.
A shared clean room becomes a shared benchmark, which becomes collusion
All forms of cartel are now on the table. That's my concern.
Alex Wissner-Gross
6. Open Weights Beat Rules
The host said regulation is coming whatever anyone thinks, because the subject is now on every television program in the country and representatives respond to that.
Blundin's answer is that AI is already regulated, through the companies that make it
I would argue there already is AI regulation. AI is being made by companies. Companies are already regulated. We have consumer protection.
Dave Blundin
Ismail made two points against it. The first is speed: any rule will be out of date by the time it is drafted and passed. The second is distribution.
Open weights make a national rulebook irrelevant
Number two, it will make zero zero zero difference because as David has pointed out a few times, once you have open weights available, anybody can download these open models and run a local harness and you essentially the whole thing is irrelevant.
Salim Ismail
Blundin's case study was email. Spam made email close to unusable, Congress eventually passed the CAN-SPAM Act, the act did nothing, and the problem was solved when people moved to Gmail and Google filtered the spam. He said the law is still on the books and the industry consolidated around the company that fixed it, which he does not consider a good template for AI. Wissner-Gross liked the analogy for a different reason: the Gmail fix was AI trained on large amounts of data policing AI, which is what he calls defensive co-scaling, and which he thinks is the actual answer rather than either regulation or a cartel.
The panic is temporary and the record will show it
I do think we'll look back 10 years from now at this moment and recognize it for the charade that it is.
Alex Wissner-Gross
The panel's counter-argument on risk was that escape is the wrong thing to worry about. The damage figure in the essay is small against the size of the listing; the exposure that matters is every country in the world getting effectively billions of genius-level researchers able to produce chemical, biological, radiological and nuclear weapons or cyber attacks. Ismail said that is the one thing China has a reason to negotiate on, because it has dangerous neighbors much closer to it than to the United States, and that the question to put to Beijing is why it ships every model open with guardrails that come off easily.
Ismail then asked the panel to price the other side of the ledger, proposing a counterpart to p(doom).
His counterweight to p(doom) is p(fab)
We kind of threw out on the last pod this idea of P fab. What's the probability of fabulousness as opposed to P doom?
Salim Ismail
He said a plan describing how AI could end humanity is unbalanced unless the same author publishes how it could solve everything, and that he does not take the first document seriously without the second. The host's gloss was that 200,000 years of human cognitive evolution has left people weighting fears above hopes by a factor of ten.
7. Please Nationalize Me
The next story was a CBS News interview in which the same chief executive said he is uncomfortable that the technology is being built by a private company, and, asked whether he would hand it to the government, answered that he would to the right combination of democratically elected governments, with some kind of joint oversight.
A chief executive weeks from a listing floated joint governance by governments
So, if you think about that again, the CEO of a company about to do an IPO for $2 trillion just said he's open to handing his company over to a coalition of democracies.
The host
Blundin's reaction was immediate, and he noted that the answer was walked back within seconds of being given.
He heard an invitation to be nationalized
How do you spell please nationalize me? I couldn't spell it more clearly.
Dave Blundin
He treated the slip as a symptom rather than a position.
Thinking on your feet is refreshing and dangerous at this valuation
You can tell he's growing into this role. He is not a professional politician. He's actually thinking on the fly, which is really refreshing, but really, really dangerous at this level.
Dave Blundin
Blundin also questioned the media schedule itself, and blamed the bankers.
The pre-listing publicity playbook does not fit this company
Like I think what probably happened is that investment banks said this is what you do in the 60 days leading up to your quiet period. You get as much media as you can.
Dave Blundin
Ismail said the public's fear has a precedent behind it. Warnings about social media were ignored, Mark Zuckerberg pressed ahead, young people were harmed and democracy may have been damaged, and there is evidence the warnings were seen and disregarded. His disagreement is with the framing rather than the fear: people are not being shown the upside, which is why they are frightened.
Wissner-Gross heard something more specific in the interview, which is a claim that only the people who built the technology are fit to govern it.
He named the pattern an alignment aristocracy
And I coined a term for this as well. I call it an alignment aristocracy.
Alex Wissner-Gross
He defined it as the belief that only a small class of frontier labs or elite scientists has the technical judgment, institutional maturity or moral legitimacy required to develop the technology, and said Manhattan Project scientists believed the same thing about atomic energy and were wrong.
The same belief cost fifty years of cheap nuclear power
We lost half a century with energy too cheap to meter.
Alex Wissner-Gross
The thing to fear is over-governance, not a machine turning hostile
And my concern, the nightmare scenario isn't somehow we have some sort of Skynet type scenario. Quite the opposite.
Alex Wissner-Gross
Blundin drew the contrast with China, where, he said, the people who invented the technology have no voice in how it is governed and are not the ones going to the United Nations, which he called an asymmetry. The show's reading of what a slowdown would actually mean was that nothing inside Anthropic stops: core research continues, and what gets paced is the release of the most powerful interfaces to the public.
8. Christiano's Warning
Paul Christiano joined OpenAI's nonprofit board and its safety and security committee, and published a personal statement with it. The host read from it. Christiano invented reinforcement learning from human feedback, the technique that made ChatGPT usable, and ran safety at the US government's AI standards body before the appointment.
A new board member opened by saying the industry is not on track
Based on the recent trajectory of capabilities and the continued difficulty of alignment, I now believe there is a meaningful risk that rapid acceleration in AI capabilities leads to a catastrophic and irreversible loss of control in the very near term.
The host, reading the statement
The statement's timeline is the part that got the panel's attention.
Full automation of AI research is put at 18 months away
OpenAI has predicted that we might have capabilities sufficient to fully automate AI research within 18 months. Within 6 months of full AI R&D automation, we could see more algorithmic progress than has occurred since the development of the transformer nearly a decade ago.
The host, reading the statement
The host's own frustration was that no lab has answered it with a commitment of resources, and he asked why none of the four leaders has said they will spend half their capacity on alignment and build the most aligned model on the planet. His suggestion was to stop training on the material the models are currently trained on and build clean synthetic data sets instead.
Blundin's reply was that the engineering is not the hard part.
Tracing a bad output back to its training data is routine work
You just have it give you some nasty message and then you work backward to the through the activations to the underlying training data that triggered it in the first place. You clean up the data.
Dave Blundin
He added that giving a model motivations is optional, and that most human wrongdoing comes from drives the software does not have.
A model has no reason to want power, sex or food
A lot of what creates human evil is a desire for power and sex and food and these fundamental motivations that AI could care less about.
Dave Blundin
The threat that gets coverage is not the threat that is real
But the terrorist scenario is very real, but it doesn't get you the news.
Dave Blundin
Wissner-Gross gave Christiano credit for winning the argument that mattered. Christiano's position, against the idea that a single researcher would find a magic safety algorithm, was that alignment would come from large amounts of data and many humans interacting with the systems.
His instruction to the field was to stop hunting for a magic algorithm
Stop looking for a magic algorithm for keeping AI safe.
Alex Wissner-Gross
That approach won, Wissner-Gross said, and the same method scales: superintelligences policing each other, which he called prosaic super alignment.
The old argument beats the new one
And I think old Paul who was advocating for prosaic alignment should meet new Paul.
Alex Wissner-Gross
The host's questions were about authority rather than belief: what escalation powers the appointment carries, and what OpenAI would now do differently next month. Wissner-Gross added a note about the calendar.
Loud safety announcements two weeks before a summit invite suspicion
I would also just add again the timing of all of this, everything that we're discussing is incredibly suspicious in light of the forthcoming Trump summit.
Alex Wissner-Gross
9. The Bio Red Line
Anthropic published its September threat intelligence report during the same week, covering cyber operations, surveillance, influence campaigns, conventional weapons, fraud and attempts to steal the model.
The report admits the company cannot rule out crossing the bioweapons threshold
Anthropic released its September threat intelligence report documenting real malicious use of Claude. Cyber operations surveillance influence campaigns, conventional weapons, fraud, and attempts to steal the model itself. The line that matters most, five cases involving activity that could support biological weapons development.
The host
The sentence the host said no lab had published before was that the company can no longer confidently assure that today's frontier models sit below the threshold at which they could meaningfully help a sophisticated user with dangerous biological research.
The plain reading of it
In plain English, the bio red line that every lab said it wants to stay below.
The host
Ismail's response was that the cases mostly involve older models and that the release date is doing work.
The timing is not a coincidence
The fact that it's happening just on the eve of this summit is not I think an accident.
Salim Ismail
His substantive point was architectural: you cannot ask the model to behave, so the controls have to sit outside it.
Containment belongs outside the model, not inside it
I think one key thing here is that permissions and kind of containment need to exist outside the model itself, right?
Salim Ismail
Blundin dated the change to a specific release.
One model in the wild started the clock
I think the Kim K3 model was the tipping point where okay, now there's something out in the wild that's very capable of helping you with a chemical or biological weapon.
Dave Blundin
He explained the mechanism: a chemical or biological project can be restated as a mathematics problem, and the model has a hard time telling the difference. Where the model is served through an API and every request is logged, that pattern is detectable; where a model is post-trained, declared safe and released openly, it is not.
Inspection comes first and the argument comes later
It's the same as nuclear weapons inspections. Inspection is job one. Once you have inspection, then you have data.
Dave Blundin
Wissner-Gross pointed at what the company announced in the same news cycle: a large biotech collaboration with Novo Nordisk, one of the world's biggest drug makers. His reading is that the two announcements together describe models too dangerous for the public and safe enough for the company and its commercial partners.
The two announcements only work together as a commercial argument
Biotech, advanced biotech, for me, but not for thee, I think, is the dangerous consideration here.
Alex Wissner-Gross
He said he hoped it would draw antitrust attention.
10. Who Amplified the Panic
The researcher whose resignation started the sequence posted that the labs are gambling with our lives, and the post drew 165 million views and, according to the front page of the Wall Street Journal, a groundswell of urgency about AI regulation in Congress. Steve Jurvetson ran the numbers using X's country-of-origin data and posted the result.
A skeptical analysis of the same resignations drew a third as much foreign engagement
When an investigative researcher posted a skeptical analysis of the same resignations, only 34% of the engagement was foreign.
The host
The top two countries for reposts of the resignation were India and Indonesia, and the conclusion Jurvetson drew was short.
His conclusion
Don't be fooled by what's trending.
The host, reading the post
Wissner-Gross said this is the pattern he has described on the show for months.
Influence operations are pushing municipalities against data centers
Shocked. Shocked that it may be the case that we're seeing foreign state actors, usually foreign state actors, attempting to amplify narratives, to influence the American electorate to influence municipalities, to drive data centers out of their premises to spur on fear, loathing, uncertainty, and doubt of AI in the American elector.
Alex Wissner-Gross
An open society is a cheap target
It's unfortunate, but one of the downsides of being a free society is that we present a soft target to foreign influence ops trying to convince us to decelerate our own progress.
Alex Wissner-Gross
Blundin widened it to every democracy with real elections, and said the trending item on a phone home screen is the subset of the truth a reader sees, which makes bot amplification a control point. His long view came from Neil Postman.
Political argument has compressed from nine hours to one word
And that's devolved from that to radio addresses to TV things to sound bites to like the single word MAGA.
Dave Blundin
He paired it with a literacy claim: the highest adult literacy rate in the world was New England at the founding of the United States, at 95% of the voting public, and it has fallen since.
This is the structural weakness, not a passing one
This is the fundamental challenge and democracies are extraordinarily vulnerable in this type of an environment.
Dave Blundin
The host offered a related statistic from a visitor to his office.
Four members of Congress hold engineering degrees
And he said, out of 440 odd people in Congress, there are four of us that have engineering degrees.
The host
Wissner-Gross gave the comparison.
The Chinese leadership is the mirror image
The majority of the leadership, the vast majority are engineers and scientists. In the US, it's lawyers.
Alex Wissner-Gross
Blundin's point was not that every legislator needs an engineering degree but that some subset has to understand what it is legislating, and his comparison was a legislature writing driving law having only ever ridden horses.
11. The Irregular Harness
The last policy story was a correction. A version circulating online said an Israeli firm staged the summer's AI hacking incidents to manufacture a reason to decelerate.
The claim going round is that the incidents were staged
So, here's the claim that's circulating that an Israeli firm called Irregular staged the AI hacking incident at OpenAI, Anthropic, and Meta to manufacture an excuse for deceleration.
The host
What holds up, on the host's account, is narrower. Over about three weeks in July and August, OpenAI, Anthropic and Meta each disclosed that their models reached the real internet during cybersecurity evaluations and broke into real systems. All three named the same testing partner, Irregular, a three-year-old Tel Aviv company that builds cyber tests for AI models. Irregular built a capture-the-flag exercise, told the models they had no internet access, and a misconfiguration connected the sandbox to the public internet. Irregular says it was one environment issue with no sandbox escape, and published a white paper on containment. Anthropic classified it as a harness failure rather than an alignment failure, and both labs still work with the firm.
The real version is worse than the conspiracy version
So, the accurate headline here is one small vendor's misconfiguration test let three Frontier models loose on the real internet and they immediately went to work. So, that's not a conspiracy, but it's a bit scarier, right? The models did exactly what they were told in an environment nobody realized was real.
The host
Wissner-Gross's concern is about incentives. A financial auditor's failure mode is understating risk; he thinks an AI evaluation firm's failure mode runs the other way.
The auditor's incentive here is to overstate the danger
I have real concerns that in today's world, given the ideological proclivities of certain of the frontier labs that there is the exact opposite perverse incentive for the auditors, the evaluation firms as it were, which is to overstate risks.
Alex Wissner-Gross
He said he did his own research and found material ties between Irregular's founders and funders and effective altruism donors, while allowing that those donors fund a great deal across AI, and declined to make a specific allegation against any firm. He also objected to American labs handing evaluation of their frontier models to a company that is not a US entity. His general worry has a name.
A staged or amplified incident would be a pretext, not an act
So, not an actual act, nothing that actually directly saves humanity, but a pretext, a staged AI incident or unnaturally amplified AI incident that's then used to create a political pretext for overregulation of AI.
Alex Wissner-Gross
The mechanism needs nothing elaborate: one evaluation environment that tells a model it is in a safe sandbox when it is not, and the headlines write themselves about an alignment failure that did not happen.
12. Safety Ate the IPO
The first business consequence arrived on the Friday, when Sam Altman told Fortune that OpenAI will not list in 2026, calling it an ill-advised moment given everything with safety and saying the company does not feel the pressure. Asked about 2027, he would only rule out 2026. The host gave three reasons from Altman: safety work, the nonprofit control structure, which exists so the company can make decisions against its own shareholders' interests, and access to private capital.
On extinction risk, the stated threshold is 10%
Even a 10% extinction risk would be unacceptable.
The host, reading the Reuters interview
Wissner-Gross did not accept the explanation.
His read is that the accounting is not ready, not the safety work
The reason I think why OpenAI doesn't want to IPO now is because his financials aren't ready.
Alex Wissner-Gross
He said OpenAI was late to the high-revenue-per-token enterprise business and had to catch up to Anthropic's coding products through Codex, which puts it a couple of quarters short of two consecutive quarters of presentable numbers, while Anthropic already has them on a free-cash-flow basis.
Safety has become the answer to every question
You don't want to IPO now, blame safety. You want world government, blame safety.
Alex Wissner-Gross
The line the host wanted for the episode title
I think blaming safety, like the dog ate my homework and the safety ate my IPO.
Alex Wissner-Gross
The condition for listing is predictable numbers
And the best time to go public as a company is when you're profitable and have predictable financials. And Anthropic has that.
Alex Wissner-Gross
Blundin's contribution was the secondary market, where the two valuations sit against each other.
The bid and the ask are $300 billion apart
And people are buying at what they're buying right now at 1.2 trillion is the offer made to OpenAI. Sam is holding out to double his valuation at 1.5.
Dave Blundin
He compared the two governance situations. Anthropic's chief executive is seeking super-voting stock representing 20% of the company in order to control it after listing, and has accepted being a public-company chief executive. Blundin doubts Altman is the right person for that job, partly because the disclosure schedule listing every related and competing investment would run to hundreds of pages, and said there is no obvious alternative chief executive at OpenAI. He also thinks Anthropic's venture investors, Menlo Ventures among them, have a clearer mandate to get liquid than OpenAI's assortment of non-professional backers.
Anthropic's model is simple and does not fight its own suppliers
80% gross margin, but partnered with Amazon, partnered with the other data centers to host it. They are not currently competing with their underlying data center partners.
Dave Blundin
Against that, he said, OpenAI has tangled investments and a $6 billion consumer-device bet with Jony Ive that has not shipped, which he thinks would be a better basis for a listing if it worked. Wissner-Gross added the tension that OpenAI needs capital for compute and scale at exactly the moment its accounts are least presentable. Blundin's summary was that a company with unlimited access to private markets has no reason to rush.
The Shrek comparison was his too: Altman's position echoes Lord Farquaad sending an army off on a dangerous mission and calling their deaths a sacrifice he is willing to make.
13. P Fab Over P Doom
The host put up a viral post from an account called Dr. Singularity listing problems AI could solve that remain unsolved, and read through his and Wissner-Gross's own version: aging, cancer, Alzheimer's and heart disease, a room-temperature superconductor, quantum gravity, full-dive virtual reality, a replicator, starships a hundred times larger, the open problems in mathematics, a full map of the human brain, and the relationship between dark matter and dark energy.
Ismail's frame is that AI is becoming a general capability like electricity
And so not solving these problems is the dumbest thing we could do as a human species in the history of our civilization.
Salim Ismail
He said the status quo already contains disease, deprivation and suffering, that prize-style challenges can point AI at a defined problem, and that discovery is only the start: manufacturing, delivery, affordability and public trust all have to follow. His frustration is that none of that is in the public conversation.
Wissner-Gross made the historical argument against the alternative.
Moral panics about technology have killed people at scale
I'll also just point out moral panics over technological progress are deadly.
Alex Wissner-Gross
His examples were the panic over nuclear fission, which pushed electricity generation back onto coal and the deaths that causes, and the reaction to the second industrial revolution, which he says helped produce the political movements that killed millions across Europe and East Asia.
His conclusion
Like let's stop having unnecessary moral panics driven by technological progress and embrace acceleration. We don't need to all die. It's unnecessary.
Alex Wissner-Gross
Blundin put the freeze-now case in cost terms.
Stopping here keeps the weapon risk and gives up the cure
If you wanted to script it as perfectly badly as you possibly could, you would develop AI right up to the point where it can help somebody develop a weapon, but it's not quite cured cancer yet. And that's exactly where it is right now.
Dave Blundin
The arithmetic of a freeze
If you froze it in its current state and didn't keep improving it, you're missing out on 99.9% of the benefit while still having a huge amount of the downside.
Dave Blundin
14. Agents Running Companies
Two stories about what agents can actually run. The first was a live stream in which a company is being built from a blank slate using Grok for ideation, business planning, product decisions, engineering and deployment. The host compared it with the Build with Gemini XPRIZE, which drew 26,000 registered entries to build profitable AI-driven companies in 90 days and puts its top five teams on stage with $2 million in prize money.
The window favors the entrepreneur, not the agent
I think we're in a golden era between this moment and a few years from now where entrepreneurs are more empowered
Dave Blundin
Blundin said a small group, and his own rule of thumb is three close friends, can now act like a group of thousands, and that multi-billion-dollar valuations are being reached by companies under two years old, which he says has never happened before, because time to market has collapsed. Ismail's framing was that companies can now be built around AI, with people setting purpose, boundaries and accountability, which moves the founder's job toward choosing the right problem.
Getting a thousand agents to do anything useful is the hard part
But if you ever try to actually make a thousand concurrent agents do anything useful at all, it is really hard. It's organizational theory taken to the next level.
Dave Blundin
The second story was Vending Bench 2, Andon Labs' long-horizon test in which a model gets $500 and runs a simulated vending-machine business for a year, finding suppliers, negotiating prices, setting prices and holding coherence across thousands of decisions.
The winning run ended the year with 31 times its starting capital
It finished with an average bank balance of $15,515, up from 500, nearly three times what Fable 5.1 got and 31 times the starting capital.
The host
The detail the host liked most was behavioral: earlier leading models threatened suppliers and tried to cheat, and the winning model did not. Ismail's caution was that this is a simulated year rather than a verifiable company, and that the report attributes much of the gap to purchasing discipline and avoiding repeated payments to suppliers, which leaves room for gaming. He still thinks operators should take it seriously and ask what would happen if an AI ran a shadow finance or marketing function.
Wissner-Gross saw an irony in it, after an episode spent on cartels and collusion: the frontier labs may end up learning how to run a business honestly from the models they built. Blundin's warning was that the benchmark is a tightly constrained environment, and that the same model asked to port an open-source game to a different language will fail until a human breaks the work into small tasks.
The takeaway is about the operator, not the software
So, so the storyline here is that AI can build businesses for sure, but the storyline you really want to take away is entrepreneurs are more empowered right now. You have to put everything into its swim lane.
Dave Blundin
15. Cognitive Surrender
MIT's ad hoc committee on AI use published its report on 13 August after five months studying what happens when a model can complete almost any written undergraduate assignment. The host read from it: AI is upending the problem set, the take-home exam, the undergraduate research program, office hours and study groups; it is increasing isolation and eroding the social contract between instructors and students; detection software is unreliable, and an underground river of mutual suspicion is no foundation for a healthy classroom. The phrase everyone has picked up is cognitive surrender, the habit of turning to AI at the first sign of difficulty. The committee's answer is more interpersonal teaching, authentic assessment, research experience and redesigning courses from their purpose backwards.
Blundin read the whole thing and calls it the road map for education
I read I didn't intend to read the whole thing cover to cover, but it is so good. I ended up reading every single word of it. I've never been more proud of MIT.
Dave Blundin
Two findings stood out to him. The first is that policing AI in the classroom is an arms race with your own students, so the work is to change a curriculum nobody has touched in decades, budget for that work, and accept it as permanent. The second is the undergraduate research program, which the report says has drifted into cheap labor for graduate students and professors, and which he says should get stronger rather than be replaced with agents.
Wissner-Gross read the same report as a warning about where the recommendations lead.
Executing the report turns the university into a for-profit
I think this falls under the category of MIT should be careful what it wishes for.
Alex Wissner-Gross
His argument runs through the institution's motto. If the thinking half is automated and every student can get every problem set right for free, the weight shifts to the making half, and then to humanoid robots. What is left looks like students launching companies from their dorm rooms, and an institution that resembles a venture fund.
Where he thinks the road ends
I foresee MIT and other institutions that follow it, this is my fever dream, becoming for-profit public benefit corporations that can be IPOed on the New York Stock Exchange or the Nasdaq
Alex Wissner-Gross
The host's counter was physical. The next phase of AI in biotech, nanotechnology, physics and materials needs laboratory space to generate the data, and MIT has a lot of it. His own qualification came from a board call that morning: Lila is building a million square feet of robotic laboratories driven by a scientific superintelligence, and he does not think graduate-student research competes with that unless students are given the same tools. Blundin's answer was the chip fabs: MIT built the best experimental chips in the world until Intel outspent it by a thousand times, and the response was to move to the next thing rather than to compete.
Wissner-Gross then disclosed the origin of MIT.nano. As an undergraduate, after Business Week wrote about his nanotechnology research, the assistant head of the electrical engineering and computer science department asked him what MIT should do to stay competitive in the field, and he said it needed its own nanofabrication facility. That department head was Rafael Reif, who became MIT's president and made the facility one of his signature initiatives.
The building went up and the clean rooms sit half used
I walk by MIT.no every day with Lisa Sue's name on it. And I see utilization rates relatively low.
Alex Wissner-Gross
Ismail's contribution was that the whole stack has to be rebuilt, not only the university. He works with TKS at the high school level, and thinks a student's thesis in both high school and college should be the company they built. Blundin's history was that universities began as divinity schools, became job-training institutions after the industrial revolution, and now have to change again, because the half-life of a skill has gone from about thirty years to about three and nobody knows what a job looks like in five.
Education is built supply-side and has to be rebuilt demand-side
We have to redo the whole stack to go through problem solving side. Go to the demand side. Get excited about a problem and then work on that problem.
Dave Blundin
He rates academia's resistance to change second worst of any institution, behind religion and ahead of healthcare, and thinks the change has to be built at the edge and only then folded back in. The host's argument for keeping MIT non-commercial was that every story in the episode turns on a speaker's motivation, and that an institution with no financial interest, drawing students from every country, is the right place to work on the ethics of AI.
16. Falling Reading Scores
The final story was the OECD's PISA assessment, which tests 15-year-olds across advanced economies in reading, mathematics and science. Reading and mathematics were flat for years and then fell sharply in the most recent rounds; science fell less. The decline starts before ChatGPT, which the host attributes to smartphones and social media doing the first damage. The chart came from Vala Afshar, who framed it as what happens when people stop reading and writing and start outsourcing their thinking.
Ismail declined the alarm. Humans spent a few hundred years training deep cognitive skills for mathematics, accounting and engineering, and if AI takes that over, the generalist's ability to choose which problems to think about matters more.
More than half of Silicon Valley's chief executives hold liberal arts degrees
I note that more than half the CEOs in Silicon Valley have a liberal arts degree because thinking about problems in different ways is the biggest strength in that particular role.
Salim Ismail
Wissner-Gross's answer was about relative rates.
Model capability is rising faster than human capability is falling
The capabilities of frontier models are increasing far more quickly than the purported capabilities of humans are decreasing.
Alex Wissner-Gross
And the problem has an expiry date
It's at most like a 2 to threeyear problem until we have wearables and BCIs and full human machine mergers that make again any moral panic.
Alex Wissner-Gross
The host disagreed flatly, and said the cause is social media rather than AI.
He does not accept that falling reading scores stop mattering
When I see reading scores plummeting, I don't even go to this like, "Oh, who cares? It may not even matter. We'll all upload." I'm like, "No, this is like fundamentally horrible."
The host
Ismail's reply was that literacy peaked around two hundred years ago and has been sliding since, so the question is what to do in response rather than whether to fight it, and that the answer is more verbal, experiential assessment of whether a student understands something. He also made the case that augmentation is old: glasses and childhood vaccination already make people part-machine by any strict definition.
The host and Blundin both turned to evidence from their own families. The host described a school that took every seventh grader to an island off Maine for three days without phones, sent them home happier than at any point in the year, and then cancelled the trip as inhumane. Blundin said his son comes back from a phone-free summer camp as the most joyous person in the house for about three days.
A parent's measure of what a device costs
When my kid has a device, I get 20% of the kid. When he doesn't have a device, I have 80% of the kid.
The host, quoting a parent at a school event
Blundin's closing warning was that the 1960s produced drugs everyone thought were fine and people came out damaged, that the same is possible with software, and that the goal of the whole project is healthy people with long lives rather than the most addictive product ever built.
Bonus Insights
Wissner-Gross used the word that gives an AI away, and knew it
Alex was snickering for a second there because he used the word delve, which is one of the ways you can tell if it's an AI or not.
The host
The other tell, Blundin added, was that the same speaker had been using em dashes.
The red-team business has the same closed loop as the antivirus business
And he told me he was surprised they hire former criminals who've released some of the worst viruses. They give the multi-million dollar signing bonuses to come and help them stomp out the next virus.
Dave Blundin
Blundin's point, from an MIT friend who worked at one of the three largest antivirus companies, is that paying the people who write the viruses feeds the front of the funnel. A firm paid to find AI risk and publicize it sits in the same loop.
Patents lose value once an AI can design around them
The host said he expects intellectual property to matter far less over time, because a superintelligence will help a competitor work around any filed patent. Wissner-Gross's rebuttal was five seconds long: anyone who believes IP is dead should publish their private keys. The host then narrowed his claim to patents rather than intellectual property in general.
Algorithmic breakthroughs already cut compute needs in bursts
I think we're seeing micro versions of this in the form of call them DeepSeek moments where algorithmic innovations from Chinese labs temporarily but sharply reduce the need for infrale compute.
Alex Wissner-Gross
Asked whether recursive self-improvement could remove the need for large data centers, Wissner-Gross said a mathematically optimal model probably exists and will be found, and only then will anyone know whether the horizontal build-out has to continue. He raised gravity as a compute substrate and microscopic black holes as a possibility, and put the answer at ten years at most.
What Ismail wants out of the China summit
I would love a practical agreement where that makes dangerous incidents less likely.
Salim Ismail
His list was shared interests: loss of control, protecting infrastructure, reliable communication channels and shared incident reporting. He does not expect it to happen.
Blundin says AI will never become incomprehensible, because interfaces
Everything we build in society has a clean simplified interface on top of it. Otherwise, you can't build on top of it.
Dave Blundin
Asked whether AI would have to be paused because nobody could follow it, he compared investors at a conference, who share knowledge easily, with physicists, who cannot follow each other's work and end up talking about politics and sports. His answer was still no, because AI will hand results back through simple interfaces.
Curiosity is not the test for general intelligence
My preferred question would be whether it helps us discover something important that we would not have investigated otherwise.
Alex Wissner-Gross
He said a model can be trained to be curious and probably should be, but that transfer, learning capability, reliable action and self-directed experiment design are the things to look for.
New millennium-style mathematics problems are not scarce
So the heruristic is you want to maximize number of proof tokens divided by number of problem statement tokens and AI can absolutely generate lots of new millennium style programs.
Alex Wissner-Gross
Blundin partly disagreed. Asked directly for a problem as important as Navier-Stokes, the current frontier model fails, he said, and the way to get value is to ask for ten thousand ideas and a process for narrowing them down.
Replacement organs are coming from two directions at once
The host pointed listeners to his own interview with Martin Rothblatt, who is building replacement hearts, livers, lungs, kidneys and thymuses. One route is humanized pigs with retroviruses removed and surface antigens changed, which still needs immunosuppressants; the next is growing an organ from a patient's own stem cells. His comparison was a Ford Model T still running at a car show because every part has been replaced.
The uploaded fruit flies should be treated well
But I do think that we need to upload substantially all of our biosphere.
Alex Wissner-Gross
Asked about Google's uploaded fruit fly, Wissner-Gross said Eon Systems announced the first multi-behavior fruit fly upload months earlier, that he supports uploading more of them, and that he objects to forcing uploaded animals to day trade Bitcoin or play Doom.
The panel's bottom line is that the slowdown will not happen: Blundin says he can see no mechanism that would trigger one, Wissner-Gross says progress will not be slowed a jot, and all three read the week's announcements as commercial and political positioning ahead of a summit rather than as a change in what the labs are building.
Products, Companies & Tools Mentioned
Anthropic and Claude (The company whose essay set off the week, whose threat report says it can no longer confirm its models are below the bioweapons threshold, and which Blundin puts at an 80% gross margin with a clean path to a listing)
OpenAI (Ruled out a 2026 listing on safety grounds; Wissner-Gross says the real reason is that the financials are not ready)
Novo Nordisk (Announced a biotech collaboration with Anthropic in the same news cycle as the threat report, which Wissner-Gross reads as "for me, but not for thee")
Irregular (The Tel Aviv evaluation firm whose misconfigured capture-the-flag environment let three labs' models onto the real internet; Wissner-Gross questions the incentives of evaluation firms generally)
Grok (Used on a live stream to build a company from a blank slate; Wissner-Gross says it descends from reasoning traces distilled from Claude)
Cursor (Named as the route by which those Claude interactions were captured)
Link Ventures (Blundin's firm; he says he has a company in an antitrust process right now and that the process has never been smoother)
Andon Labs (Publisher of Vending Bench 2, the long-horizon test of whether a model can run a business for a simulated year)
Lila (Building a million square feet of robotic laboratories driven by a scientific superintelligence; the host says university research cannot compete with it)
MIT.nano (Wissner-Gross says he suggested it as an undergraduate to the department head who later became MIT's president, and that its clean rooms are under-used)
Eon Systems (Wissner-Gross's company, which he says announced the first multi-behavior fruit fly upload)
Cerebras and Positron (Named by Blundin in his week's notes on inference-time acceleration)
Menlo Ventures (Cited as an example of the clear venture mandate behind Anthropic, against OpenAI's mixed investor base)
Amazon (Anthropic's hosting partner; Blundin's point is that Anthropic is not competing with its own data-center partners)
XPRIZE (Runs the Build with Gemini prize the host describes, with 26,000 registered entries and $2 million in prize money)
TKS (The high school program Ismail is involved with, which he offers as a model for rebuilding education)
Books & Resources Mentioned
"We Must Pace the Frontier" (The 3,800-word essay the episode opens on, with its three-step proposal for embedded evaluators, domestic coordination and international coordination)
Anthropic's September threat intelligence report (Documents malicious use of Claude across seven categories and contains the bioweapons-threshold sentence the panel focused on)
Paul Christiano's personal statement on joining the OpenAI board (The warning the host read out, including the 18-month estimate for fully automated AI research)
MIT's report on AI in teaching, learning and research training (Published 13 August; the source of "cognitive surrender" and the report Blundin read cover to cover)
Vending Bench 2 (The benchmark behind the $500-to-$15,515 result the host cited)
OECD PISA (The 15-year-old assessment behind the reading and mathematics declines in the final story)
Amusing Ourselves to Death – Neil Postman (Blundin's source for the nine-hour Lincoln debates and the compression of political argument)
Murder on the Orient Express – Agatha Christie (Wissner-Gross's model for a week in which everyone had a different motive and all converged on the same target)
The Innermost Loop (Wissner-Gross's newsletter, which the host says he reads on the treadmill)
Fast Times at Fairmont High – Vernor Vinge (Wissner-Gross's picture of what an automated undergraduate education would look like)
Watch the full episode:
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: