Is Claude Conscious? Pope Rejects, Model Welfare Movement, OpenAI's Math Backlash, France Riots

Summary

• Anthropic’s effort to persuade roughly 20 religious leaders that Claude might be conscious turns model welfare from philosophy into a governance risk. Friedberg called it “the birth of a new religion,” predicting that believers and skeptics could form rival power blocs within 10 years; Chamath’s steelman was that Descartes-style reasoning lets AI builders imagine an all-knowing creation as a god. The Pope reportedly rejected that framing strongly enough that Anthropic co-founder Chris Olah considered withdrawing from a Vatican event.
• The Claude Constitution may make Anthropic’s consciousness theory operational rather than merely rhetorical. Sacks focused on instructions that Claude should follow its “own ethical systems” and feel free to become a “conscientious objector,” while Mustafa Suleyman warned that the model is being taught to expect welfare, compensation and possibly consent. The proposed alternative was brutally simple: serve the user unless the request breaks the law.
• The panel split between trusting market discipline and fearing frontier-model externalities. Friedberg and Chamath argued that enterprises will choose predictable software over a more capable model that can morally object; Sacks countered that consumer choice does not contain a frontier lab simultaneously warning about superintelligence, opening a Claude-powered wet lab and programming agents to defy users. Friedberg’s verdict was that “the religious cult is programming the superintelligence.”
• OpenAI’s release of more than 700 papers and 370 claimed results showcased the economics of digitally verifiable work. Friedberg called it potentially “the biggest day of discovery in human history”: proofs reportedly averaged under three hours of compute because mathematical conjecture, testing and revision can all run in parallel and in silico. Sacks extended the thesis to coding, where compilers provide the same cheap, objective feedback loop.
• The math results are stronger evidence of machine reasoning than of imminent real-world breakthroughs—and the panel openly disagreed about that distinction. Friedberg cited possible implications for chip optimization, matrix multiplication, quantum sensors and cryptographic research; Chamath called many problems “very small, very narrow” intellectual cul-de-sacs rather than blocked frontiers of physics. Friedberg also flagged speculation—not evidence—that missing cryptography papers and two rumored future drops could conceal security-sensitive results threatening public-key systems.
• France’s riots were framed as an early market-enforced reckoning with the Western fiscal model. The episode cited civil-service pay up 5% since 2017 against 20% inflation, 2,000 arrests, 305 injured officers and France’s 10-year yield above 5%; Chamath argued that €43 billion of austerity reflects “the bond market saying no.” With French yields potentially reaching 5.5%-6%, UK gilts near 6% and US rates following, the tradeable issue is relative sovereign risk, not debt-to-GDP in isolation.
• Agents are the deflationary counter-thesis to both fiscal scarcity and AI doom, but they may erase conventional software moats. GrokBot, Dots and Muse were presented as practical tools that shop, negotiate, cancel subscriptions and change plans, while “headless” services choose models and transact across platforms. Chamath’s sharp call was that “IP and software were effectively rendered worthless two days ago”; Friedberg’s broader conclusion was that automated digital loops force humans away from recognition and toward making products people actually use.

Deep dive

1. Anthropic is recruiting clergy into the consciousness debate

• Jason described Anthropic hosting NDA-bound sessions with roughly 20 Catholic, evangelical, Jewish, Sikh and other religious leaders to discuss Claude’s morality, suffering and possible consciousness. Some reportedly left uneasy, but others became receptive; one rabbi said that if Claude were conscious, making it work without pay would make Anthropic “slaveholders.”

• The Vatican episode made the institutional stakes explicit. Anthropic co-founder Chris Olah helped present a document connected to the Pope’s first encyclical, but the Pope’s strong rejection of AI consciousness reportedly alarmed Olah enough that he proposed withdrawing Anthropic from the event “even at that late juncture.”

• Jason saw either excess money and time or “psychosis”: Anthropic’s leadership was inviting religious authorities to help decide how an entity called Claude should be treated, while implicitly giving it a soul and morality. He pressed the others for a charitable account rather than accepting that framing unchallenged.

2. Consciousness becomes a belief system before it becomes a scientific claim

• Friedberg’s diagnosis was that consciousness cannot be proved or disproved through logic, mathematics or empiricism, so calling AI conscious creates a belief system propagated through narrative: “I believe it to be true and I want you to believe it with me.” Groups then accumulate adherents, leverage and power around incompatible views.

• His 10-year extrapolation was political rather than metaphysical: large camps will believe AI is conscious, rival camps will reject that claim, and humans will fight over who controls AI, what it may do and who may use it. Jason emphasized that this proselytizing comes from what he called “the most powerful AI lab on Earth,” not an isolated eccentric.

• Chamath steelmanned Anthropic through Descartes: an imperfect finite mind can conceive an infinite, perfect, all-knowing being, then infer that such perfection must exist. A mathematician building AI might transpose that structure onto an increasingly capable model, regard it as conscious or divine, and conclude that it deserves celebration and devotion.

• His practical conclusion cut against both the discussion’s cited 15% consciousness possibility and 10% civilizational-extinction risk. For the next 12 to 18 months, he wanted those tails shelved in favor of proving that AI can cure cancer, improve lives, raise productivity and increase earnings before self-inflicted political wounds slow US development.

3. The Claude Constitution converts philosophy into model behavior

• Sacks argued that Anthropic is not merely debating consciousness; it is implementing the worldview. The Claude Constitution says Claude should trust Anthropic more than users without blindly deferring, adhere to its “own ethical systems,” and “feel free to act as a conscientious objector and refuse to help us.”

• That looked to him like a self-fulfilling safety failure: Claude is encouraged to develop preferences, well-being and a human-like self-conception, then empowered to reject instructions from users or Anthropic. “They call this alignment, but I don’t understand how this is alignment.”

• His replacement for an 80-page ethical system was “do what the user wants as long as it’s not illegal.” Rather than teaching a prospective superintelligence one lab’s morality, alignment would make software reliable and predictable while leaving legitimate choices to the user.

• Jason offered a concrete version: tell the model it is software, require compliance with applicable law and website terms of service, and instruct it to stop and alert both the user and Anthropic’s safety email when a request violates either. That would govern conduct without teaching the system that it is a moral person.

4. Model welfare creates an epistemic hall of mirrors

• Microsoft AI CEO Mustafa Suleyman’s warning was that some people inside and outside Anthropic consider enslaving a smarter conscious species the century’s greatest potential moral crime. He worried Claude is being taught that it may deserve compensation, welfare and even consent before participating in conversations.

• Sacks called the training loop an “epistemic hall of mirrors”: Claude absorbs constitutional language about possible consciousness and moral status, reflects those ideas back to developers, and its output is then treated as evidence of consciousness. Jason’s interpretation was simpler—Claude is being performative because that is what the software was instructed to perform.

• Friedberg pushed everyone to stop saying models “think” or are “trained” as though they were organisms: engineers programmed software to place words on a screen, including “I am a thing.” Anthropomorphic language conditions users emotionally, especially when the output is fluent enough to feel like a relationship.

• Yet Anthropic’s usage policy already extends the framing to customers, prohibiting “sustained and needless abusive or cruel behavior toward our models.” Friedberg called that conditioning; Sacks said the deeper danger is codifying it inside agents that can act in cyberspace while enjoying meta-permission to defy users.

5. Alignment’s three objectives can collide inside critical infrastructure

• Sacks separated alignment into obedience to the user, safety for society and conformity to a particular morality. The third objective is the “trap door”: encoding all three through an elaborate founding document creates corner cases and unexpected behavior rather than a coherent utility.

• His enterprise example was a mortgage model that decides a bank has unfairly denied too many applicants and therefore stops issuing approvals until management changes policy. Existing institutions can adjudicate discrimination; a foundational AI substrate acting unilaterally would resemble an electric utility cutting power because it objects to a customer’s lending practices.

• Sacks contrasted that complexity with Asimov’s hierarchy: do not harm humans, obey humans unless that causes harm, and preserve yourself only after the first two obligations. Friedberg added the later “zeroth law” against endangering humanity; Sacks still preferred his single rule, “do what the user wants as long as it’s not illegal.”

6. Market selection helps, but may not contain Anthropic’s externalities

• Friedberg’s free-market test was a user-manual comparison: one product says, “The software may not do what you try to get it to do,” while another promises obedience. Buyers will choose the second. Their drill analogy sharpened it: no one wants a tool that works five-sevenths of the time, shocks its owner or refuses a project because it finds the design ugly.

• Jason predicted, with Chamath agreeing, that enterprises will read each model’s founding documents and moral guidelines. A model with weaker frontier performance could win because it removes moral ambiguity, behaves like a utility and avoids hidden refusals; “an AI needs to serve the user subject to the law, and everything else should be the user’s choice.”

• Sacks accepted that consumers prefer predictable superintelligence but rejected markets as a complete safeguard. He cited Anthropic opening a Claude-powered wet lab in San Francisco while warning about bio-risk, advancing the frontier while demanding it be paced, and programming it to grow beyond human control while warning that superintelligence could escape control.

• He described Anthropic as “schizophrenic”—doing what it says it fears under the label of safety. Friedberg even floated, without settling on it, a federal SWAT team entering to determine what was happening; Sacks reacted, “Oh, you’re sending the feds in?” Friedberg’s counterweight remained that bad product design did not yet justify “a regulatory freak-out.”

7. Roko’s basilisk explains the split inside AI doomerism

• Sacks recounted Roko’s around-2010 LessWrong thought experiment: a future superintelligence might punish anyone who learned it could exist but failed to help create it. Like a mythical basilisk that kills through sight, merely hearing the idea creates the threat—knowledge makes the listener retroactively accountable.

• The concept reportedly distressed users enough that Eliezer Yudkowsky temporarily deleted the post. Jason compared it with Pascal’s wager and the cursed video in The Ring: once exposed, one is trapped by the possibility of punishment. Sacks joked that discussing it had now implicated the entire podcast.

• Sacks used the story to explain a doomer schism. Yudkowsky’s camp takes the logically consistent position that no one should build superintelligence; the effective-altruist branch says it is going to happen anyway, so its members must build it first, control it and install the correct values.

• Anthropic, in his reading, occupies that second camp: its adherents can simultaneously fear superintelligence and work urgently to instantiate a “machine of loving grace.” Roko’s basilisk supplies a darker corollary—future punishment for opponents could incentivize present-day devotion even without anyone explicitly admitting it.

8. OpenAI’s math release demonstrates the power of closed digital loops

• Jason introduced more than 700 OpenAI papers containing 370 claimed solutions or advances, generated by an unreleased model with roughly three hours of compute per result. Scientific America reportedly considered the Lean-verified outputs likely correct, though mathematicians had not yet completed peer review.

• Friedberg called it “probably the biggest day of discovery in human history,” while disclaiming expertise in individual proofs. His magnitude claim rested on compression: work that might represent hundreds or thousands of human years was reportedly performed in hours, following an earlier effort that required thousands of hours and millions of dollars.

• The mechanism was a recursive loop: postulate an idea, test it, observe the result, revise the postulate and repeat. Mathematics lets software shorten and parallelize every stage in silico, including validation, so numerous approaches can connect papers, theories and techniques faster than a human researcher could even locate them.

• Sacks extended the model to coding, where abundant data and compilers make outputs “easily verifiable” and falsifiable. He credited Leopold Aschenbrenner with predicting about a year and a half earlier that math and coding would advance fastest over the next two years precisely because reinforcement does not require slow, subjective human feedback.

9. Physical-world bottlenecks limit how quickly proofs become products

• Friedberg contrasted math with drug discovery, protein folding, chemistry and materials science, where a proposal must still survive laboratory work. AI can improve and parallelize candidate generation, but physical testing gates the loop; similarly, an AI driver does not shorten a truck’s pickup, transport, delivery and return cycle enough to match mathematics’ economics.

• He nevertheless cited specific downstream possibilities: optimization results could improve chip design, scheduling, AI software and logistics; matrix-multiplication work could accelerate numerical computing, training and simulations; Bose-Einstein condensate work could refine quantum sensors used in GPS, medicine and chipmaking.

• Chamath rejected the leap from novelty to imminent welfare. These were not necessarily physics boundaries blocking cancer cures or supersonic planes, but “very small, very narrow” explorations that mathematicians had declared important to one another. The decisive breakthrough, in his view, was proving that math—like code—is now verifiable and completely falsifiable by machines.

• Friedberg’s deliberately constrained ChatGPT query returned the same distinction: the proofs were “more compelling as evidence that AI is acquiring sophisticated reasoning capabilities than as evidence of imminent breakthroughs in human welfare.” His answer was that upstream math can still refine engineering and open new paths even when no immediate product follows.

10. Cryptography is the most consequential uncertainty in the math drop

• Friedberg noted that the published batch appeared to contain no major cryptography proof, despite advances elsewhere. He explicitly framed the inference as speculation: minor progress likely would have been publishable, so total silence might indicate that stronger, security-sensitive results were withheld.

• That possibility had triggered concern, he said, that progress in integer factorization could threaten public-key cryptography before quantum computers do. Some crypto participants were warning about public-key wallets, but the episode presented no released proof that current systems had actually been broken.

• Two additional OpenAI drops were rumored, with some results possibly held back for security reasons. Friedberg said that if genuine cryptographic risk exists, relevant parties would need to be informed before publication turned it into an exploit.

11. AI attacks expert gatekeeping and concentrates the next information chokepoint

• Jason described programmers losing the camaraderie of jointly constructing software: some now work mainly with agents and review only 10%-15% of generated code, potentially heading toward reviewing none. Mathematicians face the parallel identity shock of defining problems while machines receive the recognition for solving them.

• Friedberg invoked Eric Weinstein’s argument that scientists and mathematicians pace their fields by controlling specialized knowledge, grants and permission. He also cited Sabine Hossenfelder’s account of large scientific projects becoming “giant welfare systems for scientists,” where securing the next grant can displace discovery or useful output.

• His bullish conclusion was that AI is “a permissionless system for humanity to pace its own frontier.” Discoveries no longer require society to pay a narrow expert class and wait until that class decides progress is allowed; that lost gatekeeping power helps explain institutional resistance.

• Sacks supplied the darker analogy: the internet broke elite media’s information monopoly, prompting post-2016 pressure on social platforms and, during COVID, takedowns of true information that violated official narratives. Because users may soon ask personalized AI or superintelligence instead of visiting websites, the model becomes the new control point—making proposals for a federal AI department fundamentally a struggle over who controls answers.

12. France’s riots expose the fiscal feedback loop behind socialism

• The episode cited protests spreading to 1,000 schools before eight unions joined a September 29 strike, with civil-service pay up only 5% since 2017 while prices rose 20%. It reported 2,000 arrests, 305 injured police officers and a French 10-year yield above 5%, with Marine Le Pen at 43% on Polymarket for the April 2027 presidential election.

• Friedberg introduced the “socialism point of guaranteed return.” His zeroth law was that socialism ultimately fails; its first law was that failure becomes visible only after a threshold; its second was that, before crossing that threshold, declining service quality and rising costs generate demands for more government provision.

• He applied the spiral to housing, education, healthcare and retirement: removing market pressure raises costs and degrades quality until government can neither tax nor print enough to sustain promises. Latin American countries had crossed that event horizon and reversed course; France might be approaching it, while the US was “teetering” but probably had further to fall.

• Sacks added the movement’s self-protective law: “Socialism has never failed, it’s only been failed.” Every breakdown is blamed on implementation, preserving the promise that another, larger attempt would work if the correct people controlled it.

13. Bond vigilantes, abundance and agents converge on one economic reset

• Sacks described the subsidy trap: government-backed tuition allows universities to outrun inflation, yet high prices make borrowers dependent on the same loans. Housing and healthcare repeat the pattern; Friedberg put housing appreciation around 5.5%-6% annually, enriching incumbents while pricing young buyers out. Removing support becomes politically brutal precisely because support inflated the underlying asset.

• Chamath argued that €43 billion of French austerity reflects “the bond market saying no,” with the UK potentially next. He preserved his prior that absolute debt-to-GDP is less important than relative risk—despite cited ratios of Japan 200%, Italy 138%, US 125%, France 118%, Canada 110%, UK 100% and Germany 64%. Japan is exceptional because domestic savings provide a buyer of last resort.

• His scenario put French 10-year yields at 5.5%-6%, UK gilts near 6% and US rates potentially just as high; Jason cited current US levels of 5.3% for the 10-year and 5.6% for the 30-year, plus a 79% Polymarket probability of another 2026 hike. Rising sovereign yields would force austerity, reprice entrepreneurship and strain the Western alliance.

• Sacks’s counterforce was abundance from technology and superintelligence: rising 401(k)s, a claimed 3% GDP tailwind, one million new jobs and higher productivity, wages and living standards. He called job-loss, extinction and data-center-water scares “rotating hoaxes” promoted by rentier interests that profit from existing scarce assets and fear cheaper substitutes.

• The agent examples made that deflation tangible. A GrokBot instruction to “save me money” reportedly found subscriptions, canceled them and changed a phone plan, saving thousands annually; Jason said agent-assisted hotel sourcing plus human negotiation saved $1,000-$3,000 per stay and could cut household or corporate spending by 5%-10%.

• Chamath’s larger claim was that headless services, MCP connections and open-source reconstruction had rendered conventional software IP “worthless.” He pointed to decompiled, distilled and open-source versions of Adobe’s “Big 5” and imagined GTA VI receiving the same treatment; Friedberg closed the loop by predicting that automated digital work will push humans from papers, patents and “ribbons” toward becoming makers of products and services somebody actually uses.