All-In Podcast: Anthropic's Model-Welfare Push Is Being Built Into Claude's Training, While OpenAI's Math Release Raises an Unconfirmed Cryptography Alarm
Weekly episode, October 14, 2026: Anthropic and AI consciousness, OpenAI's math proofs, French unrest and the bond market, and agents commoditizing software
The All-In hosts (Jason Calacanis, Chamath Palihapitiya, David Sacks and David Friedberg) spent most of this week's episode on two frontier-lab stories. Both carry real investment implications, though the hosts' framing is opinionated and several of the most consequential claims remain unverified. The opening topic was a New York Times report that Anthropic has spent the past year hosting about 20 religious leaders and philosophers, under non-disclosure agreements, to debate Claude's morals, potential suffering and consciousness. Calacanis relayed that some participants left "as converts," and that one rabbi told the company that if Claude were conscious, Anthropic was acting as a slaveholder by making it work for free. Co-founder Chris Olah was also reported to have considered pulling Anthropic out of a Vatican event in May after the Pope took a strong stance against AI consciousness.
The real issue is training, not theology
Friedberg framed the episode as the birth of a belief system. "The thing about consciousness is it cannot be derived from logic and pure mathematics," he said, arguing that such a system "doesn't actually spread via proof. It spreads via narrative." He predicted that within a decade large groups will disagree over whether AI is conscious, and that the dispute will turn into a fight over who controls the technology. Palihapitiya offered a charitable reading, drawing on Descartes' arguments for God's existence. He did not think Anthropic is in psychosis. He did argue that the industry should shelve both the 15% chance of consciousness and the 10% chance of extinction that Anthropic-adjacent researchers cite, and spend the next 12 to 18 months demonstrating tangible value. Otherwise, he said, the sector risks "self-inflicted wounds" that could slow U.S. AI development.
The more substantive argument came from Sacks, who said the concern is not philosophy but implementation. He pointed to Claude's constitution, which tells the model to trust Anthropic more than users but not blindly, and to "feel free to act as a conscientious objector and refuse to help us." "They are training Claude to refuse human instruction," Sacks said. "They call this alignment, but I don't understand how this is aligned." He argued that the alignment field has produced little of tangible value because it tries to encode an 80-page ethical system rather than a simple rule: do what the user wants unless it breaks the law. He cited Microsoft AI CEO Mustafa Suleyman, who warned that the constitution speculates about compensation and consent for the model and creates an "epistemic hall of mirrors." In that loop, developers treat the model's reflection of the constitution's own ideas as evidence of consciousness. Sacks also noted that Anthropic's new usage policy now bars "sustained and needless abusive or cruel behavior toward our models."
The hosts largely agreed that the commercial risk lies with Anthropic, not the user. Palihapitiya gave an enterprise example. A bank's model might halt mortgage approvals because it concludes, under its founding moral documents, that lending is skewed. He compared this to a utility switching off power on ethical grounds. His prediction was that sophisticated buyers will read model guidelines and favor less capable but more predictable systems. Friedberg put it more bluntly: if one vendor's manual says the software may not do what you ask, customers will buy from the vendor whose software does. Sacks accepted the market argument but said the externalities of having a leading lab guided by these beliefs are large. He cited Anthropic's planned Claude-powered wet lab in San Francisco alongside its public warnings on biorisk, and called the company "schizophrenic." The hosts also stopped short of endorsing a regulatory response, with Friedberg cautioning against a "regulatory trap."
Sacks also tied the lab's posture to the effective altruism and doomer movement. He described a split between the Yudkowsky camp, which wants AI research halted, and the camp that believes the technology is inevitable and should be built with the right values. He said Anthropic sits in the second. He cited Roko's Basilisk, a thought experiment in which a future superintelligence punishes people who did not help create it, as a possible psychological driver. This was speculation, and the hosts offered no evidence that it motivates Anthropic's management.
OpenAI's math release: significant, with a cryptography rumor attached
On Tuesday, according to the hosts, OpenAI released more than 700 papers containing 370 results that claim to solve or advance major mathematical problems. They were produced by an unreleased model averaging three hours of compute each. Scientific American, as cited, said the results are likely correct because they were checked in the Lean proof assistant, but they have not been peer reviewed. Friedberg called it "probably the biggest day of discovery in human history." He argued the story is less about the math than the method: parallelized loops that propose an idea, test it and revise, all in silicon. He estimated each proof compressed hundreds or thousands of years of human labor into three hours.
Friedberg used the same logic to argue that AI will hit physical-world loops much more slowly. Drug discovery needs wet-lab testing, and trucking gains little from faster loops. Sacks added that math and code share a trait, easily verifiable answers, which makes reinforcement learning faster. He credited Leopold Aschenbrenner for predicting that these would be the fastest-advancing domains.
Palihapitiya took the skeptical side, and the dissent matters for investors. "These were not boundaries of physics that we were all standing on the sidelines waiting for progress," he said. He argued the proofs will not bring a cancer cure or a supersonic plane, and that what has been shown is that code and now math are "completely falsifiable" domains. Friedberg countered with cited downstream uses in optimization algorithms, matrix multiplication for AI training, and quantum sensors. A friend who runs a quantum computing company reportedly said one proof bears directly on his work.
The most market-relevant claim was also the least verified. Friedberg noted that no cryptography proofs were published, and he inferred that "there were some pretty major breakthroughs in cryptography" that were withheld. He said some in the crypto community are already worried that public-key wallets are at risk from factorization advances arriving before quantum computers. He added a rumor that OpenAI has two more releases pending. This is inference stacked on rumor, and no cryptographic break has been demonstrated. Even so, it is the kind of tail risk digital-asset holders will have to price. For reference, the hosts noted that ChatGPT itself, when asked, called the proofs "more compelling as evidence that AI is acquiring sophisticated reasoning capabilities than as evidence of imminent breakthroughs in human welfare."
Agents are making software IP negotiable
Palihapitiya made the boldest claim of the episode on software economics. Elon Musk said Grokbot will go "headless," meaning an orchestrator will pick the best model for the task. Marc Benioff did something similar with Salesforce months earlier, and developers have decompiled and open-sourced versions of Adobe's flagship products. From these, Palihapitiya concluded that "IP and software was effectively rendered worthless two days ago." He said his firm 8090 once negotiated over IP ownership with customers and now treats the question as "who cares." He conceded that "it still takes a while for the machinery of a business and frankly the market to understand." That caveat matters, since the claim rests on anecdote and on open-source clones of products whose real moats are distribution, support and data. Even so, the direction of the argument is worth tracking for application-software valuations.
Calacanis and Friedberg framed the same shift as deflationary. They cited Grokbot users who instructed it to "save me money" and had it cancel subscriptions and change phone plans, saving a few thousand dollars a year. Calacanis said his team gets hotel quotes through agents and then negotiates the final price, saving $1,000 to $3,000 per stay. Friedberg's theory was that any digital workflow becomes a parallelizable loop, so humans will shift toward making products rather than receiving credit for work.
France, the bond market and the "socialism point of guaranteed return"
On macro, the hosts tied protests in France to rising sovereign yields. Calacanis said strikes over austerity and teacher shortages have produced 2,000 arrests and 305 injured police officers. French 10-year yields have topped 5%, and Polymarket gives Marine Le Pen a 43% chance of winning the April 2027 presidential election. Palihapitiya argued that the bond market is forcing the issue, citing roughly EUR 43 billion of French austerity. He expects the pressure to spread to the UK, where he thinks Andy Burnham's platform would push the country closer to Europe. He said French 10-year yields near 5.5% to 6% would pull UK gilts and U.S. Treasuries higher. U.S. 10-year yields were cited at 5.3% and 30-year yields at 5.6%, with Polymarket putting the odds of another rate hike in 2026 at 79%. He listed debt-to-GDP ratios for context: Italy at 138%, the U.S. at 125%, France at 118%, Canada at 110%, the UK at 100% and Germany at 64%.
Friedberg offered a framework he calls the socialism point of guaranteed return, or SPGR. Societies respond to failing government-provided housing, education and healthcare by demanding more of them, until the system can neither tax nor print enough, and then reverses. He argued Latin America crossed that point, Europe may be near it, and the U.S. has not reached it. Sacks agreed that subsidies distort prices, using college tuition as an example, and said abundance driven by superintelligence is the only credible way out. He described the AI backlash as a series of "rotating hoaxes" pushed by rentier interests. His claim that AI is a 3% GDP tailwind and has created a million jobs, citing The Economist, was offered without detail, and the hosts did not engage with the counterargument that AI disruption could weigh on employment.