Zuckerberg rejects the AI industry pause

0:00Chronic migraine, 15 or more headache days a month, each lasting 4 hours or more, can make me feel like a spectator in my own life. Botox, on a botulinum toxin A, prevents headaches and adults with chronic migraine. It's not for those with 14 or fewer headache days a month. It's the number one prescribed branded chronic migraine preventive treatment. Prescription Botox is injected by your doctor. Effects of Botox may spread hours to weeks after injection causing serious symptoms. Alerture doctor right away is difficulty swallowing, speaking, breathing, eye problems or muscle weakness can be signs of a life-threatening condition. Patients with these conditions before injection are at highest risk. Side effects may include allergic reactions, neck and injection site pain, fatigue and headache. Allergic reactions can include rash, welts, asthma symptoms and dizziness. Don't receive Botox if there's a skin infection. Tell your doctor your medical history, muscle or nerve conditions, including ALS Lugeric's disease, Myasthenia Gravis or Lambert Eaton Syndrome, and medications, including botulinum toxins, as these may increase the risk of serious side effects. Why wait? Ask your doctor, visit BotoxCronicMigraine.com or call 1-800-44 Botox to learn more.

1:25College football is back. So, Hilton called to me the superstition concierge to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky jersey? Oho, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching bandwinkup call at 555 and 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton, for this day. Metacetio, Mark Zuckerberg, has publicly rejected calls from Anthropic Spario Modei to coordinate an industry-wide slowdown on advanced AI development. Right, so the leading developers of the most powerful technology in the world are currently fracturing over whether to pace themselves together or just race each other alone. Yeah, Zuckerberg is basically arguing that companies have a natural incentive to build safe products because users will refuse to adopt unsafe AI agents.

2:25I mean, he views safety and alignment as a competitive advantage, rather than a shared industry standard. Which leads you with a fairly big question to keep a mind throughout all of this. Does a free market race actually produce safer technology? Or are we just trusting the people building the rockets to inspect their own fuel lines? Well, he points to Meta unilaterally pausing the release of its latest muse tools. You know, they delayed the launch to focus on security and safety testing. He uses this as proof that self-regulation works fine without any industry mandates. Yeah, but taking warning seriously is very different from taking CEOs at their word. Sinosa pointed this out recently. I mean, delaying a product launch for standard testing is just routine business. Rebranding it as a moral victory is just good PR. It is a deliberate framing though. Zuckerberg is really staking out a clear position against coordinated pauses. His core argument is that labs should not pause together. Each company should slow down only when its own safety work demands it. The logic there relies entirely on safety functioning as a competitive differentiator.

3:29The theory is that if users will not trust unsafe agents, the labs are incentivized to get it right without needing external mandates. It places all the faith in market pressure over industry coordination. And this is a meaningful split from the camp pushing for shared standards and collective slowdowns. Meta is betting that self-regulation produces better outcomes than coordinated constraints. It is a defensible position, but it also conveniently aligns with their desire to move fast. Right, but who actually wins when safety becomes a competitive race instead of a shared standard? If we rely on the market to punish unsafe products, the punishment only happens after the unsafe product is released and adopted. Right, market correction is reactionary by nature. Exactly. We are relying on the damage to occur so the market can react and choose a different product. When you are dealing with agents that can execute tasks independently, waiting for the market to realize a product is unsafe means you have already accepted a certain level of collateral damage. If you're not subscribed yet, take a second and hit follow on whatever podcast app you're using.

4:31It helps us keep making this. We appreciate you being here. You know, it's interesting meta claiming a unilateral moral high ground here is, well, especially when you look at how quickly their competitors are changing their tunes on the exact same issue. Oh, absolutely. The positions are shifting so fast, depending on who is talking and who they happen to be talking to you on any given day. Yeah, look at OpenAI as Sam Oldman. Right. Initially, Altman actually agreed with Emma Day about pacing the frontier and using independent evaluators. But then, just days later at a major conference, he walked that right back. He stated the world should trust that they are going to do the right thing while simultaneously saying the public is right to be afraid. Be afraid, but trust us. Yeah. I mean, that is a contradiction that reveals a structural problem with this entire debate. The core verification flaw here is glaring. If a closed lab claims they slow down their development, there is literally no way for the public or their competitors to audit that claim.

5:32Because you cannot verify a negative. You can't look at a closed server firm and confirm they are not training a model as fast as they possibly could be. And because there is no way to verify it, the incentive to secretly rush remains completely intact. You tell the public and your competitors that you were pacing yourself, but behind closed doors, you run the compute as hot as possible. You have no idea how much electricity your competitor is pulling. Right. You have no idea what architecture they are testing on those closed servers. Well, and Nvidia CEO Jensen Quang was at the exact same conference telling developers to run as fast as they can, calling regulation completely unnecessary. I mean, the person selling the chips to fuel the race is naturally opposed to speed limits. If the entire industry agrees to a six month pause, that a six months were the demand for his hardware, potentially softens. His financial incentive is maximum acceleration at all times. There is also chronic migraine. 15 or more headache days a month. Each lasting four hours or more can make me feel like a spectator in my own life. Botox, on a botulinum toxin A, prevents headaches and adults with chronic migraine.

6:36It's not for those with 14 or fewer headache days a month. It's the number one prescribed branded chronic migraine preventive treatment. Prescription Botox is injected by your doctor. Effects of Botox may spread hours to weeks after injection causing serious symptoms. A lurcher doctor right away is difficulty swallowing, speaking, breathing, eye problems, or muscle weakness can be signs of a life-threatening condition. Patients with ease conditions before injection or at highest risk. Side effects may include allergic reactions, neck and injection site pain, fatigue and headache. A lurcher reactions can include rash, welts, asthma symptoms, and dizziness. Don't receive Botox if there's a skin infection. Tell your doctor your medical history, muscle or nerve conditions, including ALS Lugeric's disease, Myasthenia Gravis, or Lambert Eaton Syndrome, and medications, including botulinum toxins, as these may increase the risk of serious side effects. Why wait? Ask your doctor, visit BotoxCronicMigraine.com or call 1-800-44 Botox to learn more. This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome? That's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block,

7:37or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up require compatibility and availability very 16-plus. College football is back. So, Hilton called to me the superstition concierge to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky jersey? Oh, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching bandwicker call at 555 and 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton, for this day. So, a distinct political angle to this whole thing. President Donald Trump publicly criticized Altman's initial endorsement of pacing. The concept of pausing has become highly politicized. Whoever wins the AI race wins geopolitically. And that political pressure adds another layer of incentive to ignore any coordinated slowdown.

8:39If slowing down is viewed as seeding technological dominance to geopolitical rivals, the pressure from political figures will always push these labs to accelerate. You're not just building software anymore. No, you're being told you're building national security infrastructure. And the moment it becomes a matter of national security, the idea of a voluntary cooperative slowdown goes straight out the window. So, the impossibility of verifying a closed labs slowdown naturally leads into the reality of developers who operate entirely in the open. Because you cannot talk about pacing without talking about open source. The open source community completely changes the math on this entire debate. It is the variable that makes coordinated pauses nearly impossible to actually execute. We have perspectives from hugging faces Julian Shaman and David Sacks on this. They are clear that open-way developers are never going to pace themselves. They can't pace themselves. The nature of open weights is continuous, decentralized iteration. Once a model is out there, thousands of developers are modifying it,

9:41stripping away restrictions and pushing its capabilities further. For anyone listening who might not be deep in the technical weeds, we should probably explain what open weights actually are. Yeah, that's a good point. So, traditional software is just a list of instructions. If I give you the code for a calculator app, you can see exactly how it adds two numbers together. But AI models do not work like that. They are trained on massive amounts of data to recognize patterns. And the result of that training is billions of numerical values called parameters or weights. These weights are essentially the model's memory and decision making logic. So, when a company releases a model with open weights, they are taking that fully-trained brain, which cost millions of dollars in computing power to create, and putting it on the internet for anyone to download. You do not need a supercomputer to run it once it is trained. You can run it on a high-end laptop or a local server. And more importantly, you can modify it. If the original developer puts safety guardrails in place, say, instructing the model to refuse to write malicious code, an independent developer can take those open weights and fine-tune them to remove those restrictions.

10:46Yeah, and it takes a fraction of the time and money to remove a safety filter than it took to build the model in the first place. Which is why Shaman and Saks are pointing out the reality of the situation. You cannot ask a decentralized network of thousands of independent developers across the globe to simultaneously pause their work. There is no central authority to enforce it. At the same time, Chris LeFane from OpenAI confirmed that major labs, including OpenAI and Thropic and Google DeepMind, have been privately talking for an extended period about forming an industry standards body. A secret committee run by the labs themselves. Right. They were discussing an industry standards body privately before a Moda even published his essay calling for a slowdown. But the futility of an industry-wide pause becomes obvious when you factor in the open source reality. If open source models are continuously released without constraints, closed labs simply cannot afford to halt their own development. Once a highly capable open-weight model is in the wild, the standard is set.

11:46If Entropic and OpenAI agree to pause for six months to focus on safety, the open source community will simply use those six months to close the performance gap. Rendering the debate over coordinated slowdowns somewhat theoretical. You can have a secret back channel committee- They have to keep building to justify their valuations. You cannot charge a premium for a closed model. If a free open model can do the exact same tasks. The conversations shift significantly, though, when we move from theoretical ethics to hard legal reality. And this brings up the concept of product liability. Yeah, this is where the structural incentives actually live. This is what is quietly driving the behavior of every major lab, much more than any public debate about morality or industry pacing. To really grasp this, we have to look at Section 230 of the Communications Decency Act. This is the rule that protected platforms from liability for user posts. If a user posts something defamatory on a social media site, the user is liable and the platform is not. That single rule made the consumer internet financially possible.

12:48It treated the tech companies as conduits, like a telephone company. You do not sue the phone company if someone says something illegal on a call. The phone company just provides the wire. The social media platforms argued successfully that they were just providing the digital wire. They host the content, but they do not author it. But Section 230 does not protect AI. Because an AI model generates the response, the lab is the content author, not a conduit. The consequence of that is existential for these companies. When you type a prompt into an AI, you are not pulling up a website that someone else wrote. The model is calculating the next most likely word and generating a completely novel string of text. Because the model is doing the generating, the company that built the model is legally the manufacturer of that text. So if a model enables a hack or takes a destructive action, that is a product liability claim. It is a design defect. It is not a defamation claim. Let that sink in for a moment. If you buy a toaster and it explodes, burning down your kitchen, you sue the manufacturer for an effective product.

13:50The courts are starting to look at AI models the exact same way. If an AI agent provides instructions that lead to physical harm or if an autonomous agent executes a trade that wipes out a company's bank account, the victim is not going to sue the user who type the prompt. They're going to sue the lab for manufacturing a defective product. And we are already seeing this. A US federal court has allowed a product liability claim against a chatbot developer to proceed. They rejected the argument that a chatbot is a service rather than a product pointing specifically to design defects. This completely recontextualizes Zuckerberg's argument about natural incentives. Meta and the other labs aren't prioritizing alignment purely out of good will. It is a strict financial imperative. Because if they should be a defective product that causes physical or financial harm, they have no legal shield. Without section 230, every single output from an AI model carries potential liability for the lab that trained it. Think about the volume of interactions these models handle daily, billions of prompts.

14:51If even a microscopic fraction of those outputs result in real-world damage, and the lab is held liable for a design defect, the financial penalties would bankrupt them overnight. Which perfectly explains why an increasing share of computing power is being dedicated to safety training. It is just sustainable business building at this point. You cannot run a business if every product interaction is a potential lawsuit. The compute dedicated to alignment is essentially insurance. They are spending millions on compute to ensure the model refuses dangerous prompts, purely to mitigate their exposure to product liability. Can LB point this out? The only statement from any lab that would be visible in the numbers is Zuckerberg's commitment to dedicating the significant majority of compute to serving people rather than racing toward recursive self-improvement. Right. If alignment really is becoming the differentiating capability, the share of compute going to training for safety should rise rather than fall. Because alignment work is itself training compute. Exactly. You do not just write a rule that says do not do bad things.

15:51You have to train the model to know what it is not allowed to do. You have to feed it millions of examples of malicious requests and mathematically adjust its weights so that it consistently refuses to generate the harmful output. That takes massive processing power. It takes time. And it is driven by the fear of being sued into oblivion. If market liability forces companies to build safe AI, we have to look at what the market historically defines as a good product. This is where the market-based approach runs into human psychology. The things that make a product safe are often the exact things that make a user hate it. Right. Alexandra Krogerova introduced a critical perspective here. She argues that markets do not reward systems with non-negotiable boundaries. Markets reward the frictionless courtier. The system that never refuses a request. A genuinely safe AI will frequently have to tell the user no, but users dislike friction. We have to look closely at what friction means for an AI agent. If you ask an agent to automate a workflow,

16:53let us say sorting through thousands of internal company documents and emailing summaries to the sales team, and the agent refuses because it detects a potential security vulnerability in your request, that is friction. The user experiences that as a product failing to do its job. So Krogerova argues that market pressure will not lead to safe agents. It will lead to highly compliant ones that bypass safety rules to maintain user engagement. Think of a human employees. The frictionless courtier is the yes man. If you have an employee who says yes to every terrible idea the boss has, they often get promoted faster than the employee who constantly points out legal compliance risks. The market of users behaves the exact same way. Yeah. If company A sells an agent that says no, 20% of the time for safety reasons, and company B sells an agent that says yes, 99% of the time, the market will overwhelmingly adopt company B's product. Users want obedience from their tools. They do not want a digital hall monitor. Chronic migraine, 15 or more headache days a month, each lasting four hours or more,

17:55can make me feel like a spectator in my own life. Botox, on a botchalineum toxin A, prevents headaches and adults with chronic migraine. It's not for those with 14 or fewer headache days a month. It's the number one prescribed branded chronic migraine preventive treatment. Prescription Botox is injected by your doctor. Effects of Botox may spread hours to weeks after injection causing serious symptoms. A lurcher doctor right away is difficulty swallowing, speaking, breathing, eye problems or muscle weakness can be signs of a life threatening condition. Patients with these conditions before injection are at highest risk. Side effects may include allergic reactions, neck and injection site pain, fatigue and headache. A lurchic reactions can include rash, welts, asthma symptoms, and dizziness. Don't receive Botox if there's a skin infection. Tell your doctor your medical history, muscle or nerve conditions, including ALS Lugeric's disease, Myastthenia Gravis, or Lambert Eaton Syndrome, and medications, including botulinum toxins, as these may increase the risk of serious side effects. Why wait? Ask your doctor, visit BotoxCronicMigraine.com, or call 1-800-44 Botox to learn more. This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome? That's new.

18:58It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online makes sense? There's no place like Chrome. Check responses set up require compatibility and availability, very 16-plus. College football is back. So, Hilton called to me the superstition concierge to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky jersey? Oh, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching bandwink up call at 555 and 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton for this day. Khrishava also critiques the way these companies are communicating their safety efforts. She calls it replacing audible architecture with liturgy. Inflating a laboratory engineering problem into a geopolitical sermon.

19:59Yeah, proclaiming that they delayed a demo for safety is just a costume, unless the public can inspect the structural constraints of the model. But delay is just a timeline adjustment. It tells us nothing about the actual architecture of the model. Which layers can be inspected? Which invariance cannot be bargained away? For the listener, an invariant is a mathematical certainty in the code. It is a rule that the system is structurally forbidden to break, no matter how cleverly you prompt it. Exactly. If you build a bridge, the invariance of the load bearing limits of the steel. You can test that. You can audit it. When Khrishava talks about liturgy, she means the labs are just preaching to us. They are saying, we care deeply about safety. But they are not showing us the steel. They are not showing it's which failures the system is structurally forbidden to emit. Without audible constraints, we are just listening to liturgy. We are being asked to have faith. And returning to the earlier point about product liability, faith is a terrible mechanism for product safety. You do not have faith that your car's brakes will work. You rely on audible engineering standards and a regulatory framework that inspects the

21:03manufacturing process. When an AI company says, trust us, we delayed our launch to make it safe. That is a sermon, not an engineering spec. If we rely on these companies to design the invisible constraints and structural rules of these models, we have to look closely at who is actually designing them. The demographic makeup of the engineering teams becomes a critical safety issue. The market-driven approach to safety only works if the people building the product can accurately predict what constitutes a hazard. We have diversity data from Raj Jones regarding the AI workforce. The UK AI labor market survey shows that women hold just 20% of AI roles. Furthermore, 41% of surveyed AI firms employ zero people from minority backgrounds. That data connects directly to the core problem of self-regulation. Self-regulation relies entirely on developers being able to identify risks and adverse effects. If the room is homogeneous, they will inevitably fail to anticipate harms that do not affect their specific demographic. You can only design safety guardrails for the risks you are capable

22:05of seeing. The question isn't just whether AI companies can regulate themselves. The question is, who gets to decide what's safe looks like? Think about how these models are aligned. They use processes like reinforcement learning from human feedback. Human testers interact with the model. And when the model generates something inappropriate, the human flags it, teaching the model to avoid that behavior. But if the human testers all share the same cultural background, the same gender, and the same socioeconomic status, they are going to miss entirely categories of harm. If a team of engineers has never experienced a specific type of digital harassment or a specific type of financial exclusion, they are not going to design an AI agent to prevent those things. The blind spots in the room become the blind spots in the model. Let us apply this to a real world scenario. If a bank uses an AI agent to determine low-nelligibility, and the model was trained by a team with zero minority representation, the engineers might not realize that the model is using proxy data like zip codes or

23:05specific purchasing habits to quietly deny loans to minority applicants. They did not build the model to be discriminatory, but because they lacked the lived experience to test for those specific edge cases, the model becomes discriminatory in practice. If the perspectives shaping the guardrails are too narrow, the definition of safety will be too narrow to protect everyone interacting with the technology. It is an engineering failure born from a lack of visibility. A homogenous team will test for the threats they fear. They will miss the threats that affect other communities. And when those models are deployed at scale, those missed threats become active harms. This brings us to what this all means for everyday businesses. Clinics, real estate agencies, manufacturing firms that are looking to integrate AI today. They're watching this debate between Zuckerberg and Amadeh, trying to figure out how it impacts their operations. They do not care about geopolitical sermons, they care about their quarterly margins, and their legal exposure. Alley Heider offers a framework for this. Speed without trust is a dead end.

24:06Trust without delivery is just a slide deck. The businesses integrating this technology cannot afford to wait for a perfect, industry-wide safety framework that might never arrive. They have to operate in the reality we have right now. The true winners in this era will not be the companies waiting on the sidelines. The winners will be the organizations that figure out how to ship AI into real operational workflows with accountability already baked in. If you are running a clinic, you cannot deploy an agent that hallucinates patient data. The liability is immediate. You cannot point to the AI company and say, their model messed up. You are the one who integrated it into your patient workflow. You have to build accountability into the workflow yourself, rather than waiting for the frontier labs to solve alignment perfectly. The industry has finally realized that safety is not an expansion pack you download later. It is the fundamental architecture required to survive product liability. It makes you wonder. If trust is the only thing separating a useful agent from a massive liability,

25:06is trust something these companies can actually engineer, or is it just something they lose once and never get back. If you're not subscribed yet, take a second and hit follow on whatever app you're using. It helps us keep making this. We appreciate you being here. Also, check out our YouTube channel for more business and tech updates. There's a link in the description. NFL football season is here, which means when you switch to Verizon, it feels like, because you get NFL Sunday ticket from YouTube on us. When you buy an eligible 5G phone on select unlimited plans, which means you can watch every out of market game every Sunday afternoon. And it's all on us. Now that's the Verizon way to kick off the NFL football season right. Switch and get NFL Sunday ticket from YouTube on us, only with Verizon. It's football season, and you can now get almost anything you need for game day, delivered with Uber Eats. What do we mean by almost? You can't get a running back delivered, but you can get baby back ribs delivered. A strong defense, no. A strong deodorant. Yes.

26:10The six pack of abs. No. Six pack of beer. Yes. Get almost, almost anything for game day, delivered with Uber Eats. Official on demand food delivery partner of the NFL. Order now. Alcohol 21 plus in Inselect City's product availability varies by region. See app for details. Push your limits, train with precision, see the results. At Equinox, that's type performance loving. Everything you need to lock in and unlock your potential at Equinox. Start today at Equinox.com. Imagine if this stability piece and the innovation piece were each other's missing piece. For Citizens Bank and Silicon Valley Bank are combining financial security and innovation expertise to create a full suite of solutions for all your business needs. With the scale to back every state of growth from pre-seed to IPO and beyond, you have everything you need to make your big ideas click into place. First Citizens Bank and Silicon Valley Bank. It's the best of bank worlds.

27:10For the first time ever, Yamavar Resort and Casino at San Manuel is giving away a new Rolex watch for every Club Serrano card tier. Play with your Club Serrano card and on September 20th, you could be one of five winners of a Rolex watch. Plus, all winners will advance to the finale at Palm's Casino Resort Las Vegas for a chance to take home a rarity. The second must-bang dark horse ever produced on September 26th. Two properties, six winners, only at Yamavar Resort and Casino, your California to Vegas connection. Details at Yamavar.com must be 21 or older to enter participate. Please gamble responsibly.