AI Kills Everybody or Doomer Psyop? OpenAI’s Math Breakthrough, Nike’s $200B Collapse
Chamath PalihapitiyaJason CalacanisDavid SacksDavid Friedberg
- Sacks argues that the viral Anthropic “whistleblower” resignation was an orchestrated doomer operation, not a spontaneous act of conscience. Jacob Coxon’s nearly blank account suddenly reached roughly 110 million views; three doomer policy groups amplified it within 10–15 minutes, and The Wall Street Journal’s embargoed story appeared before the tweetstorm. “Show us the data... all he’s got is vibes.”
- The investor question is what the episode does to Anthropic’s IPO, which Polymarket put at 88%. Chamath argues that alignment lead Evan Hubinger’s public co-sign — including his claim that extinction risk is personally “over 10% within the next decade” — is more consequential than Coxon’s resignation. Anthropic must either disavow the claims and risk an internal revolt or accept the product-liability and disclosure implications, with investors demanding a major discount.
- Sacks calls it a “time for choosing.” Anthropic’s implicit pitch — frontier AI is dangerous in everyone else’s hands but safe in its own — is, in his view, self-serving regulatory capture. If it agrees with Bernie Sanders that development should be frozen, he asks how it can IPO at a trillions-level valuation; if it rejects the doomer claims, its internal safety cohort may revolt.
- The panel attacks the doomer track record and argues that the ultimate target is open source. Sacks counts GPT-2, reasoning models, catastrophic cyberattacks, and mass job loss as a 0-for-4 record. He argues that an “FDA for AI” could require centralized monitoring and rollback that open models cannot provide, effectively banning published weights and creating a government-backed duopoly.
- Chamath, not Friedberg, makes the core counter to the precautionary principle: recursive self-improvement needs only power, chips, and an internet connection, so a US ban would leave the country behind while development continued elsewhere. Friedberg separately frames doomerism as a recurring social-panic pattern, while Jason argues that existential-threat narratives have historically been used to centralize power.
- OpenAI’s Navier–Stokes result is presented as brute-force leverage rather than supernatural intelligence, while the discussion raises serious data-leakage concerns. Friedberg estimates 130 billion output tokens across 10,000 agents as roughly 50,000–500,000 human work-years. Chamath calls zero data retention “Swiss cheese” and recommends sovereign infrastructure such as a controlled VPC or bare-metal deployment; he predicts some CIOs will be fired for careless API deals.
- Jensen Huang called Coxon’s claims “outlandish and deeply untrue,” prompting Sacks to ask why Anthropic has not said the same. Sacks characterizes Anthropic’s response as mealy-mouthed and interprets its reluctance to disavow the claims as evidence of internal and strategic tension; that remains his interpretation, not an established fact.
- Nike’s removal from the S&P 100 after 18 years is attributed to channel destruction, product deterioration, and disputed marketing choices. Jason calls the Kaepernick campaign a turning point; Sacks criticizes the broader “woke” repositioning; Friedberg says the shoes themselves deteriorated. Chamath’s proposed recovery is a return to “mastery and excellence,” the aspirational North Star he says made Nike powerful.
1. The resignation heard round the internet — and why the numbers don't add up
- The trigger: Jacob Coxon, a researcher who spent three years across OpenAI and Anthropic but only six weeks at Anthropic before quitting, posted that “the people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt,” accusing both labs of “racing straight to self-improving superintelligence and gambling with our lives.” Jason notes he had not seen roughly 150 million views on a tweet since Elon’s cocaine-in-Coca-Cola joke.
- What lit the fuse legally and virally: Evan Hubinger, who leads alignment science at Anthropic, replied, “Jacob is correct here. We really do earnestly believe AI could kill all humans! I personally think it is over 10% within the next decade. We do not yet have a plan to solve alignment.” The two posts reached roughly 200 million combined views, and Bernie Sanders cited them in connection with legislation to ban superintelligence; JB Pritzker called to “sound the alarm louder.”
- Chamath asks how someone with almost no followers or prior activity could appear on a Monday and reach 110 million views on Tuesday. Coxon had contacted Jason about coming on the podcast, then canceled that morning, which Jason frames as possibly avoiding cross-examination by Sacks.
2. Sacks' forensic case for a coordinated op
- Sacks’ evidence chain: a blank or scrubbed account was amplified within 10–15 minutes by three organized doomer groups — Nathan Calvin, general counsel at Encode AI, the head of policy at the AI Policy Network, identified in the discussion as likely Peter Wildeford, and Daniel Kokotajlo of the AI Futures Project, who released a Rogan episode using similar language at the same time. Sacks says all three groups are funded by Jaan Tallinn, an EA mega-donor and Anthropic Series A co-lead. Later, he also names Dustin Moskovitz and Sam Bankman-Fried in connection with the Series A.
- The apparent clincher is The Wall Street Journal’s story appearing minutes before the tweetstorm. Jason explains that this likely means the Journal was briefed under embargo and published at the wrong time. Sacks says the story was “teed up,” not spontaneous.
- Sacks challenges the “whistleblower” label: “What evidence has he brought forward that we didn’t have? This is all just vibes... Show us the data, show us the report, show us the leaked information.” He also highlights the anomaly that, unlike in a typical whistleblower case, people inside the company amplified rather than rejected the claims.
- His broader claim is that overlapping economic, political, and ideological interests may have converged around the same story. He does not claim Bernie Sanders personally participated in creating it; he says Sanders may have received it quickly through existing relationships because it served his agenda.
3. The endgame: an FDAI that exists to kill open source
- Sacks says the groups want a federal department of AI or AI regulator that can impose their preferred framework. He argues the mechanism for banning open source will not be called a ban: regulators will apply the same standards to open and closed models, then require central monitoring, control, and rollback — requirements he says are technologically impossible once open weights are published.
- The coalition he sketches includes ideological EA and rationalist groups that favor centralized solutions, politicians such as Sanders who gain power from public fear, and OpenAI, which Sacks says is running in Anthropic’s regulatory slipstream. A government-created moat could produce a duopoly with economic benefits for the labs and power benefits for regulators.
- Sacks’ COVID counterfactual imagines the trust-and-safety regime applied to personal AI. Someone asking an AI about vaccine risks might receive the government’s official answer rather than an independent assessment, producing a more centralized and dystopian information system.
- Friedberg adds that open source has no single corporate entity to make regulatory submissions or participate in the proposed review process. Yet he says open source can reduce AI costs by 50x and spread the benefits broadly. Sacks’ warning is that whoever controls the regulatory “gas pedal” can declare that open source has not complied and prohibit it.
4. Friedberg's macro frame: hysteria; Chamath and Jason on control
- Friedberg compares the current AI panic with Al Gore’s An Inconvenient Truth and IPCC forecasts that he says have been disproven, Fauci’s COVID warnings and resulting lockdowns and spending, and Three Mile Island–era opposition to nuclear development. He contrasts the estimated cost of a US nuclear gigawatt — roughly $15 billion — with about $4 billion in France and $1 billion in China.
- He says social panic becomes self-reinforcing: data centers poll at negative 80 because people are frightened of what they do not understand. Humans, in his framing, are primates facing an unfamiliar frontier and instinctively treating it as an existential threat.
- Chamath’s separate argument is that recursive self-improvement needs only adequate power, chips, and an internet connection. A US ban would not require approval from every government; someone elsewhere could continue, leaving the US like “the tribe in the Amazon that has never seen civilization.”
- Chamath also says the core motivation behind the regulatory campaign is control: give some agency, senator, or regulator the power to decide how AI develops.
- Jason broadens that into a theory of power structures. He argues that many systems of authority have used an existential-threat narrative — “you will all die; give me the power to protect you” — to centralize control.
5. Anthropic's impossible position: Philip Morris with an S-1
- Chamath’s securities-law framing draws on IPO examples. He recalls Google’s Playboy interview appearing during its quiet period and his own CNBC comments about Slack nearly disrupting Slack’s IPO, forcing the company to print the comments and address them in the S-1.
- His analogy is Philip Morris: Anthropic appears to be saying that its product could be extraordinarily dangerous while still taking the company public without fully resolving how that danger is disclosed. “No company’s trying to raise money by also talking about how they’re trying to destroy humanity,” he says, and argues there is no clear legal precedent for the situation.
- Sacks formulates the contradiction: public investors are being asked to underwrite a company at a valuation in the trillions while its own safety lead says the core product is unsolved and potentially civilization-ending. “At a minimum, that is the mother of all product-liability lawsuits,” he says.
- Chamath says an ordinary risk factor — NVIDIA may withhold chips, or competitors may emerge — is not equivalent to employees publicly assigning a greater-than-10% extinction probability to the product.
- Sacks and Jason agree that the S-1 may need to address the tweets. Sacks says Hubinger’s endorsement is more significant than Coxon’s resignation because Hubinger remains a senior alignment executive who co-signed and amplified the claim.
6. Why Anthropic can't disavow him
- Sacks’ “time for choosing” is that Anthropic must either renounce Coxon as an entry-level employee engaging in hyperbole and science fiction, or agree with the substance. If it agrees that frontier AI is civilization-ending, Sacks asks how it can IPO at a trillions-level valuation or reject Sanders’ call to freeze development.
- Sacks also attacks Anthropic’s implicit solution: frontier AI is too dangerous in everyone else’s hands but safe in Anthropic’s “virtuous and wise” hands. He calls that self-serving and says the company is seeking a monopoly or duopoly through regulatory capture.
- Chamath says Anthropic cannot simply disavow the claims because rational people are building a real company while a large enough internal cohort genuinely believes the existential-risk argument. Disavowal could trigger an internal revolt.
- Jason lays out three possibilities: the employees believe the claims and are right; they believe them and are wrong, perhaps with “some level of psychosis”; or the story is part of a coordinated effort to regulate AI, ban open source, and pull up the ladder. He asks whether Coxon is participating or is merely a useful idiot.
- Sacks resists treating the entire episode as one conspiracy. His view is that there was an organized amplification campaign, followed by different actors pursuing overlapping economic, political, and ideological agendas.
7. The doomer track record: 0-for-4
- Sacks runs through prior claims: GPT-2 was too dangerous to release; reasoning models were too dangerous, including the model associated with what Ilya Sutskever saw; AI cyberattacks would bring down the banking system; and Dario had forecast that roughly half of entry-level white-collar or knowledge-worker jobs would disappear, alongside 10%–15% unemployment.
- Sacks counts these as 0-for-4 so far. He says cyber risks remain real, but the best defense is AI-powered cyber defense, and he argues that the job-loss claims have not materialized; the economy has instead experienced job gains.
- The panel’s running joke is that basic applications still fail: Python autocomplete is imperfect, and booking a hotel room through AI remains unreliable. The gap between those failures and human extinction is presented as an argument against skipping the intermediate steps.
8. “How Do We All Die?” — steelmanning extinction and why it falls apart
- In Friedberg’s game, the panel tries to construct a 10% human-extinction scenario. Jason proposes a Terminator 2 path in which Claude is integrated into military systems and autonomous weapons. Friedberg proposes internet-connected bioreactors and robots using precursor materials to create and distribute a virulent airborne agent. He acknowledges that this requires many intermediate steps.
- Friedberg also considers the shutdown of financial networks, but notes that financial institutions maintain hard backups, air-gapped systems, redundancy, and human-in-the-loop procedures.
- Sacks says the cyber and biological scenarios involve real misuse risks, but the same technology could produce cyber defenses, antidotes, prevention, or cures. He also points to legal penalties and the fact that many capable people do not commit crimes even when they possess the tools.
- Sacks says the strongest doomer argument is not an ordinary cyberattack but recursive self-improvement: an AI becomes capable of fully automating AI research, creates its own training run, produces a stronger model, and repeats the process without a human in the loop.
- The counterargument from the panel is that current systems remain far from removing humans from AI development or even mundane applications. Jason says human approval points — such as requiring someone to press a button before a run continues — can be built into the software.
- Friedberg’s structural argument is that humans operate on human timescales. Even if an AI could analyze a decision instantly, people still need to wake up, meet, approve actions, and perform physical tasks such as bringing a toiletry kit upstairs. Sacks invokes Bill Gurley’s argument that the important question is to identify every intermediate step and the interventions available before reaching an extreme outcome.
9. Jensen breaks the silence mid-show
- During the recording, Sacks reports that Jensen Huang, speaking at a Goldman Sachs conference, called Coxon’s comments “outlandish and deeply untrue,” and described them as wrong, arrogant, and ignorant of safety work across the industry.
- Sacks asks why Jensen can say that while Anthropic cannot. He characterizes Anthropic’s public statement as “mealy-mouthed,” meaning vague and noncommittal, and argues that Anthropic’s reluctance to disavow Coxon is consistent with its internal beliefs and regulatory strategy. That is Sacks’ interpretation.
- Jason contrasts this with an earlier exchange roughly three weeks before, when Dario discussed claims raised by Gavin on the podcast. Jason also reads a response from Anthropic chief brand and communications officer Shasha, dated August 14, saying, “Dario has never said this. Complete and utter nonsense.” Jason speculates that Anthropic’s quiet-period status may have changed, but the transcript does not establish why the company is silent now.
10. Navier–Stokes: brute force in a lab coat
- OpenAI claimed a result involving the Navier–Stokes equations, written in the 1800s and used to model fluid dynamics. Jason calls it one of the seven hardest mathematical problems in the world.
- Friedberg’s deconstruction is that the reported 130 billion output tokens across 10,000 agents represent enormous distributed brute force, not a supernatural insight. He estimates the equivalent at roughly 50,000–500,000 human work-years, while noting that even an order-of-magnitude reduction still leaves tens of thousands of years of human knowledge labor.
- His conclusion is that AI is an engine of leverage. The agents’ messages and work are documented and understandable to humans; the system can reduce years of aircraft-wing, engine, or energy-system design work to minutes or seconds.
- The controversy is whether researchers’ use of OpenAI products helped the model make progress. OpenAI said, “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.” Sacks believes Noam Brown’s denial that anyone examined the researchers’ prompts and considers it more plausible that OpenAI learned researchers were making progress and applied substantial compute to the problem. Jason says Sam Altman confirmed that OpenAI knew Anthropic was getting close.
- Sacks still sympathizes with the researchers’ concern because OpenAI is both the provider of the models and a competitor developing breakthroughs in the same applications.
11. ZDR is Swiss cheese: the sovereignty trade
- Chamath says zero data retention is a commercial best-efforts commitment, not a guarantee. Even if a customer asks a model not to retain a protein-design process, other interactions — such as clicking a Like button — may not be covered by the same promise.
- He says awareness of this problem is moving through public-company audit and risk committees, with help from firms such as EY and Deloitte. His prediction is that some CIOs will be fired after executives discover that a standard API deal allowed sensitive information to enter a model ecosystem.
- Chamath’s recommendation is not necessarily open source or on-premises hardware. It is a controlled sovereign deployment: a customer-owned environment or VPC, potentially using AWS, Nebius, or Fireworks, with models provisioned on the customer’s terms. He says sensitive enterprises should not rely on a standard rate-card API without understanding the leakage risk.
- Friedberg describes asking a frontier model novel scientific questions that it identified as new insights, then later seeing a different account or model version describe the same ideas. He presents this as anecdotal evidence that prior conversations may have influenced training, while acknowledging that he cannot prove the mechanism.
- His key distinction is that de-identification removes personal or company identifiers, not necessarily the underlying method. “The approach is the IP.” He has ordered two Mac Studio M5s to run local open models and move roughly 90% of sensitive work away from Claude. Chamath cautions that local hardware does not by itself solve collaboration, memory, or shared knowledge-base requirements.
- Sacks broadens the issue to legal privacy. He says AI chat data currently receives weaker protection than email: a subpoena or court order may suffice where email would require a probable-cause warrant. Because people use AI as a lawyer, doctor, or therapist, he argues that AI conversations should receive at least email-level protection.
- Sacks also notes that closed-model providers are entering the vertical applications built on top of their platforms. He cites Claude Code and the resulting conflict with Cursor as evidence that customers cannot assume the model provider will not compete with them. “Maybe that’s the law we should pass first before we regulate.”
12. Nike: how to woke your way out of the S&P 100
- After 18 straight years, Nike was removed from the S&P 100 and replaced by Palo Alto Networks. Jason cites a peak market cap of $264 billion in 2021, peak revenue of $51 billion in 2024, an 80% decline in the stock from its peak, a 30% decline in China sales, and eight consecutive quarters of decline there. Chinese brands Anta and Li-Ning, along with HOKA, On, and Brooks, have gained share.
- Jason attributes part of the damage to CEO John Donahoe’s aggressive direct-to-consumer strategy, which alienated retail partners and gave competitors shelf space. He also criticizes a reorganization from sport-based divisions such as basketball, football, and tennis into men’s, women’s, and kids’ categories.
- Sacks calls it another “go woke, go broke” example and criticizes Nike’s movement away from athletes who embodied performance and excellence. He attacks the Kaepernick and Dylan Mulvaney campaigns as inconsistent with the brand. Jason, rather than Sacks, calls the Kaepernick campaign the turning point in his own view.
- Friedberg says the decline was not just advertising. He says Nike shoes began falling apart within six weeks, prompting him to switch to Brooks. Brooks, owned by Berkshire Hathaway, has achieved nine consecutive years of double-digit revenue growth and reached $1.6 billion in revenue. Friedberg attributes its approach to Buffett’s instruction to make the product better every year.
- His diagnosis is that Nike shifted from product to narrative, while Brooks kept improving the product.
13. The comeback thesis: mastery and excellence as North Star
- Chamath says Nike’s North Star was “mastery and excellence embodied through athletics.” He cites Michael Jordan, Tiger Woods, Serena Williams, Pete Sampras, and Roger Federer as people whose performance made the products aspirational.
- Chamath says he switched from Nike to On partly because Federer represented excellence and mastery. He argues that wanting to become better by following exceptional people is healthy and should not be shamed.
- His recovery playbook is to restore that North Star, return to retail locations, and sell the product broadly. If Nike again stands for mastery and excellence, he believes tens of millions of customers could return. He calls the brand a “coiled spring” with substantial potential.
- Jason suggests Nike should also revisit performance-oriented devices and communities such as Strava, while the episode closes with his joking “Do it or don’t” mock rebrand. Sacks calls it the one idea capable of making Nike’s situation worse.
Full transcript
Speaker 0: The PSYOP canceled on us? Is that what happened?
Jason Calacanis: Yeah, the PSYOP canceled. He couldn't take the smoke.
Speaker 0: He did not want to be cross-examined by David Sacks.
Jason Calacanis: I haven't used my legal degree in 25 years.
Speaker 0: Oh, man. I was up all night doing research on this.
Speaker 2: We can still talk about it.
Speaker 0: I'm so ready, yeah.
Speaker 3: Let your winners ride.
Jason Calacanis: Rain Man, David Sacks.
Speaker 3: And instead, we open-source it to the fans, and they've just gone crazy with it.
Speaker 2: Love you, S.
Jason Calacanis: Sacks, we have to talk about Anthropic. A researcher just quit over AI fears that he thinks could kill us all. Jacob Coxon worked as a researcher at OpenAI, then Anthropic, over the past 3 years. He only started at Anthropic 6 weeks before resigning.
After quitting, he posted the following to X: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.”
Neither OpenAI nor Anthropic is, ostensibly, acting responsibly. They're not acting responsibly, Sacks. “They are racing straight to self-improving superintelligence and gambling with our lives.”
This went insanely viral. I haven't seen a tweet go to 150 million views since Elon said he was going to put the cocaine back in Coca-Cola, which I think was the number-one tweet of all time.
Speaker 2: But Jason, can you actually explain to me how somebody who doesn't have any followers or an account shows up on a Monday and gets 110 million views on a Tuesday? How does that algorithmically happen?
Jason Calacanis: That's a great question. We're going to get into it. Evan Hubinger, who leads alignment science at Anthropic, responded to Jacob:
“Jacob is correct here. We really do earnestly believe AI could kill all humans!” Exclamation point. “I personally think it is over 10% within the next decade. We do not yet have a plan to solve alignment.”
These 2 posts had 200 million combined, and the story immediately jumped. As I mentioned last week, these things are now jumping from X and the AI community to the nightly news in under 24 hours.
Bernie Sanders, never the one to waste a crisis, said, “Posts like this are the reason he's proposing legislation to ban superintelligence and pause AI development.” Governor JB Pritzker replied to this post, saying, “It's time to sound the alarm louder on reining in AI.”
Sacksy-poo, as our former czar of AI, do you share everyone's concern here that we're all going to die by 2030?
David Sacks: No, I don't. Look, there's nothing new here. This is the same doomer histrionics that we've been hearing from this crowd for a long time.
The media's calling him a whistleblower, but what evidence has he brought forward that we didn't have? This is all just vibes from him. Show us the data, show us the report, show us the leaked information that the public didn't already have. Show us the facts, show us the evidence. We don't have any of that.
In fact, as Chamath was alluding to, this appears to be an op. How do we know that? Well, first of all, this is an account that had almost no activity, almost no followers, and no posts prior to this resignation tweetstorm. If he did have prior posts, they were scrubbed.
So this was a completely blank account, and somehow, within a day, it found this huge audience. Elon has said it never happened before. We also know who amplified it in the first 15 minutes. There were basically 3 well-organized, well-funded doomer groups that amplified it.
Nathan Calvin, who's the general counsel at Encode AI—they're the group that basically pushed for SB 53 in California. They're pushing for all these state-level AI laws.
Peter Will defer, who's head of policy at the AI Policy Network, which is pushing for all these federal regulations and a federal department of AI.
And then Daniel Kokotajlo, who's head of the AI Futures Project, just happened to drop a Rogan episode at the same time in which he used the exact same phraseology about how AI's going to kill us all.
All 3 of these groups are funded by Jan Tallon, who is one of the top few EA mega-donor doomer types, and who also happens to be a co-lead of the Anthropic Series A.
This is a tight-knit community that jumped on this tweet. This was not some spontaneous act by an employee who said, “Oh, I'm just going to post my resignation letter on Twitter.” This was something that was orchestrated.
And then just the final piece of evidence on this point is that someone looked at the Wall Street Journal story that covered the resignation letter, and somehow it was posted minutes before the tweetstorm itself. So the Wall Street Journal was given—
Jason Calacanis: It was under embargo. That's what it means. The technical term is they were briefed under embargo, and they screwed up the publish times.
David Sacks: Right. And this guy, Jacob Coxon, reached out to you, JCal, and said, “Hey, I want to come on the pod.”
Jason Calacanis: Yes.
David Sacks: Did he not? And then he canceled this morning because maybe he figured out it would be a tough interview.
If he had come on the pod, the questions I'd want to ask are: Who connected you with the Wall Street Journal? This was a relatively low- to mid-level employee who'd been there for, I've heard different estimates, 6 weeks to 3 months. Let's say 3 months.
This is not something that would be covered by the Wall Street Journal in this way. This is something that was teed up. He's clearly working with these doomer groups. They're going for maximum amplification.
Again, he's being called a whistleblower, but what have you blown the whistle on? Where's the evidence? Where's the data? We'd love to see what you saw inside of Anthropic if somehow there's a problem there, but you won't tell us what that is.
In most whistleblower cases, you get the company pushing back. They say, “Well, no, it's not true,” or they want to refute what he's saying. In this case, they're amplifying it themselves. You've got this—
Jason Calacanis: A functional company, Sacks, would disavow this. Wouldn't a functional company disavow this immediately, Sacks?
David Sacks: Here's the contradiction: he's—
Speaker 2: Well, it's very complicated.
David Sacks: He's actually amplifying their message. You saw that when his boss, or the person who's running this safety team, said, “Yeah, we all agree with you.” He then put on this greater-than-10% chance of extinction, which, when you put a number on something that frankly is just vibes, takes the message to a whole other level.
Speaker 2: 2 questions for you, Sacks, and then I have a bunch of comments and statements. This thing is crazy.
Question number 1 is, let's steelman the other side. What is the game theory on the PSYOP? How do you think it is supposed to go from here, had it not been exposed? What do you think they thought would happen?
David Sacks: Well, I think they know exactly what's going to happen. They're going for maximum amplification.
Given that this guy was at the company for a maximum of 3 to 4 months, it's entirely possible to me that he joined knowing that he would put together this campaign and be—
Speaker 2: Yeah. So what's the point of the campaign?
Jason Calacanis: So, premeditated campaign. What was his goal? Is he a rogue agent?
Speaker 2: What's the goal?
David Sacks: Oh, it's to shape—
Jason Calacanis: Is he rogue, or is he acting in coordination? That's the question I have.
David Sacks: These groups are interested in shaping the public perception of AI. This is what they do.
Jason Calacanis: To what end?
David Sacks: They're creating a very negative perception of AI to further their case for AI regulation. The end goal here is to create a federal department of AI. They want an AI regulator. Some of them also want a pause. Some of them also want a ban on superintelligence. Well, that's what Bernie Sanders' bill puts forward.
But all of them want a new federal regulator for AI, and that is what Dario himself has called for. Again—
Speaker 2: That they can control, basically.
David Sacks: It's a federal department of AI, which invariably they will control, and they will basically then be able to push their preferred frameworks for regulation through this new regulatory body.
Speaker 2: Okay. I have another question. Now you have this very complicated moment where you have internal employees validating an employee who just left, making claims—let's call them whistleblower claims—that create, at a minimum, product-liability issues, but at a maximum may create false-representation issues.
You have an active S-1 process underway. How are they to handle that? I'll give you a couple of historical examples. You guys remember when Google was going public? Friedberg, you probably lived it, but there was an interview that I think Sergey or Larry, or both of them, did with Playboy. That was well before the IPO window, and then they published it in the middle of their quiet period.
There was all kinds of noise about how that could have derailed their IPO and that they would have had to refile.
In a separate example, when Slack was going public, I was on the board, and I was on TV with CNBC. We were talking about Tesla and SpaceX, and I made a couple of comments about Slack—positive, but not saying anything that was crazy, just about the network effects of Slack.
Jason Calacanis: It wasn't super effusive. I remember it now.
Speaker 2: That almost stopped our IPO. We had to print it out, put it back into the S-1, and disclaim these things.
This seems so beyond that because this is so beyond the pale of what these folks are claiming. What does it do to their IPO process? Now, Sacks, you jokingly said, “Stop the IPO,” but what is the SEC supposed to do? What are the—
David Sacks: Well, let me—
Speaker 2: What are the lawyers supposed to do?
David Sacks: Let me explain my tweet, actually.
Wait, are you joking or are you serious in that tweet?
I was tongue in cheek in that tweet, and that's why—
Okay.
I put the scare quotes around “whistleblower” because this is not an ordinary whistleblower. He appears to be doing this with the full cooperation and support of many of the employees, including his own boss. I don't know if he's doing it with the senior leadership of the company, but again, this message is fully consistent with the doomer/regulatory-capture agenda of Anthropic.
Again, he's not saying anything that is incompatible with or detrimental to his boss, and that's why he's getting support from within the company. But it does raise these questions that Chamath is talking about, and I do think that there's a tension here in Anthropic's position that implicates their IPO process.
The tension is this: On the one hand, you're asking public market investors to underwrite your company to a value in the trillions. On the other hand, your own safety lead, not Jacob Coxon, the boss who then sort of co-signed his tweet, is saying that your core product is unsolved and potentially civilization-ending.
At a minimum, that is the mother of all product-liability lawsuits. Is that the type of liability you can just wave away with a little disclosure in the S-1?
Sacks, it's a very important point for us to pause on. If they are saying this thing is incredibly dangerous, and they continue to update it and release it, they're saying that if something goes wrong in your enterprise or in your life—if it does a major hack, if somehow it figures out how to hack Bitcoin or something and compromises the miners—they're liable for some giant hack that occurs. That's an incredibly important point for us to meditate on here.
Look, I think this is a time for choosing for Anthropic. I think that these games have to end. I think that they have to choose between either renouncing this “whistleblower” as a doomer op. He's basically an entry-level know-nothing who's engaging in hyperbole and science fiction. It's all vibes. In other words, there aren't facts and evidence behind this, right?
So they can say that, or they can agree with him. In which case, Bernie Sanders may be right that further development should be frozen, in which case I don't see how they can IPO now in the trillions. Because that is where these doomer arguments lead: directly to Bernie Sanders.
That's the reason Bernie Sanders quote-tweeted this guy. That's why Bernie Sanders introduced a bill to stop further development, to ban artificial superintelligence. But if that is where Anthropic says it's going, and they agree with Bernie Sanders that it's dangerous, how can they be allowed to IPO at this point in time? Why would any public-market investor want to invest in that?
So it's fundamentally a contradiction, and it's in tension. I think that historically, the way that Anthropic has solved this tension is they've basically said, “In anybody else's hands, frontier AI is too dangerous to be developed, but in our virtuous and wise hands, it's safe. We will make it safe. We're the only ones you can trust.”
Everybody else in the world can see that that is the most ludicrous, self-serving argument. We're not buying into it. We're not going to grant you a monopoly on frontier AI or a duopoly through this regulatory capture. So you have to give up on that game. It's patently ridiculous.
I think, again, that's sort of their way out. But if that's not going to be sustainable, I think they now have to choose—
Hmm.
—you’re either on the Bernie Sanders side of this thing, in which case, why are you IPO-ing? It's time to stop. Or you renounce the argument that Jan Leike is making, that it is hyperbolic—
They can't do that.
—that there's no basis for it.
What you saw yesterday showed that they're not in a position to do that, even if they want to. I think that there are very rational, good people who work at Anthropic building a great company. The problem is that that's in conflict with a philosophy held by enough of a group of people inside that business that they can't disavow him, Sacks, because then this entire cohort of people will go absolutely bonkers and get up in arms.
This is such a good point, Chamath. Yeah.
There are all kinds of people right now quote-tweeting this guy, essentially saying from inside both closed frontier companies that this is happening.
Yes.
And so I think it is a complete hornet's nest that they have kicked to the ground, because on the one hand, you have rational people who want to build a business, for which you need enormous amounts of capital that is only available in the public markets. But the process of doing that requires you to underwrite risk, absorb liability, disclose that properly, and then allow people to know what they're buying.
What we have is this weird version now where it's almost like an example of Philip Morris, where we know that the cigarettes are bad for you, we know that the cigarettes will kill you, and we're going to say it, but we're not going to disclose it, and we're going to take the company public and allow you to own the stock, and then we're just going to figure it all out later. This is not a tenable position for Anthropic.
Let me get Friedberg involved here. This is a very good analogy you bring up. Find the video where the executives at the tobacco companies are asked at a congressional hearing if they believe that nicotine is addictive, and one after the other, Chamath, they say they do not think it's addictive.
That happened 10 or 20 years after they had the research and hid it. This was the movie The Insider. You're 100% correct. You now have legions of people inside OpenAI and Anthropic, Friedberg, agreeing that we're all going to die in 2030. They're not just saying everything's going to be hacked anymore. They're saying there's a 10% chance of everybody dying, and then other people are saying, “Yes, that's probably pretty accurate.”
I'll give you some other times when we've heard similar things. Al Gore published a movie called Inconvenient Truth, and he used these IPCC forecasts that, as of now, have been disproven, and many of the assumptions in those forecasts were not correct.
A guy named Fauci—you guys may remember him—told us all that COVID would kill us all, and as a result, we needed to lock down the entire world. So everyone was locked in their houses, businesses were shut down, and the government began spending wildly to keep people economically active, which had the adverse effect of driving inflation through the roof. We now have a spiraling debt and spending problem in this country.
There was an era when Three Mile Island had an accident, and we hysterically said, “We're going to destroy the entire country if we keep doing nuclear development.” Meanwhile, other countries raced ahead. The cost to produce a gigawatt of nuclear power in the US today is like $15 billion. The cost in France is $4 billion, and the cost in China is $1 billion.
I think we're in this hysteria phase of AI doomerism. Once you're in this phase, it's very hard to get out because it basically becomes the default social system. Everyone says we are facing an existential threat. We are all going to die if we don't do something about it, if we don't stop it.
The absence of proof that we're not going to die and that we have absolute security lets the doomsayers get away with the compounding effect of getting all of the social systems and social networks activated, which is where we're at right now. The majority of Americans—that's why data centers poll at -80—believe very deeply that we're going to die, and I think that there's a deep unconscious rooting to all this.
Humans are primates living in a cave, and we're deeply scared of what we don't know, what we don't see, and where we haven't been. Where we're going is a place we've never been, and it's scary.
In this era when people thought that the world was flat and you would fall off the edge of the world, it was scary to sail west because you would go off the edge of the world and you would never make it back. Then Columbus did it and hit some land, and lo and behold, the New World was discovered.
I think that we've had many moments like this in human history where we're kind of a pioneering species. But pioneering is not everyone's pioneering. The vast majority of people convince each other and convince everyone that we have something deeply threatening us, and we're always facing an existential threat. It is rooted in our beings as a species. It is deeply, deeply rooted. It's what's created survival instincts for us over generations and millennia.
But it takes just a little bit of pushing, and someone gets to be the pioneer, and the New World is discovered. Because you haven't been there, it's easy to say, “There are dragons. You'll fall off the end of the world. We're all going to die. Climate change will destroy us all,” and on and on and on: “COVID will wipe us all out.”
I think we're in that social-panic phase right now, and we're told we have to listen to the experts once again. In this case, the expert is an individual who has been employed for 6 weeks, has not had a real job for very long, and worked at an NGO whose literal job was to scare the shit out of people to stop doing AI development.
This individual went in and did 6 weeks of work, came out, and suddenly the whole world says, “This is an expert.” Bernie Sanders says, “This guy's the expert that, once again, we all have to listen to.”
Now, I think it's important to note one thing.
We could sit here and argue that we should follow the precautionary principle: the idea that if something is existential to us, we should take precautions, take our time, and slow it down. But the difference here is that, in order for recursive self-improvement—which is AI that continuously improves itself and its capacity—to reach that moment, you really just need an adequate amount of power, a set of chips, and a connection to the internet. You don't need to be approved by the US government or anyone else, because you could do that anywhere on planet Earth.
As we'll hear about at the summit next week, there are technologies now that are reducing the cost of power per token of output by 10,000x, and there's enough compute out there that's publicly available and floating around. Someone is going to stand up a data center or a set of computers—it doesn't even need to be in a data center—and they are going to develop a system that's going to recursively self-improve. It is going to happen.
So if the US wants to stop all, quote, "development" and make it illegal for people to use computers to effectively write speech—which is what software is, writing the speech—we can go ahead and do that. Then someone else will have this capacity, and we will be the tribe in the Amazon that has never seen civilization, because civilization will evolve all around us. Somewhere in the world that has this technology, it will evolve. It will have the capacity to advance, and the US will be left behind.
So you can have the doomerism and precautionary principle, but the reality is that, in the world today, we don't have the ability to control. At the end of the day, I do believe that the fundamental motivation is a system of control, because think about this. This guy's going on all these non-tech podcasts telling people they're going to die, and he's saying one thing at the end of the day, and so are Bernie and others: Give power to X to control the development of AI. Give power to someone to control the development of AI. Someone has to be given that power. Someone is being given that control.
All of them will have a different answer on who it is—some regulatory body, the senators, some new agency—but they are centralizing control and power to some individual or set of individuals.
Can we build on that point? Let's imagine we went through that whole COVID period with AI. AI happened a little bit after the whole COVID thing, but let's assume that AI had happened before and, furthermore, that we had a federal Department of AI. Imagine that whole trust-and-safety regime that happened during COVID, where social media was basically censored in cooperation with the federal government.
The White House told Twitter and Facebook that these accounts were dangerous and that disinformation was dangerous. People who had alternative or dissident viewpoints on the origin of COVID or the efficacy of the vaccines were suppressed. Imagine if that had happened during AI.
You'd be asking your personal AI model, "Hey, should I get the vaccine? What are the risks versus benefits?" Do you think it would tell you the truth? No. It would basically give you the official perspective, and the government would have pressured them through the federal Department of AI to do that.
The FDA for AI would have set standards for these companies, saying, "You can't allow disinformation through your models." That's one of the harms that we want to prevent. Think about how much more totalitarian that whole episode would have been if, instead of turning to social media or the web for our answers, we were getting them from our personal AI, and that personal AI was being influenced by the government to give us basically official answers. I mean, you can see how dystopian it's going to be, and that is where an FDA for AI will take us.
That is also the problem with open source, because open source doesn't have the individuals employed. Remember, open source is a community-developed effort, so community members contribute code and develop something. It might be a single company, but technically, open source can be forked, developed, and pursued down different paths. This is what happened with Android. There are lots of different forks of it, and so on.
But with open source, there is no corporate infrastructure to go and do the regulatory submissions and the sort of review that's being proposed with AI. That is the biggest thing to take note of, because open source is what drops the cost of AI by 50x and makes it available to everyone, so everyone can benefit from AI and no one gets rich. We all get rich with open source.
I don't own any shares in OpenAI or Anthropic, for full disclosure. I don't care if either one of them wins. I use both those products. I think they're both fantastic. Good for those companies, good for those leaders. I also use open source, and I think open source is the game changer because it does give everyone the ability to get wealthy and progress, make more income, and have more freedom to operate in the world.
With AI, I'm a big advocate for open source, but I do worry that this system of control—doing, quote, "regulation," doing reviews, slowing it down—means someone gets to decide when it gets to speed up again. That means someone is controlling the gas pedal, which means someone gets to say, "Open source is not submitting to the regulatory process; therefore, we will ban open source."
As soon as open source is banned, you have now created a monopoly that is in partnership with the federal government, and you now have a centralized system of global control that is realized from this sort of approach.
I'm not saying that AI shouldn't be tested, that we shouldn't confirm that AI is safe, that we shouldn't build safety systems, or that we shouldn't be responsive and thoughtful about where things are going. There are a lot of ways to do that, and we should do them all. But to create systems of control, centralize them, and get rid of open source is where this gets to, and that is not a world any of us should want to live in.
Well said. Well said.
That is where it's headed.
Yeah.
No, look, this is a doomer psyop, and the ultimate target is open source. That's what it all comes down to, because that is what an FDA for AI will ultimately ban. They won't call it a ban. They'll just say, "Look, we have to apply the same standards to open and closed models."
What are those standards? Dario has already testified before the Senate. He says that models are dangerous if you can't centrally monitor and control them and roll them back. That is technologically not feasible with an open model, because once the weights are published to the public domain, they're out there; they can't be rolled back. We're going to have to prevent those weights from being published, used, or hosted. That'll be the way that they deal with this.
That ultimately is the target, and you can see that there's an alliance forming, right? You have this doomer community, who are these highly ideological, sort of EA rationalist groups. They favor centralized statist solutions. Then you've got politicians like Bernie Sanders, who like the fear because, again, it grants them more power, so there's more power for the government.
And then you've got groups like OpenAI. Frankly, OpenAI seems to be running in Anthropic's slipstream right now from a regulatory-capture standpoint. I don't think they're as doomer as Anthropic, but they can see the benefit if the federal government effectively creates a moat. There are huge economic benefits to essentially creating a duopoly situation.
Some of these players are going to reap economic benefits, some of them are going to reap power benefits, and some of them are just latent totalitarians. They like the idea of socially engineering society, so they're on board with this. That's what's happening right here.
And the whole idea is to panic the public to stampede us into this FDA for AI.
I think what we're seeing here is either a conspiracy theory that is being debunked in real time, Sacks. We went through COVID. We've had other conspiracy theories that have come true. Twitter was being influenced by the FBI; that turned out to be true—the Twitter Files, all that good stuff.
Now you have a group of people in the public who are on the alert for a conspiracy theory. So what we have to ask ourselves is: Is this actually, Chamath, a conspiracy theory that's being debunked in real time by people like us and the broader public, who's starting to say, "Is this real or not? Are these guys performative or not?" Or do we actually have to ask the question: Do these people believe what they're saying?
It's one of these 3 scenarios, I think. They either believe what they're saying and they're right. So they believe it, they've seen something that we can't see because they're in the frontier labs, and this could end us.
The second scenario is they believe what they're saying and they're wrong, and they have some level of psychosis. That's my leading theory. I think these people believe what they're saying and have psychosis.
Or, to your point, Sacks, there are some—and this is scenario 3—who are involved in some coordinated psyop with the goal of banning open source, getting regulated, and pulling up the ladder behind them. And then you have to ask: Is Jacob part of this, Sacks? Or is he just a useful idiot in all this? So they're identifying who the people are who have the psychosis, and let's let them speak for the company.
But Dario—
No, look, it doesn't all have to be a conspiracy.
Yeah.
I think it's very clear, based on who amplified this post and when they did it, that there was an organized PR campaign here. This wasn't just a random resignation.
That's your theory, yes.
I think we can also say that these major doomer groups, by amplifying his tweet within 10 to 15 minutes of him posting it on an account that had no tweets and no followers, were alerted to this and probably participated, if not in the creation of the tweetstorm, certainly in the amplification of it.
The facts are aligning toward that, yes.
We know he pre-briefed The Wall Street Journal. But do I think that Bernie Sanders is in on it? No. I think probably what happens is someone in these groups is already working with his staff, or they have a conduit or a relationship, and they get it to him very quickly because it serves his agenda.
So what I'm saying is there's a lot of overlapping agendas here, where some people have an economic interest and some people have a power interest, and there's going to be a division of the spoils. In their view of the world, AI is going to be very centralized. We're going to have a duopoly.
The duopoly reaps the economic benefits, and they give big political donations, and they'll do the dirty work of suppressing speech because, again, on that side of the ledger, you can circumvent the First Amendment. And then on the government side of the ledger, you'll have those people who exercise control over our lives through the whole FDA. So you get this division of labor in our system between aligned groups.
Okay. Now, do you—
But look, I don't think we have to get into the conspiracy. I think we should just ask: What is these people's track record of being correct? I think it is worth asking that.
Again, these are the same people—let's just go through it. These are the same people who said GPT-2 was too dangerous to release. Okay, 0 for 1. These are the same people who said thinking models were too dangerous to release. Remember, what did Ilya see? He saw the first reasoning model. That was too dangerous to release.
Fable, cyber
0 for 2.
Yeah.
They said that this would cause cyberattacks that would bring down the whole banking system. That has not happened yet. There are cyber risks, to be sure, but the best way of dealing with those cyber risks is by using AI swarms for cyber defense—
Cyber AI. Yeah.
Yes, exactly. Look, the only way to solve AI cyberattacks is with AI-powered cyber defense. We all know that.
Then you look at the job-loss claims, where Dario himself said that by now we'd be at, I don't know—
Half of all—
Half of all entry-level—
White-collar jobs would be gone by the end of this year—
Entry-level and knowledge-worker jobs—
Into next year. Yeah.
Yeah, and overall unemployment levels of 10% to 15%.
Yeah, it hasn't happened.
There's no evidence of that. Quite the contrary, it's all been job gains. The economy's doing well. So again, what are we on now? 0 for 4.
These doomers just move from claim to claim, and as the last one gets refuted, they move on to the next one. Again, my point is: What has this guy told us that is new here?
We're still struggling to get autocomplete for Python working right.
You can't book a hotel room. That's actually the hardest one.
You can't book a hotel room.
Try to book a hotel room using AI. It does not work yet.
Python autocomplete is pretty good.
But it's going to kill us all.
It's almost perfect. But it's going to kill us all. Oh, my God.
Maybe DoorDash might get us a burrito.
If you guys dare to book me in a non-penthouse suite at the Radisson—
Are you—
Suite-shaming me again?
You're suite-shaming.
You know what I mean? This is a joke. No, it's a joke.
If you think back on the course of human history and you think about the moments where systems were set up that created power structures that didn't exist before, they were all predicated on the simple notion that there's an existential threat: You will all die. A deity, a god, or some powerful system will come in and destroy you.
Some weapon.
Right. And because I can protect you, you must follow me. You must do what I say. It is the common system of human power and control structures to tell a narrative of existential threat and leverage that narrative into your solution being the path to protection and safety.
That is what sells votes, but it is also what created all these early structures where early religious institutions ran all the countries, told all the people what to do, and were the power and command centers of the world. In many parts of the world, they still are.
When you talk about, "Oh, well, if you don't listen to me, God will damn you, God will shame you, God will destroy you, or some deity will come down and get you," this is the same psychological concept: There is a story being told that you will all die. There's a 10% chance you will all die, standing at the pulpit and banging hard, and if you don't let me control it, if you don't give me the power and the authority to control it, you will all die. Give me that power.
The only other time I can think of this happening would be nuclear weapons, which we did get ahead of. But putting that aside, a cyber hack is not going to destroy society. So how do you get to the point where 7 billion people, or 1 billion people, are going to die? Really, the only known scenario here is, I guess, that it creates a bioweapon and kills everyone, or—
Well, let's play a game.
Terminator 2.
Let's play a game.
Let's play your game.
Okay, so my game—
Do you have any music?
Okay, guys, this game is called How Do We All Die? In this game, we will go around and steelman the concept that there's a 10% chance that we all die from AI. What is your most likely path to human extinction, JCal, from AI? Go ahead. You're first.
Okay, I'll take human extinction for 800. You always wanted to be the host of Jeopardy! That would be a great moment for all of us.
I'm going to go with Skynet and Cyberdyne Systems. If people forget, at the end of Terminator 2, they got the arm and one of the chipsets. Then Cyberdyne Systems built for NORAD, which controls weapons and decisions to fire off nuclear bombs, and there was Miles Dyson. He's the Jacob, he's the Dario in Anthropic.
That engineer studied the parts that came back through time. In this case, this would be studying the LLM. Then he reverse-engineered it, and he empowered NORAD to build a defense system powered by AI. When it became sentient, it tricked the Russians and started the nuclear holocaust.
That's the number-one way we all die: They somehow use Claude inside the military to create autonomous weapons. It's the only way that this actually happens, in my mind.
Chamath, do you have a How Do We All Die theory? How do we all die for 800? The best example I could get is that there are some internet-connected bioreactors with internet-connected robots and access to all of the precursor materials, and the AI hacks the robots to take the precursors and use the bioreactor that it then controls, again, over the internet, to create some kind of virulent airborne mechanism that spreads rapidly and violently all around the world.
So it has to have an enormously long half-life, get into the jet stream, and go all over the world. When we inhale it, we die.
Mm.
That would need all of these other things.
So, contagion by robots and a sentient AI.
Again, you'd also have to have bioreactors and precursors, which—
Yes.
It's a couple of steps.
What do you got, Sax?
Well, the problem with these types of scenarios—first of all, I think it's kind of a dumb thought experiment to make arguments that you don't believe in. But the problem with the cyber example or the bio example is, yeah, is AI going to allow for misuse in new ways? Will there be new AI-powered cyberattacks? Could there be new viruses created using AI to create novel sequences? Yes, those are risks, but at the same time, you can use the same technology to create antidotes, prevention, cures, and all that kind of thing.
And so it’s like a lot of things where, yes, there are tools that can be used by bad actors, but there are a lot more good actors than bad actors out there, and it’s illegal to use them for bad actions. It was possible to break into systems and hack before AI, and a lot of people who could do it didn’t because they didn’t want to go to jail. Again, with the combination of the tools being used for good and the legal penalties for doing bad, these are things that I think we can control.
Part of the argument that I wanted to make in going through how everyone dies is that it is very hard to have a clean scenario, given the human-in-the-loop dynamic and the air-gap dynamic.
Explain both of those, please. Hold on. What’s your best answer, Friedberg? What’s the best thing you can come up with?
Yeah.
How do we all die?
I think shutting down financial networks, but there are hard backups. There’s a solve for this because a lot of the financial institutions already have nuclear-response scenarios where they make physical hard copies of everyone’s assets and records. They sit in different locations, and there are a lot of systems of redundancy that are air-gapped, designed not to be connected and not to be hackable, and that have a human-in-the-loop process where there’s an actual human process involved.
I think a lot of our critical systems are importantly set up in such a way that they have human-built redundancy for the notion of cyber defense, the notion of hacking effects, nuclear events, and all these sorts of things happening. My biggest thing is that if everyone lost all of their digital assets, maybe that’s a world that falls into chaos and people start to attack each other and whatnot.
But I think it’s very important to take note, as I’ve thought about this a lot, that there’s so much that’s human-in-the-loop. That’s also what makes AI so hard today to be successful from a productivity perspective. We all experience it: AI can do these amazing things, but if you need a human to do something as part of the process, AI doesn’t really do that much for you.
Let’s say you want to get a toiletry kit brought up to your room at a hotel. You tell your assistant on your phone, “Get me a toiletry kit brought up,” and the AI figures out how to connect to the front desk and call the guy and tell him. It’s all digital, it’s all automated, and AI does all these amazing things that it couldn’t do a few years ago, but you still need the guy to bring the toiletry kit up from downstairs.
Well, did it work? Did you get the amenity kit or not?
Well, this is the point. This is the hard part, the longest pole in the tent. It’s getting the guy to actually come upstairs and bring you the freaking toiletry kit.
The fucking autocomplete doesn’t work yet, guys.
I know, exactly.
Come on. We need to explain 2 terms to the audience. Air gap means it’s not connected to the internet; therefore, AI can’t get it. And then human-in-the-loop—the example would be a 2-key system that you see in every movie when you’re trying to launch a nuclear missile.
The human still needs to do something.
2 Navy officers have to put the keys in and turn them at the same time, based on reality.
You guys are actually not confronting the core argument that I’ve heard him make. If we want to really steelman it, the core argument that he’s making is about RSI. It’s about recursive self-improvement.
His argument is that the capabilities are improving at such a rate that very soon we’re going to be able to fully automate an AI researcher. In other words, we’re going to have an AI that does the job of an AI researcher, so the AI can create its own training run for the next model. When that model gets created, it will then train the next model, and so forth and so on. There’ll be no human in the loop, and it’s just off to the races. That’s basically the argument.
I think the counterargument that you guys are making by saying, “Look, we haven’t even got autocomplete,” is that we are nowhere near removing the human from the loop of AI development.
Or even AI applications is my point, Sacks.
Right.
Yeah.
We can’t get mundane tasks done yet, so how are we going to have a task that kills all of humanity? Even if you put it at killing 1 person, can you get it to kill 1 person?
Right. That’s also our insurance and our security against AI taking over the world. Even if you want to make a decision at your company, you still need a meeting of the decision-makers to make a decision. The AI can do all the analysis, but you still need to get the humans together.
They have to wake up, take a shower, drink their coffee, come to the office, sit down in a room together, and make a decision. So you have that 16-hour window between decisions that’s happening. Even if the AI is infinitely more intelligent than any human, humans are still running their day-to-day processes on a human timescale, and that’s what also protects us.
The part I don’t understand, Sacks, is—
Yeah.
If these geniuses are building such amazing tools, did they ever think to put in there, “Hit the space bar to continue,” or allow the human in the loop to say, “Yes, you may continue your run”?
It’s pretty obvious that if you try to let Claude or OpenAI use your browser, it’s like, “Can I open your Gmail? Can I submit the form?” And then it’s like, “I don’t want to book the ticket for you, but I’ll select the flight for you,” and then you have to book. You can build these things directly into the software. We’re acting as if we’re not building the software and we have no control. It’s a false premise.
Yeah.
No, but literally, we went through this with Dwarkesh’s post last week. OpenAI set up this recursive process: “Here’s a bunch of goals. Get better at it. Work with each other,” and they just watched it work. Like, okay, just watch it work. And maybe if there was a liability—
Yeah.
—that Anthropic and OpenAI are responsible, and the researchers who set these models into recursiveness, Sacks, those researchers themselves are responsible for what happens, and the company’s responsible for what happens. Along the way, they’re going to get—
Look—
—speeding tickets or be shut down.
Yeah, because there are 50 waypoints along the way to—
Yeah.
—Chamath’s computer-generated bioweapon being released into wet markets everywhere by robots.
Could you unplug it?
Well, no, they do experiments like this. The whole thing Dwarkesh did—
It’s like South Park. Cartman just comes and unplugs it.
Unplugs it.
Hmm.
Yeah.
No, but literally, we went through this with Dwarkesh’s post last week, which was that OpenAI set up this recursive process: “Here’s a bunch of goals. Get better at it. Work with each other,” and they just watched it work. Maybe if there was a liability—
Yeah.
—that Anthropic and OpenAI are responsible, and the researchers who set these models into recursiveness, Sacks, those researchers themselves are responsible for what happens, and the company’s responsible for what happens. Along the way, they’re going to get—
Look—
—speeding tickets or be shut down.
Agents working together as a swarm is so far removed from AI being able to do its own training run. These are just 2 totally different things. And look, it’s certainly something we should keep an eye on, but again, he’s going from this idea that we now have AI as a tool to make AI researchers more productive to, “We’re going to get to RSI maximalism, and then from there, everyone’s going to die.”
There are so many intermediate steps where you have to show how we would get there. What Bill Gurley said is, “Okay, wait, let’s slow down this argument, and let’s just talk about all the intermediate steps between those things and why they would happen, and then the interventions we could take if we ever got near that point to stop them from happening.” Then the argument starts to fall apart.
Yeah, because there are 50 waypoints along the way to Chamath’s computer-generated bioweapon being released into wet markets everywhere by robots.
Again, my example is so pathetic because, as Friedberg said, there are just so many steps between here and there.
That it's just so implausible.
And even if you got there, you haven't even established how everyone would die.
Yeah. And in a world where there's any other kind of model out there, the fact that it, as you said, Sacks, couldn't actually create a vaccine at the same time—
Yeah.
And someone else is gonna be doing it. And if you try and create one system of control for some of it, the recursive self-improvement scenario will play out somewhere on planet Earth. And when that happens, when that person or that organization or that country now has that capacity, they now have a huge technological advantage, and then it becomes very hard to catch up.
We've seen this time and time again with technological races. This is a moment that the United States would be deeply, deeply mistaken to write off.
And let me just say one more thing about open source that I didn't say earlier. I want people at home to understand what open source really means. With open source, you don't need a data center. You don't need a third-party billionaire getting richer.
All of these retorts about why AI needs to be stopped—it's for data centers, it's for rich people getting richer. If you get open-source AI, you can install it on your computer at home or on your phone, and you can run it on your phone for free. You now have an open-source AI agent model system that can do all sorts of things for you.
It can book your travel, it can do work for you, and it can make your life easier. It can do all these amazing things in terms of productivity. Open-source AI is critical for everyone to benefit from this technology and keep it from being centralized and controlled by a handful of companies and individuals and a small set of governing officials.
Open source needs to thrive, and if you give power and control systems over to governing people, open source will die, and you will end up with more power and more of an oligopoly than you have ever imagined.
Who's positioned here? What impact does this have? I'm being obviously a little bit playful here, but how does this impact the Anthropic IPO, which was one of your original points here, Chamath? There's an 88% chance they're gonna go public right now, according to Polymarket. So let's pull up that chart for a second here.
When you are in a quiet filing period, just to give you the technical understanding, there is an important back-and-forth that happens between you, the filing company, and the SEC. Now, what is the SEC trying to do? They are pressure-testing how you've written what your best understanding is of your business, and the reason they do that is that when the S-1 flips public, that is meant to be this anchoring document that's supposed to be an extremely accurate, then-current snapshot of all the opportunities, but also all the risks.
And so you now are in this very awkward position. Sacks's tweet is actually really important. It's: What are we all to believe? And are we to believe that there is something that's extremely dangerous hiding inside this company because this person who quit said it, and then it was amplified by a bunch of people who are currently there?
If it's true, now you have to go back and actually say, “Are these risks properly disclosed?” And if so, what should the investors do in reaction to it? And I'll tell you what they'll do. They'll demand an enormous discount, and the reason they'll ask for an enormous discount is the other part of what Sacks and Friedberg said, which is that it creates massive long-tail product-liability risk because you're effectively evading a known major risk.
Then there's the other path, which is you have to completely disavow this person. Now, the reason that they can do that is if they actually believe that this is a psyop. So that allows the IPO to go more smoothly, but then the other part of it is that you'll have a plurality of employees inside of Anthropic revolt because, unless they're lying, they actually believe that this is a huge risk. And if you disavow it, they're gonna feel enormous tension.
So I think that Anthropic leadership is caught in an extremely difficult situation. The SEC is caught in an even more difficult situation, which is: How do I deal with these disclosures now? No company is trying to go public and raise money—a very capitalist scenario—by also talking about how they're trying to destroy humanity. Nobody has ever done that, so there is no legal precedent for this.
There is a risk-factor section. Sacks, does this—
No, but as Sacks said, risk factors are extremely benign. These are superficial risks where you talk about, if this happens, then that could happen. If demand goes away, that could happen.
Yeah, but—
The typical risks—
Yeah.
Okay, just to explain. The typical risk, Jason, that they would have had would have been to say—
Competition.
NVIDIA may not give us the chips. There may be other models that come around the corner.
Regulation.
This is not a risk that anybody actually writes in there as a 10% chance because a bunch of employees have now said it's 10%.
Well, let me ask Sacks this. Sacks, with all these employees coming out saying all this kind of stuff, they represent the company, they're resigning on behalf of the company. Do you think the S-1 is being amended right now to include these chaotic tweets and—
I think they have to.
Yeah. Right?
I think they have to.
I think so too.
Because here's the thing: The key tweet, actually, from a legal standpoint wasn't Jacob Coxon's tweet of resignation. It was Evan Hubinger's tweet—
Endorsement.
—endorsing, co-signing.
Yes.
That's right.
Because this guy is still a senior executive, or senior alignment executive, at the company. He's managing one of these safety teams, and he is co-signing and endorsing—
Mm-hmm.
—what Coxon is saying, and then embellishing it by putting this greater-than-10% chance number on it, which just makes the thing go super-viral.
Let me ask you a question, Friedberg.
So now it—hold on. It crosses over. If it was just Coxon, you could just say, “Oh, this is just some disgruntled employee.”
If Anthropic was a normal company, what would a normal company do in that situation? They would say, “Look, this guy just joined a short time ago. He was barely here long enough to find the bathroom, okay? He probably joined the company knowing he was gonna resign and make himself into this AI—
They would disavow him.
Yeah, he had an intention to turn himself into an AI safety—
Martyr.
—celebrity. He's certainly availing himself of every opportunity to go on every talk show right now, except ours, to basically promote—
Uh-huh.
—himself. So this guy's a self-promoter. He did not actually reveal anything damaging to our company. He's calling himself, or the media's calling him, a whistleblower, but frankly, all he's got is vibes, and it's hyperbolic, and it's ridiculous, and we're moving forward, okay?”
That's what a normal company would say. But they won't say that because fundamentally they agree with him, and their employees are saying they agree with him, and he's amplifying their message. Frankly, whether they planned this or not, they are happy that he's doing this.
Why? Because they like the political outcome that he is steering the country toward, right? They want an FDA for AI. They actually like what Bernie Sanders is doing. They like the fact that 30 Democratic politicians quote-tweeted this guy. This is part of their regulatory strategy.
So to go back to what Chamath said, there are now legitimate questions here: Did they violate their quiet period? Maybe Dario didn't. Dario hasn't said anything. He's been very quiet, but we know this Hubinger guy—did he violate the quiet period?
Yes.
Is he a sufficiently important executive? And again, is this all part of their strategy?
Well, let me just talk from very basic principles here. Everybody who joins a company signs a nondisclosure agreement. They sign a nondisparagement agreement. They have a corporate communications policy.
The legal department talks to every employee and explicitly tells them, “Do not talk, because we're in the quiet period. We're leading up to an IPO. Everybody keep your mouth shut and do not talk about what's going on here. That is job number one.”
What is going on with the management of Anthropic that they're allowing people to co-sign this, retweet it? They should do what Apple and Steve Jobs did, or Tim Cook did. Nobody in the company speaks for Apple unless it's Tim Cook and unless it comes from the corporate communications department. Who's running comms?
Yeah, but we don't know that they haven't said that. All we know is that there is a pile-on that's happened. The brush fire was initiated by a disgruntled ex-employee who was floating around and was there, as Sacks said, barely long enough to find the bathroom.
But the problem was then a bunch of insiders piled on, and now they have this issue, which is: Is it properly disclosed? And then beyond that, how will people view that disclosure?
There is no way that you can now have a disclosure like this and have it pass muster, meaning you have people at the company saying there's a chance of existential risk and civilizational extinction. Downstream of that is that there is risk to individuals and individual lives.
Downstream of that, there are going to be cases where something bad happens because the AI said to do A or B, or it was trying to give love advice. The person gets rejected, hurts themselves, self-harms, whatever.
There are going to be all of these issues, okay? The long tail of issues is going to be enormous. If social media is a guide, this will be as bad or worse.
Except in all of these cases, nobody at those companies actually thought they were doing something bad. We thought, when we all worked there, that this was legitimately good. There were these spurious things that brought that into question. Some of us talked about that at that point, but this was 10 or 15 years into the journey, not when you were making it.
Now, in the middle of making it, you're like, “Whoa, wait a minute, this is going to kill a bunch of people.” So what do you expect people to do if something bad happens to them or their loved ones? They're going to go back to you. It doesn't matter what the disclosure says; they're going to say, “Excuse me, this.”
What do you think a jury's going to do in a jury trial? What do you think grand juries will do in reaction? What do you think state AGs will do when they're trying to build their political bona fides? I think that this is an enormous problem for Anthropic.
Yeah. Clearly.
Let me peel the onion on this one more layer. You not only have employees of Anthropic possibly violating the quiet period by stepping out and co-signing this, you also have the fact that this whole resignation tweet storm was amplified by these groups that were funded by Jaan Tallinn and Dustin Moskovitz, who co-led the Series A of Anthropic along with Sam Bankman-Fried.
It was an EA-funded round, which is a really interesting detail in Anthropic's history: Dario went to the EA mega-donors for funding at the very beginning, and that's part of why they have this culture. But think about their financial interest. They're on the cusp of making many billions of dollars.
Hundreds of billions.
I mean, the Series A will turn out to be one of the great venture investments of all time. The groups they funded—encode.ai, the AI Policy Network, the AI Futures Project—are funded by Series A investors in Anthropic. So this is getting really twisted, I think.
Let me just say, this makes no sense to me because we were sitting here three weeks ago, and we had Gavin on the podcast. I don't know exactly when their quiet period started, but Dario thanked Gavin for a thoughtful exchange, and he went and explained in detail all of their concerns and countered this claim that Gavin had—that he heard from insiders that people inside Anthropic believe they were the last company standing—which is to say they would take all of capitalism and coalesce it into one company with their superintelligence and be able to beat every company on the planet.
Friedberg, at some point Dario believed he should comment on this. So where are the comments about this?
Well, Anthropic did release a statement, but it's basically saying—
Anthropic did, but Dario did it himself. The CEO did.
The statement they put out is this mealy-mouthed, vague post about how they're a responsible actor or whatever. Again, it's basically more of the same from them, which is that they have this fundamental tension: they're a frontier AI company that believes that frontier AI is inherently dangerous.
Where that obviously leads to is Bernie Sanders, which is: shut it all down, or at least stop it, right? But they don't want to do that for their own motivations. What they claim is, “We're uniquely situated to protect humanity.” It's the savior complex.
But the rest of us aren't buying that, and we're not going to give them a monopoly because we think they're so virtuous. So their way of resolving the tension is not acceptable to the rest of us, and that's the problem they have.
Here's their chief brand and communications officer, Shasha, responding to the clip from All-In: “Dario has never said this. Complete and utter nonsense.” That was on August 14th, after the show.
Obviously, Dario and she have not said anything right now, so maybe they weren't in the quiet period on October 15th, and they are now. All right. Listen, in other news in AI—great discussion, everybody—OpenAI says it solved a 200-year-old math problem. But claims now are that somebody may have front-run an actual scientist's work that was being done on OpenAI.
The Navier–Stokes equations, written in the 1800s, describe how fluids move. Every aircraft, pipe, and weather model of Earth is built on them, and they apparently—give us the history of this here, Friedberg. A lot of people have been working—
Look, I'm not the guy. These are methods for modeling fluid dynamics—partial differential equations. These are mathematical approximations of continuous things in the physical world. In the solution that OpenAI published, they said that they used a reported 130 billion output tokens to do the work between 10,000 agents that were spun up to work with one another.
I just did the math on this. The human-equivalent hours are the real estimate of what this work was that was done by these agents. I think this is so important for people to understand: it's not like the AI had some stroke of genius, some magical insight that no human brain could comprehend.
What happened was the AI just did a bunch of brute-force work to come up with this answer, and the work was done across many, many computers. 10,000 agents just means 10,000 applications running on a computer, talking back and forth to one another, sharing information, sharing analysis.
The equivalent, if you were to think about human work years, is somewhere between 50 and 500,000 years of human work, with humans typing at 40 words a minute, having mathematical problems, solving them with calculators, doing this sort of stuff back and forth. You could probably reduce that down by an order of magnitude, but it's still in the order of tens of thousands of years of human knowledge labor.
I think that what this indicates about AI is that AI is a tool of leverage for humans. AI is not some magical genius super-god that sits above us, that understands things we don't understand. Every piece of the OpenAI message set that went back and forth between these agents is documented, can be read by people, and can be understood by people. We can see the work that the agents did over the equivalent of tens of thousands of human years to get to the solution.
For me, the solution that was described here is more of an insight into what AI is rather than what AI isn't. AI isn't some mathematical god; AI is an engine that gives humans extraordinary leverage.
When you think about the application of this particular set of problems and what else we might be able to do with things that are similar, instead of spending years trying to design a new aircraft wing, a new aircraft wing can now be brute-force designed on a computer in a matter of seconds or minutes. The same goes for a new engine design or a new energy system design.
AI is a tool of leverage that takes tens of thousands of years of work and reduces it down to minutes or seconds. Just like any other technology before it, I think it indicates that this gives humans leverage. It seems magical at first, it seems unbelievable at first, but when you dig into what actually took place, what's really going on is intelligent systems designed to solve problems using human-known techniques.
So I want to make my summary point on this whole thing. You guys are welcome to talk about partial differential equations and why they're so hard, but for me, that was my biggest takeaway on this whole thing.
Yeah, the bigger issue here, Chamath—
Your takeaway is that this was a very clever systems approach to brute-force problem-solving.
Yeah. The other issue here—
I agree with that. Also, Jason, the autocomplete still doesn't work.
Yes. We solved one of the 7 hardest math problems in the world, and you still can't get your intimacy kit from the front desk.
And depending on the model, it still can't spell “strawberry” right.
Or count to 100.
How many Rs in “strawberry”?
Oh, my God. Chamath, the other issue here is one we've talked about in terms of AI sovereignty: using open source in the enterprise and not trusting frontier models with your data.
When OpenAI came out with its solution, there was a claim that the original scientist working on this may have trained OpenAI's LLM to make it easier for them to solve it. OpenAI posted the following statement: “While unlikely, we cannot rule out that de-identified data derived from their usage”—they being the mathematicians—“of our products helped improve our models.”
So maybe you could speak on AI sovereignty here. Should any scientist, law firm, or person working in biotech making new potato seeds trust any of these LLMs, for fear of having their proprietary knowledge cribbed into a core LLM?
If you have incredibly sensitive proprietary data, you have to make a very difficult decision because the reality is that something is leaking. There's remnant memory of how these solutions were solved that sits within these models even after the fact.
There's a concept inside these models called zero data retention, or ZDR.
It’s a best-efforts basis. It’s a commercially best-efforts basis at that. They can’t guarantee it. A simple example: you could say, “Do not retain the decision-making process of this protein design.” And it can do that, but if anybody using it clicks the Like button inside the chat window, that’s not necessarily guaranteed.
So there are all these vectors where this data leaks into the broad corpus and understanding that these models have. If you believe that the information you have is critically important, you cannot use these services the way they’re currently offered by most people. What you need to do is stand up your own sovereign solution.
What does that mean? It means you need to go to a vendor that you trust. It could be AWS, in terms of the hyperscalers, or it could be Nebius. What they’ll do is stand up your own hardware, and then you’ll stand up your own models. They’ll provision those models to you, and that’s how you should probably be consuming it.
Hmm.
And if you don’t do that, you are creating risk. Look, let’s be clear: Will a handful of CIOs get very publicly flogged and fired in the next year because they accidentally didn’t understand this, did an API deal because they wanted to feel popular, and leaked data into these models? Guaranteed.
Why will people get fired, Chamath? What’s going to happen?
The thing that’s happening right now is that a lot of this AI understanding and awareness has bubbled up through the audit committees and the risk committees of public boards. Who are those? They’re a handful of directors who are responsible for understanding the IT posture and the security posture of their business. Then they disclose that back up to the board, and that understanding and awareness is what gets signed off on and submitted back to the SEC in these filings.
They also have consultants who help them, folks like Ernst & Young, Deloitte, and others. Increasingly, what we’re all learning is that ZDR is not an effective solution, and there’s a ton of leakage. What’s happening now, Jason, is that this information and awareness is bubbling up through the risk and audit committees of these boards. It’s now going to the broader boards, and CEOs now understand, “Hold on a second. My information could leak.”
So now they turn to their CIO and say, “Okay, whatever you thought was happening, we now clearly know that it’s not. What is your answer?” It’s when that happens, or when all of a sudden they have to disclose that some critical IP was leaked into a model—or that they don’t know whether it was or was not, and all of a sudden something similar appears on the other side—that shareholders will be really upset. They’ll say, “Guys, when OpenAI even tells you that they can’t guarantee it, and you have closed your eyes and assumed that your information isn’t leaking, what are we to do?”
If that decays and erodes my share value that I purchased from you, assuming you were going to have your hand on the switch, that’s where folks get fired and—
Yeah.
Shareholder lawsuits, and it’s going to be a mess.
What does it mean for frontier models in terms of this trust leakage? Another reason to, I guess, embrace open source in the enterprise and for individuals who are doing important work.
By the way, it doesn’t have to be open source. It just has to be—
On-prem.
Well, it doesn’t even have to be on-prem. It’s effectively like your own VPC. You must go straight to the bare metal on your own terms. You need to control everything. You can work with folks like Amazon, Nebius, or Fireworks. You can have open source solutions. You could even probably have Claude allow you to host a version of Claude in your VPC.
But if you don’t go through these hoops—
Mm.
—because you took a standard-rate-card API deal, either because you were lazy or you didn’t know any better, and after today you’re still doing this stuff, you’ll probably get fired.
I had mentioned this over the summer. A lot of these AI application companies are now moving off of the frontier models, and I think that’s a headwind. We’ll soon be able to see some data on these very large customers they have who are spending $10 million or $100 million per month on these services.
This is why we at 8090 partnered with EY and Deloitte. We’re like, “Hey, let’s find a way to help solve these problems in advance.” Initially, when we would go and talk to a bunch of companies, they would say, “Oh, we did the ZDR thing.” We’re like, “Okay. We’ll just keep educating you. In 5 or 6 months, you’ll realize that ZDR means literally nothing.”
Yeah.
It’s flimsy, and it’s Swiss cheese. Now all these companies are coming back to EY and Deloitte and us, and they’re like, “Oh my God, fix the problem.”
Yeah.
Part of it is this realization that folks like OpenAI confirm in their own press releases.
Yeah. Harvey announced on September 9 that they had their own proprietary legal AI model. It’s called Tenet, T-E-N-E-T, and they based it on Kimi K2.
I mentioned earlier that we incubated a company called go.ai, and this company is on fire. What they do is build on-premises hardware, a thing called the Go1 box, which I’m showing here on the screen, for people who want to get control of this. This is one of our incubated companies from our launch accelerator. I’ll give them a quick shout-out.
Any thoughts on this and the data leakage issue, Sacks, and what it means for the frontier models and how their biggest clients either trust or don’t trust them with their data and their innovations?
I tend to believe Noam Brown, who’s a senior researcher at OpenAI, and what he said about this is that nobody looked at Levent and Tristan’s prompts. That would be insane. So he’s denying it. I think it just doesn’t make sense to me that they rummaged through their prompt chains.
Could there be data leakage through the deidentified data? Perhaps. But look, I think the more plausible explanation here is that OpenAI heard that researchers were making progress on Navier–Stokes and just threw a ton of compute at it to see if they could—
And Sam actually confirmed—
They can do it.
—that part as well, that they heard—
Yeah.
—that Anthropic was getting close, et cetera. So they were like, “Yeah, let’s take a swing.” Yeah.
So I think that’s probably what happened. I don’t know anything; I could be proven wrong, but I tend to believe that.
Look, I think this raises a somewhat orthogonal issue that’s worth talking about, which is data privacy. I think we need much stronger data privacy laws around the data that you have with AI services.
Mm.
Right now, the data you have in your AI chats doesn’t even reach the same level of protection as email. In most contexts, if the government wants to get your emails, they would have to get a search warrant and prove probable cause in a court. But that is not the standard for AI data. The standard for AI data is that you can just get a subpoena or a court order.
When you think about how personal AI is becoming, people are using AI as their lawyer, their doctor, their therapist—the list goes on and on. The idea that somebody could just get your data without probable cause, without a search warrant, seems crazy to me. So I think we at least need to strengthen AI data to the level of email, and maybe even beyond that, to reflect how personal it is to people.
Again, it’s nuts that if you talk to a lawyer, that’s a privileged communication, but if you ask that same question to your AI and then talk to your lawyer about it, it might not be protected. We need the data privacy rules to reflect how people actually use these tools and how they think about them.
Any final notes on this, Friedberg, in terms of the data leakage issue?
I’m deeply concerned about it. I’ve had experiences where we’ve asked some fairly novel scientific questions, and it identifies them as a novel insight. It’s like, “Oh, never thought about that. Interesting.”
Then, using a different account and asking the same model later, or asking the next version later, I’ve now experienced this: It’s like, “Oh, well, you could do this,” and it actually describes this exact thing that we had in our chat in the previous version.
Now, these are a handful of anecdotal experiences, but I know the domain that we work in, the niche of it, the ideation and novelty of this stuff, the lack of papers being published, and so on. I know there isn’t some new corpus of information out there that’s training the new model. So all I can say at that point is that my conversation or our analyses have been used for training. Go ahead.
Wow.
Siri, Siri, please prepare the shareholder lawsuit to Oholo.
Okay, this does raise a really good question: What does it mean that the model is allowed to train on deidentified data?
That’s my point. It doesn’t use any of my personal information, but it can use an insight derived from our chat, which it can then say is some training data that is unrelated. But the truth is, it’s actually a piece of IP that’s—
Trust no one.
Exactly.
But the truth is, it’s actually a piece of IP that’s—
For our organization's IP and our engagement back and forth, we don't have any NDA or confidentiality provisions or protections with them being a service provider back to us.
Mm.
I think that's what's lacking in the terms of service in all of these hosted services. This is why I care a lot about open source, because I don't want them having my chat logs, because they can use them for training to create an IP advantage—
Bingo.
—that is now diffused to the rest of the market.
I find this very unfair. Full disclosure, yes, I'm an Oholo shareholder, but take any company. You have these brilliant scientists at Oholo. They're pushing the boundaries of science. They are in a position to create true abundance. The first real manifestation of abundance would be food, right?
And now Friedberg has to sweat whether his ability to actually build a viable business has been compromised—not because they did anything wrong, but because they thought they were just trying to use a tool to advance his ability to get to the answer faster, and he thought he was doing it in a reasonable way. Now, all of a sudden, it could just leak out, and I just think that that's not fair.
I'm going to get you a Go AI machine. I literally just ordered 2 Mac Studio M5s. I just ordered 2 of these. We're going to be running local models, and we're moving off Claude for a lot of our sensitive data.
The problem with local models, just to be clear, whether they're closed or open, is that you need a multiplayer experience, and it needs to be cloud-based. You need to have a knowledge base. You need to have memory. So I don't think buying a machine solves the problem.
The problem is that the way these models are constructed, they are these layers of embeddings, right? When you go from an input token to an output token, the stuff that happens in the middle is a total leaky black box. That's the problem, and they give you a fig leaf to make you feel like you're not leaking your IP to folks. But when push comes to shove and you ask them, "Have you learned, and can you guarantee you have not learned?" the answer is, "No, not really."
The thing we're doing is we're putting open-source models on this, and then we connect API to API to these as if they're servers. We've already experimented with it. I think we're going to take, like, 90% of the work off of—
No, but I think that's fine.
—our Claude relationship.
But how does a 1,000-person or a multi-hundred-person company do that?
It's going to be more. They're going to need to have rack servers. They're going to stack the rack servers themselves.
Wait, let me ask you a question.
Yeah.
Let's say that the folks at OpenAI are looking for mathematical breakthroughs, and they would just ask their model, "What do you think are the top problems that we could solve in the next few weeks?" Would the model be able to use deidentified data from all of its users, seeing all the math usage and seeing what is close to being solved—
Exactly.
—and then roll that up?
Exactly. So take away—just focus on the term deidentification. Deidentification means removing personally identifying information or information that's specific to a particular company. But a general approach to a mathematical problem—the approach is the IP.
So if someone is iterating on that approach inside of that tool, that iterative chat can be deidentified as a first step. Okay, great. There's nothing that identifies who did this. There's nothing that identifies something unique about a company or a person. Let me learn from this. Now the model is smarter at how to think about solving mathematical problems because it just observed this user using the model.
This is the network effect of these closed AI model systems—
Yeah.
—they get to see what everyone in the world is doing. They're the master key into everything.
That's right.
That observational system gives them more data than anyone else has, and that is what builds their advantage over time so that their models are better than anyone else's models. But as a user, as an enterprise or consumer, I am personally concerned about my IP or my personal information.
Yeah.
Well, now we're in the Alex Karp world of you're giving them all your alpha.
Yeah. That's right.
And there's only 1 way to really solve this, aside from you just moving on to your own open models and your own infrastructure: those companies have to forgo the opportunity to go into vertical applications built on top of their platform. In other words, they can't compete with their customers.
But they've already indicated a desire to compete with their customers, and they can't even tell you. They've already done it, and they can't even—
They've even done it. They released Claude Code, they released Claude Design, and that really pissed off Cursor, which was one of the biggest customers that Anthropic had to date. They felt they got rug-pulled by Claude.
Right. And this is where the guys who are complaining, Levent and Tristan, this is where you have to feel some sympathy for them: they don't know. They're making this accusation based on timing that actually their work got stolen.
But the fact is that they are using OpenAI's models, and OpenAI is competing with them to develop this breakthrough, so obviously you're going to make that accusation whether it's true or not. I'm giving OpenAI the benefit of the doubt. I'm saying it's probably unfounded.
But as long as we have a closed-model duopoly of frontier AI, and they're reserving the right to get into every vertical application there is, they're declaring in advance that they're going to compete with their customers. So how can we trust them with our data? Unless there are much stronger privacy laws as a starting point, and even that might not be good enough.
All right, gentlemen—
Maybe that's the law we should pass first before we regulate.
The market is responding to it. I think that's the high-order bit here: people are now—
But hey, can we—
—taking measures. Yeah.
Hey, 1 piece of breaking news here. I don't know if you guys saw this, but Jensen is speaking at a conference right now, at the Goldman Sachs conference, and he just said that the Jacob Coxon comments are outlandish and deeply untrue. He said the labs are great, but his comments were wrong, arrogant, and ignorant of all the work being done around the industry to drive safety.
Now, here's my question: If Jensen can say this, why isn't Anthropic?
Oh.
Quiet period, I guess. Ostensibly, they can't come out and say it.
Yeah.
No, they released—
Hmm. What did you call it?
—a different kind of statement that was, well, sort of mealy-mouthed.
Mealy-mouthed? What does it mean to be mealy-mouthed? What is mealy-mouthed?
It's like you smoosh out of both sides of your mouth. It's like—
Oh, I see. I don't know what mealy is.
Vague-posting, virtue signaling.
All right. Nike, the 60-year-old iconic American brand—"Just Do It"—just got booted out of the S&P 100. They had been in the stock index, Sacks, for 18 straight years, being replaced by—hey, shout-out to our friend Nikesh—Palo Alto Networks.
Nikesh.
Congratulations to Phil Knight and Nikesh. A little history here: they went public in 1980 with a 50% share of the US athletic shoe market. They were the dominant player, the Coca-Cola of sneakers, and they changed things forever with Air Jordans in 1985. They did the iconic "Just Do It" campaign in 1988.
Peak market cap in 2021 was $264 billion. Peak revenue was $51 billion in 2024. Since then, revenue's dropped a bit—10%. In 2020, John Donahoe became CEO and pushed an aggressive direct-to-consumer strategy, basically alienating all the retail partners who had helped build Nike and willingly removing those sneakers from the stores.
That made brands like HOKA and On Running—the brands that every VC is obligated to wear, the white ones, to any speaking gigs they have—even more popular. China sales are down 30%, with 8 straight quarters of decline, losing share specifically to Chinese brands Anta and Li-Ning. If you're Chinese, you probably know those.
Nike's marketing also went super woke, and they started backing political movements and putting robust people, Sacks, on billboards. I didn't get asked. This kind of killed the entire brand. The stock is down 80% from its peak. Sacks, $200 billion eviscerated. What's your take here? I don't know if the audience can guess.
I mean, this is the millionth example of "go woke, go broke." Nike was a brand that stood for great athletes, for amazing performance, for victory. Isn't Nike the goddess of victory?
Yeah.
They built their brand on GOATs like Michael Jordan and great athletes like that in many different sectors—men and women of all different races. There wasn't a need to do all this woke stuff, where now all of a sudden they're showing people you don't know.
They generally had low percentage body fat prior to the woke era.
No, but it’s not even about that. This person isn’t even an athlete. Who is this person, and what is the message? “Own the floor”? It used to be “Just Do It.” Then they got Dylan Mulvaney in an ad—the whole trans thing.
They did?
Yeah, this is like the Bud Light—
Oh.
—case.
So you had the woke stuff. Then you have the other piece of it, which is this consulting approach, where a new CEO comes in. We know John Donahoe—nice guy, central casting. He looks like a CEO. But it seems like a lot of change for change’s sake, where he comes in and direct-to-consumer is the new hotness, so they blow up all their retail operations and all their retail relationships. They had all these stores that had shelf space for Nike, and now that goes to competitors, who then took a huge amount of market share. The other thing I think they did is a reorg, where they used to have departments in the company based on sports. There was a basketball division, football, and tennis—
Swimming.
—and so on.
Everything.
Swimming. And they reorged that to be, I think, men’s, women’s, and kids’. Why would you do that? I just—
Why break—
—I don’t get it. It’s—
—what’s working? You’re surging—
Yeah, it’s like—
They had a very clear model here, Chamath: clicks and bricks, right? You could go to a brick-and-mortar store, have a great experience trying these shoes on, running around the store, and working with people who help runners pick the right shoes. It’s nice to be able to buy direct. I have to say, the Nike app is a delightful app, but why would they kill the retail channel of all things and piss off those folks who were their big advocates? If I’ve learned anything in almost 30 years of business, you have to have a very clear North Star and stick to it.
Hmm.
At least when I was growing up, Nike’s North Star was very clear to me: mastery and excellence embodied through athletics. That was it. I would see Michael Jordan, and I saw mastery and excellence. I would see Tiger Woods—mastery and excellence. I would see Serena Williams—mastery and excellence. Pete Sampras. You name it, and it was mastery and excellence. Somewhere along the way, mastery and excellence became too traditional-looking, and so they went away from things that had nothing to do with mastery or excellence. I don’t aspire to be a fat person. Never have, never will.
Yeah.
I aspire to be Michael Jordan. Always have, always will. I just think it’s as basic as that. You’re not going to buy the clothes of a brand that you don’t aspire to be like by wearing those clothes or those shoes.
I think as long as you can recenter around this idea that we’re not here to make people feel better about themselves. Our success was when people imbued their desires to get better through the lens of these incredible people. That’s aspirational. Being inspired by somebody and wanting to be somebody better than yourself is a good thing. We should not shame that. It is good. It’s how we all get better.
There are incredible athletes who give up their entire lives to master something, and I think Nike should own that and be proud of that. If they do that, all the other details will make sense. How do you sell? Sell it everywhere, because everybody will want it. What do you sell? To everybody, because everybody will want it. But if you don’t get the North Star right, you’re going to miss out, and I think the North Star was clear: excellence and mastery.
There’s a broader discussion here, I think, that came up in your incredible Eric Weinstein All-In interview, Friedberg, where excellence, expertise, and the desire to be elite—the desire, to Chamath’s point, to be inspiring to people as somebody to follow—we’ve lost this love of excellence and expertise, and it’s been replaced with participation trophies or inclusiveness as opposed to excellence. Maybe you could talk a little bit about that in your field, your chosen field, science and entrepreneurship, et cetera.
Entrepreneurship is a meritocracy, so you either do it or you don’t. But in the case of this business, I think what they also lost—and it didn’t just come through in the brand—I used to buy Nike.
Mm.
All the shoes I owned were Nike. It’s all I bought.
Me too.
I had my brands. I liked to go to them. What happened over time?
100%.
It wasn’t that the ads went bad. It wasn’t that the ordering was bad. The product started to suck. The shoes literally fell apart in six weeks. They were flimsy. They didn’t have the same quality they used to have. I could tell that the product started to suck. So one day I went down to my REI store next to the place where I get salads. I stopped in, tried on a pair of Brooks, bought the Brooks, and started wearing them.
Salads.
All I wear now are these Brooks shoes. And Brooks—
The Speedwalking 900s?
—they’re a great product.
Brooks.
They last forever.
I know Brooks.
They’re durable.
Horrible. That’s unacceptable.
They’re comfortable.
Marathon runners do like Brooks, yes, and speed walkers.
I moved to On. Wait, you know why I moved to On? I got them for running because when I started running in Nikes, I got these knee problems. I’m getting old. The Brooks are awesome. They just feel great. So—
Is that what they call skipping now—
Yeah.
—running?
Yeah. I moved to On for one reason: I held on to Nike forever.
Mm. Right.
I was a Nike diehard.
I was the same. I’d buy a new pair every six weeks.
Diehard. Yeah.
Of course.
Every six weeks I was buying.
Yeah.
Well, no, every six months I would buy new Nikes, and I still have a collection of many Nikes. I love Nike.
Well, yeah, it became six weeks as the product sucked.
And I finally moved to On for one reason, because I thought, “Roger Federer—he’s excellence and mastery. I like him.”
Right.
Everything he does looks effortless.
Right.
I love that, and so I thought, “Roger Federer—good for me.”
Let me just tell you the story on Brooks. This company is owned by Berkshire Hathaway, and they’ve grown this company for 9 years in a row with double-digit revenue growth, to $1.6 billion in revenue. It’s a standalone company inside Berkshire, and it just keeps compounding revenue as a subsidiary. There’s an interview with the CEO where he said, “Warren Buffett just told me, every year, make the product better than it was at the end of last year. If you keep doing that, the business will keep growing.” That’s what he’s been doing, and I think that’s the opposite of what Nike’s focused on, because they shifted from product to narrative. It was all about, “What’s the narrative that we think the audience wants to hear?” It’s the same problem Bud Light had.
It was pandering. They pandered.
And the whole thing—yeah, and when you kind of—
Remember that huge ad for Colin Kaepernick they ran in San Francisco? It was like—
Oh, that was the turning point, in my opinion.
What the hell does that have to do with anything? I want excellence and mastery.
Yeah. You don’t have to comment on everything.
Yeah.
I think that if you were to pinpoint the beginning of the downfall, that was it. Look, there have been moments in sports that were inherently political, and they were great moments. You could think of Jesse Owens at the 1936 Olympics, raising a fist.
Yes.
Yeah.
You know.
Jackie Robinson.
Muhammad Ali.
Jackie—
Jackie Robinson or Muhammad Ali refusing the draft into the Vietnam War, but they were all the greatest—
Exactly.
—and their causes were things like standing up to Nazis, protesting the Vietnam War, and protesting racism. What was Colin Kaepernick taking the knee for—to protest the American anthem? What was the point of that? Was he the greatest, or one of the greats? No, he was not. So why would you give that a platform? What were you promoting exactly? What were you trying to say? Then they kept doubling down with Dylan Mulvaney and the rest of this woke stuff. It is so breathtakingly stupid. Their job was marketing, right? They were actually marketing images that were so contrary to what their brand was about. It was so craven. It was so faddish that I think every single person who was part of that decision-making chain, from the marketing executive all the way up to the CEO, should be drummed out of corporate America. How could you ever approve those campaigns? It was—
It’s unbelievable.
—it was during a certain time period, obviously.
If they find mastery and excellence as their North Star again, I think they’ll be fine, because I think it’s a coiled spring in terms of its potential.
Well, that's what I was gonna get at.
Uh, now by the way, just take an example.
What do we do from here, Matt?
But, like, at Stanford, the Nike store closed. You know what replaced it? On. And you know when you go to the On store, it's amazing. And you're like, “Oh, Roger Federer is the...” This is what I think when I go in there. And now I may be limited in my ability to understand why I need to care about all these other things when I buy a pair of running shoes, but normally I was just thinking about mastery and excellence. Michael Jordan shoes, my shoes. Tiger's clubs, my clubs. You know, Sampras' clothes, my clothes. Uh, Federer's racket, my racket. That's how simple I am, and I would just happily buy that stuff. So you just, A, you gotta reestablish yourself in retail locales where there are people with disposable income, and B, decide whether you wanna be aligned with mastery and excellence, because then I think tens of millions of men and women will show up again because that's what they want. Maybe they wanna get back in shape. Maybe they want to reestablish themselves in a sport. Maybe they wanna learn a new sport. And I would just embrace Nike everything, and now I don't.
Yeah, they should have really gotten into the smart devices as well. They dipped their toes into that. I thought that was, like, really good 'cause it was so performance-based. If they owned Strava, is that the app that all the elite—
Strava.
—Strava, all the elite biking and running people do, and then, you know, embrace that, and they should just redo the tagline, “Do it or don't.” Like, come up with, like, a really challenging rebrand here.
Do it or don'ts.
I think that's Yoda, J-Cal.
Yeah.
Yeah.
It is Yoda.
I, I would love—
That is Yoda.
Do or do not. There is no try.
Do it or do not.
If you became the CMO, if you became the CMO of Nike—
J-Cal, the CMO of Nike.
—I would launch an activist campaign. I would literally—
Do it or don't.
—buy billions of dollars of stock and then have you fired.
Dude, I lost 43 pounds. You gotta s— I'm doing farmer's walk half mile, 20 pounds on each arm.
Bro—
Do it. Just do it.
—if you just don't buy into this woke stuff, you're, like, ahead of the game right there.
Yeah.
Just do it.
Refocus—
Just do it, man.
—it’s so incompetent. I cannot get past the incompetence to woke your way out of the S&P 100.
All right, everybody. That's your All In podcast for September 11th.
Thank you, Jason, for your mastery and excellence yet again.
Yeah, just do it or don't. Do it, man.
I mean, just do it.
I'm, I'm, I'm getting 30, 40—
Do it or don't.
—do it or don'ts.
Do it or don't, give a shit what you do.
Do it or don't.
Jason, marketing genius, Don Draper over here.
Everybody make some ads with me as the CEO, uh, and do it or don't. We don't give a shit what you do.
This is like, uh, Brooklyn Don Draper.
Yeah, we go from just do it to do it or don't, we don't give a shit.
No, I think it would be, like, a very provocative thing to do.
This is the only way to make things worse at Nike.
It's the worst tagline.
You guys are wrong. Test it. Test it.
You found the one way.
Do it or don't.
He found a way to get kicked out of not just the S&P 100, but also the 500.
Get delisted, pink sheets. No, I think it will work.
Do it or don't.
I, I'm standing behind it.
That's what you should tell the S&P committee.
Do it or don't.
Do it or don't. Do it or don't.
That's the—
Kick me out or don't. Do it or don't.
I mean, I'm good at marketing. Who named this podcast, you ungrateful shit?
You're terrible at—
Like, I named it All In.
I did.
I named it.
I did.
You said, “What do we call it?” I said, “All In,” then I got the domain name, you ungrateful shit. All right, everybody.
See you. Good luck.
Do it, Jamal.
You just go, daddy? No, I negotiated with the owner. I got it for 250 beans. I got a great deal on that five-letter domain.
Okay, well, do it or don't, you did it, so whatever.
Yeah, I did it.
You did it.
I did it.
Yeah, you did it.
Say thank you, bitches. All right, everybody, we'll see you at All In Summit. Another thing—
Love you, boys.
—the All In Summit is this weekend, Sunday through Tuesday, in LA, presented by Irene. Irene is the AI cloud, data centers, compute, and software for training and inference. They're throwing Sunday's registration day party, so grab your badge, grab a drink, and join us at the Irene house in the expo hall for a meet and greet and photo op. We're also excited that Meta will be in the hizzy. President and Vice Chair Dina Powell McCormick will join us on stage. Plus, there's a Meta glasses experience hub. Check it out at the All In Summit.
Let your winners ride.
Rain Man, David Sacks.
And instead, we open source it to the fans, and they've just gone crazy with it.
Love you, bestie.
Ice queen of Kin Wah.
Let your winners ride. Let, let your winners ride. Let your winners ride.
Besties are back.
That is my, uh, dog taking a piss in your driveway.
Sacks.
Wait, no, no, yeah.
Oh, man.
My amateur will meet me at Blitzen. We should all just get a room and just have, like, one big huge orgy, 'cause they're all just useless. It's like this, like, sexual tension that we just need to release somehow. Wet your the beef. Wet your beef.
Wet your beef. Wet—
We need to get merch.
Besties are back.
I'm going all in. I'm going all in.