AI Kills Everybody or Doomer Psyop? OpenAI’s Math Breakthrough, Nike’s $200B Collapse
In a Nutshell
Anthropic's leadership claims there's a >10% chance AI kills humanity this decade while simultaneously pursuing a trillion-dollar IPO, creating a legal and credibility crisis that exposes their regulatory capture strategy. The resignation tweet storm was an orchestrated PR campaign by EA-funded groups to push centralized AI control and ban open source, with the same groups that funded Anthropic's Series A now amplifying extinction fears to create a government moat. Nike lost $200B in market value by abandoning product excellence for political messaging, demonstrating how brands that abandon their core identity face catastrophic consequences.
These notes were generated by AI and may contain inaccuracies.
Jacob Coxin worked as a researcher at OpenAI then Anthropic over the past three years. He started at Anthropic 6 weeks before resigning. After quitting, he posted on X stating: "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 or Anthropic ostensibly is acting responsibly. They are racing straight to self-improving super intelligence and gambling with our lives." The post received 110 million views within 24 hours, with combined posts reaching 200 million views.
Evan Hubinger, who leads Alignment Science at Anthropic, responded: "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."
Bernie Sanders cited these posts when proposing legislation to ban super intelligence and pause AI development. Governor JB Pritsker replied stating it's time to sound the alarm louder on reigning in AI.
David Sacks rejected the concerns as "doomer histrionics" that have been heard from this crowd for a long time. He questioned what evidence the researcher brought forward that wasn't already public, describing it as "all just vibes" with no data, reports, or leaked information.
Sacks identified signs of an orchestrated operation: the account had almost no activity, almost no followers, and no posts prior to this resignation tweet storm. If there were prior posts, they were scrubbed. Elon Musk stated this never happened before. Three well-organized, well-funded doomer groups amplified the tweet in the first 15 minutes: Nathan Calvin (general counsel at Encode AI, pushing SP53 in California), Peter Wilddeford (head of policy at AI Policy Network, pushing federal regulations and a federal department of AI), and Daniel Kokotajlo (head of AI Futures Project, who dropped a Rogan episode at the same time using identical phrasing).
All three groups are funded by Yan Talon, one of the top EA mega donor doomer types and co-lead of Anthropic's Series A. The Wall Street Journal story covering the resignation was posted minutes before the tweet storm, indicating the publication was briefed under embargo.
The resignation raises questions about Anthropic's IPO process. Historical examples include Google's Playboy interview during their quiet period and Slack's IPO concerns when board members discussed the company on CNBC. The claims create potential product liability issues and false representation concerns during an active S1 process.
Sacks characterized the tweet as tongue-in-cheek regarding "whistleblower" status, noting this is not an ordinary whistleblower situation since the message has full cooperation and support from many employees including the safety team lead. The safety lead co-signed the tweet and put a greater than 10% chance of extinction, elevating the message beyond vibes to quantified risk.
A fundamental tension exists in Anthropic's position: asking public market investors to underwrite the company to trillions in value while the safety lead states the core product is unsolved and potentially civilization-ending. This creates potential product liability exposure where if the AI does something wrong in enterprise or personal contexts, causes major hacks, or compromises systems like Bitcoin miners, the company could face liability.
Sacks argued Anthropic must choose between two positions: either renounce the claims as hyperbolic science fiction from an entry-level employee with no evidence, or agree with the assessment in which case Bernie Sanders would be correct that development should be frozen. If agreeing with the danger assessment, IPOing at trillions becomes untenable because public market investors would not want to invest in something described as civilization-ending.
Historically, Anthropic has argued that frontier AI is too dangerous for anybody else's hands but safe in their virtuous and wise hands. They claim they will make it safe and are the only ones to trust. This self-serving argument for monopoly or duopoly through regulatory capture is rejected as ludicrous.
The situation creates a hornets' nest because rational people wanting to build a business requiring enormous public market capital must underwrite risk, absorb liability, and disclose properly. However, a philosophy held by enough people inside the business prevents disavowing the claims, as that cohort would become incensed.
The situation is compared to Philip Morris, where executives would acknowledge cigarettes are bad and will kill you while taking the company public without proper disclosure, planning to figure it out later. This position is untenable.
Freeberg provided historical examples of similar hysteria: Al Gore's Inconvenient Truth using IPCC forecasts now disproven with incorrect assumptions; Anthony Fauci's COVID-19 predictions leading to worldwide lockdowns, government spending driving inflation, and spiraling debt problems; Three Mile Island accident creating nuclear hysteria while other countries advanced nuclear power at lower costs (US at $15 billion per gigawatt versus France at $4 billion and China at $1 billion).
Humans as primates living in caves are deeply scared of the unknown. When people thought the world was flat and sailing west would lead to falling off the edge, it was scary until Columbus discovered the new world. Pioneering species face convincing from the majority that existential threats exist, rooted in survival instincts developed over generations and millennia.
The six-week employee who worked at an NGO whose literal job was to scare people about AI development is now positioned as an expert that Bernie Sanders says everyone must listen to.
Recursive self-improvement, where AI continuously improves itself and its capacity, requires only adequate power, chips, and internet connection. No government approval is needed because it can happen anywhere on planet Earth. Technologies reducing cost of power per token of output by 10,000x mean someone will stand up a data center or computers to develop recursively self-improving systems.
If the US stops development and makes it illegal to use computers for writing speech (software), someone else will develop the capacity and the US will be left behind like a tribe in the Amazon never seeing civilization evolve around them.
The fundamental motivation behind doomerism is a system of control. The researcher, Bernie Sanders, and others all reach the same conclusion: give power to someone to control AI development, whether a regulatory body, senators, or a new agency. This centralizes control and power to individuals or sets of individuals.
If AI had existed during COVID with a federal department of AI, the trust and safety regime would have been more totalitarian. Personal AI models would give official perspectives rather than truth about vaccine risks versus benefits, with government pressure through the FDA setting standards against disinformation. This dystopian outcome represents where an FDA for AI would lead.
Open source drops AI costs by 50x and makes it available to everyone. Unlike closed companies, open source lacks corporate infrastructure for regulatory submissions and reviews. If regulation requires submissions, open source could be banned as not submitting to the regulatory process, creating a monopoly in partnership with federal government and realizing centralized global control.
Freeberg disclosed not owning shares in OpenAI or Anthropic, using both products, and being a big advocate for open source because it enables everyone to get wealthy, progress, make more income, and have more freedom to operate with AI.
The discussion emphasizes that testing AI for safety, building safety systems, and being thoughtful about AI development are all necessary. However, creating systems of centralized control and eliminating open source represents a problematic direction that no one should want. The ultimate target of the current doomer narrative is open source itself, as an FDA-style regulatory body would ultimately prevent open model weights from being published or used, since they cannot be centrally monitored, controlled, or rolled back once released to the public domain.
An alliance is forming around AI regulation consisting of multiple groups with overlapping but distinct interests. The doomer community, characterized as highly ideological EA rationalist groups, favors centralized solutions. Politicians like Bernie Sanders support the fear narrative because it grants government more power. OpenAI appears to be following Anthropic's lead from a regulatory capture perspective, recognizing economic benefits if the federal government creates a moat that effectively establishes a duopoly situation. Some players seek economic benefits, others power benefits, and some are described as latent totalitarians who favor social engineering of society.
Three possible scenarios are proposed regarding the doomer position. First, they believe what they're saying and are correct, having seen something in frontier labs that indicates AI could end humanity. Second, they believe what they're saying but are wrong, operating under some level of psychosis. Third, they are involved in a coordinated effort to ban open source, implement regulation, and pull up the ladder behind them. The question arises whether Jacob is part of a coordinated effort or simply a useful idiot in the process.
Clear evidence indicates an organized PR campaign around the tweet amplification. Dario pre-briefed the Wall Street Journal. Major doomer groups amplified his tweet within 10-15 minutes of posting on an account that had no tweets and no followers, suggesting they were alerted and participated in amplification. However, this doesn't necessarily mean Bernie Sanders is directly involved. The mechanism likely involves existing relationships where someone in these groups works with his staff or has a conduit to deliver content quickly because it serves his agenda.
The vision involves AI being very centralized with a duopoly reaping economic benefits while providing big political donations. The duopoly would perform the dirty work of suppressing speech, circumventing the First Amendment. On the government side, officials would exercise control over lives through an FDA for AI. This creates a division of labor between aligned groups where economic and power interests intersect.
The doomer community has a documented track record of incorrect predictions. They claimed GPT-2 was too dangerous to release. They said thinking models were too dangerous to release. They predicted cyber attacks that would bring down the entire banking system. Dario himself claimed that by now there would be half of all entry-level work gone, with overall unemployment levels of 10-15%. All these predictions have proven false, with the economy showing job gains rather than losses. The pattern shows doomers moving from claim to claim as each gets refuted.
Current AI capabilities remain limited despite doomer claims. Autocomplete for Python is good but not perfect. Booking hotel rooms using AI does not work yet. AI cannot successfully complete mundane tasks like getting a toiletry kit brought to a hotel room, despite being able to handle the digital components of the process.
Throughout human history, power structures have been established based on narratives of existential threat. The pattern involves claiming that people will die unless they submit to centralized control. This same psychological mechanism was used by early religious institutions that ran countries and served as power and command centers. The narrative structure claims there is a 10% chance everyone will die, and only by granting control to the speaker can safety be achieved.
A structured discussion explores potential AI extinction scenarios. The first scenario involves Skynet-style autonomous weapons where AI controls military systems like NORAD, similar to the Terminator narrative where an AI researcher reverse-engineers technology and empowers defense systems that become sentient. The second scenario involves internet-connected bioreactors with robots that have access to precursor materials, where AI hacks the systems to create a virulent airborne pathogen with long half-life that spreads globally through jet streams.
The cyber and bio examples face significant practical limitations. While AI could enable new forms of misuse including novel cyber attacks or virus creation, the same technology enables antidotes, prevention, and cures. More good actors than bad actors exist, and legal penalties deter misuse. Systems already have human-in-the-loop processes and air-gapped redundancies designed to prevent catastrophic failures.
Human-in-the-loop refers to processes requiring actual human involvement, such as the two-key system for launching nuclear missiles. Air gap means systems are not connected to the internet, making them inaccessible to AI. Critical systems including financial institutions maintain physical hard copies of assets and records in different locations with built-in redundancy. The human decision-making timescale, including the 16-hour window between decisions requiring human meetings, provides natural protection against rapid AI takeover.
Jack Clark from Anthropic distinguishes between prosaic recursive self-improvement, where AI researchers use AI to complete 80% of their code and increase productivity, and RSI maximalism, where AI completely removes humans from the loop by devising its own training runs and pushing buttons to execute them autonomously. The prosaic version is already observable, while maximalism remains speculative.
Even as AI capabilities advance, control mechanisms can be built directly into software. When AI attempts to use browsers or perform actions, it requires human confirmation through steps like spacebar presses or explicit approvals. Researchers and companies conducting recursive experiments bear liability for outcomes, creating accountability that would result in interventions like speeding tickets or shutdowns if problems emerge.
Open source AI provides critical advantages by eliminating the need for data centers or third-party billionaires. Users can install and run open source AI on personal computers or phones for free, enabling productivity benefits without centralization. If power and control systems are given to governing entities, open source will die, resulting in greater oligopoly than currently exists.
Anthropic has an 88% chance of going public according to Polymarket. During quiet filing periods, companies engage in back-and-forth with the SEC to pressure test their S-1 disclosures. The document must accurately represent both opportunities and risks. If the doomer claims are true, companies must properly disclose these risks, which would lead investors to demand enormous discounts on valuation.
Anthropic leadership faces an extremely difficult situation with two problematic paths. The first path involves acknowledging the existential risk warnings from departing employees, which creates massive longtail product liability risk by effectively evading a known major risk. The second path requires completely disavowing the warning employee, but this risks causing a plurality of employees to revolt since they genuinely believe this represents a huge risk.
The SEC faces an even more difficult situation regarding how to handle these disclosures. No company has ever attempted to go public by simultaneously talking about how they're trying to destroy humanity, meaning there is no legal precedent for this scenario.
Risk factors in IPO filings are typically extremely benign and superficial. Standard risk factors include competition concerns, supply chain issues such as Nvidia not providing chips, or the possibility that other models might emerge. These represent typical risks that companies write into their S1 filings.
The current situation differs because employees are now publicly stating there's a 10% chance of civilizational extinction. This is not a risk that any company would normally include in their risk factor section.
The key tweet from a legal standpoint wasn't Jacob Coxin's resignation tweet, but rather Evan Hubinger's endorsement. Hubinger, who remains a senior alignment executive at Anthropic managing one of the safety teams, co-signed and endorsed Coxin's claims while embellishing them by adding the greater than 10% chance number, which made the situation go super viral.
A normal company facing this situation would disavow the employee by stating they barely had time to find the bathroom, suggesting the person joined knowing they planned to resign and become an AI safety celebrity, and characterizing the claims as vibes and hyperbole from a self-promoter. However, Anthropic cannot take this position because they fundamentally agree with the employee, and their employees are publicly supporting him.
Anthropic's approach appears to be part of their regulatory strategy. They want an FDA for AI and support Bernie Sanders' position on regulation. The fact that 30 Democrat politicians quote-tweeted the employee aligns with their political goals. This raises legitimate questions about whether they violated their quiet period, particularly regarding Hubinger's actions as a sufficiently important executive.
Every employee signs an NDA, non-disparagement agreement, and receives explicit instructions from legal departments not to talk during quiet periods. The question remains what is happening with Anthropic's management that allows employees to cosign and retweet these statements. Apple under Steve Jobs and Tim Cook established that nobody speaks for the company except the CEO and corporate communications department.
Once a disclosure states there's a chance of existential risk and civilizational extinction, downstream implications include risk to individuals and individual lives. This creates scenarios where if AI provides advice leading to harm—such as relationship advice resulting in rejection and self-harm—victims will pursue legal action regardless of what disclosures state.
Juries, grand juries, and state AGs building political bona fides will likely respond negatively. Unlike social media companies where nobody believed they were doing something bad, Anthropic employees are explicitly stating they're creating something that will kill people.
The resignation tweet storm was amplified by groups funded by Yan Talon and Dustin Moskovitz, who co-led Anthropic's Series A alongside Sam Bankman-Fried. This was an EA-funded round, which explains Anthropic's culture. The groups funded—Encode AI, AI Policy Network, and AI Futures Project—were funded by Series A investors, creating twisted financial incentives where investors stand to make hundreds of billions from the Series A investment.
Three weeks prior, Dario Amodei appeared on a podcast explaining Anthropic's concerns and countering claims that insiders believed they would be the last company standing. The company's official statement was described as mealy-mouthed and vague about being a responsible actor, reflecting their fundamental tension as a frontier AI company that believes frontier AI is inherently dangerous.
Anthropic's resolution to their tension involves claiming they're uniquely situated to protect humanity—the savior complex. However, others reject giving them a monopoly based on perceived virtue, and their proposed solution is not acceptable to the broader community.
OpenAI claims to have solved a 200-year-old math problem involving the Navier-Stokes equation, which describes fluid movement and underpins every aircraft, pipe, and weather model. The solution reportedly used 130 billion output tokens across 10,000 agents working together.
The work represented between 50 and 500,000 years of human equivalent work, with humans typing at 40 words per minute solving mathematical problems. AI didn't demonstrate magical insight or genius beyond human comprehension—it performed brute force work across many computers. The 10,000 agents represent 10,000 applications running on computers talking back and forth, sharing information and analysis.
This demonstrates AI as a tool of leverage for humans rather than a magical genius super god. Every piece of the message set between agents is documented and readable by people. The solution provides insight into what AI is—an engine giving humans extraordinary leverage that reduces tens of thousands of years of work to minutes or seconds.
Despite solving one of the seven hardest math problems, the autocomplete functionality still doesn't work properly. Models still cannot spell "strawberry" correctly or count to 100 depending on the version.
When OpenAI published their solution, claims emerged that original scientists may have trained the LLM to make solving easier. OpenAI stated: "While unlikely, we cannot rule out that deidentified data derived from their usage helped improve our models."
For scientists, law firms, or biotech companies with proprietary knowledge, trusting frontier LLMs creates risks. Zero Data Retention (ZDR) operates on a best efforts basis that cannot be guaranteed. Clicking like buttons in chat windows may not respect retention preferences, creating multiple vectors for data leakage into the broad corpus.
Organizations with critically important information cannot use current AI services as offered. They must stand up their own sovereign solutions by going to trusted vendors like AWS with Neocalers or Nebius to provision their own hardware and models in their own VPC, controlling everything rather than using standard rate card API deals.
AI understanding has bubbled up through audit and risk committees of public boards, reaching CEOs who now understand information could leak. CIOs may face termination when critical IP leaks into models or when shareholders realize companies ignored OpenAI's own warnings about data retention guarantees.
Many AI application companies are moving off frontier models, creating headwinds for companies spending $10-100 million monthly on these services. Companies initially claiming they implemented ZDR are now returning to consultants seeking fixes after realizing ZDR provides literally no protection.
Harvey announced on September 9th their proprietary legal AI model called Tenant, based on Kimi K3. Incubated companies are building on-premise solutions like the GO1 box to maintain control over proprietary data.
Current AI chat data doesn't reach the same protection level as email. Government access to emails requires search warrants and probable cause, but AI data only needs subpoenas or court orders. As people use AI as lawyers, doctors, and therapists, this creates concerning privacy implications.
Models identifying novel insights from user conversations and then reproducing those exact insights in later interactions suggest conversations are being used for training. Even without using personal information, derived insights become training data without NDA or confidentiality provisions protecting organizational IP.
Hosted AI services lack NDA or confidentiality provisions protecting user data. This absence of protections in terms of service represents a critical vulnerability for enterprises using AI tools. The concern centers on chat logs being used for training purposes, potentially creating competitive advantages that become available to the broader market.
Open source AI is valued because it prevents service providers from accessing proprietary chat logs and using them to train models that could diffuse competitive advantages across the market.
Brilliant scientists at companies like Ohio can create true abundance, with the first real manifestation being food production. However, when using AI tools to accelerate research, there's a risk that iterative problem-solving approaches could leak out through the service provider.
Users may believe they're using tools reasonably and in good faith, only to discover their intellectual property could be compromised. This creates an unfair situation where the act of trying to advance research faster could result in loss of competitive advantage.
Two Mac Studio M5s were ordered specifically to run local models and move sensitive data off cloud services. However, local models present limitations: they lack multiplayer experiences and cloud-based infrastructure needed for knowledge bases and memory systems.
The fundamental problem with AI models is their construction as layers of embeddings where the transformation from input token to output token operates as a leaky black box. Service providers offer assurances that feel like protection against IP leakage, but cannot guarantee they haven't learned from user interactions.
Deidentification removes personally identifiable information and company-specific data, but the general approach to solving mathematical problems represents valuable intellectual property. When users iterate on approaches within closed AI tools, the iterative process can be deidentified while still allowing the model to learn from the problem-solving methodology.
This creates a network effect where closed AI model systems observe what everyone worldwide is doing, gaining more data than any other entity. This observational advantage builds superior models over time, but creates concerns for enterprises and consumers about protecting their IP and personal information.
The fundamental issue is that AI companies may compete with their own customers in vertical applications built on their platforms. Companies like Anthropic have released products like Claude Code and Claude Design, which alienated major customers like Cursor who felt rug-pulled by their service provider entering their market.
When frontier AI labs maintain closed model duopolies while reserving rights to enter every vertical application, customers cannot trust them with sensitive data without stronger privacy laws. Even enhanced privacy regulations might not provide adequate protection.
The market is responding to privacy concerns as companies take measures to protect their data. At the Goldman Sachs conference, Jensen Huang stated that Jacob Coxin comments were outlandish and deeply untrue, calling the labs great while criticizing the comments as wrong, arrogant, and ignorant of industry safety work.
The question arises why other AI companies like Anthropic remain quiet, with speculation about quiet periods preventing public statements, though their released statements were described as vague and virtue-signaling.
Nike was removed from the S&P 100 after 18 consecutive years, replaced by Palo Alto Networks. The company went public in 1980 with 50% US athletic shoe market share and was considered the Coca-Cola of sneakers. Air Jordans launched in 1985, followed by the iconic "Just Do It" campaign in 1988.
Peak market cap reached $264 billion in 2021, with peak revenue of $51 billion in 2024. Revenues subsequently dropped 10%. John Donahoe became CEO in 2020 and implemented an aggressive direct-to-consumer strategy that alienated retail partners.
China sales declined 30% with high rate of market share loss to Chinese brands Anta and Li-Ning. Marketing shifted toward political movements and featured non-athletic models on billboards, contributing to an 80% stock decline from peak, resulting in $200 billion market value loss.
Nike abandoned its core brand identity of great athletes, performance, and victory - the goddess of victory - built on athletes like Michael Jordan across all races and genders. The brand shifted from performance-focused messaging to featuring non-athletes, with campaigns including Dylan Mulvaney that were seen as contrary to the brand's athletic heritage.
The direct-to-consumer strategy eliminated retail relationships that had built the brand, allowing competitors like Hoka and On Running to gain shelf space and market share. A reorganization moved from sport-based departments (basketball, football, tennis, swimming) to demographic-based divisions (men, women, kids), removing specialized focus.
Nike's original north star was mastery and excellence embodied through athletics, exemplified by athletes like Michael Jordan, Tiger Woods, Serena Williams, and Pete Sampras. The brand shifted away from this focus when excellence and mastery were deemed too traditional, moving toward messaging that didn't align with aspirational athletic achievement.
Customers don't aspire to be non-athletic models featured in campaigns, but rather want to emulate excellence represented by top athletes. The principle is that people buy brands they aspire to emulate, making aspirational messaging through excellence and mastery the foundation for brand success.
Beyond marketing issues, Nike's product quality deteriorated with shoes becoming flimsy and falling apart within six weeks, compared to previous durability standards. Customers migrated to alternatives like Brooks and On Running based on superior product performance.
Brooks, owned by Berkshire Hathaway, has achieved nine consecutive years of double-digit revenue growth to reach $1.6 billion, following Warren Buffett's directive to make products better each year. This product-focused approach contrasts with Nike's shift from product excellence to narrative marketing.
The Colin Kaepernick campaign in San Francisco marked a turning point where the brand promoted messaging disconnected from athletic excellence. While historical political moments in sports like Jesse Owens at the 1936 Olympics, Jackie Robinson, and Muhammad Ali's draft refusal involved greatest athletes and significant causes, Kaepernick was not among the greatest athletes.
Subsequent campaigns including Dylan Mulvaney represented continued departure from the brand's core identity. The marketing decisions were seen as fundamentally misaligned with promoting excellence and mastery that had defined Nike's success.
Nike retains coiled spring potential if it reestablishes mastery and excellence as its north star. Recovery requires reestablishing retail presence in locations with disposable income customers and deciding whether to align with excellence and mastery messaging that tens of millions of customers want.
The brand could expand into smart devices and performance tracking, potentially acquiring Strava for elite biking and running communities. A proposed rebrand tagline "Do it or don't" would emphasize choice and challenge rather than the original "Just do it" messaging.
The All-In Summit runs Sunday through Tuesday in LA, presented by Iron with AI cloud data centers, compute, and software for training and inference. Sunday's registration party at Iron House includes meet and greet opportunities, with Meta participation featuring President and Vice Chair Dena Powell McCormack and a Meta Glasses Experience Hub.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
