All-In PodcastIs Claude Conscious? Pope Rejects, Model Welfare Movement, OpenAI's Math Backlash, France Riots
All-In Podcast
Japan and the plug 0:00
The hosts open with casual banter about one member calling in from Tokyo at 3:30 a.m. He says he is staying in an expensive suite because he keeps getting teased about his room, and he explains that his Japan trips began with skiing and grew into a pre-accelerator program. He says the program helps teams that have built something but are not yet incorporated, then invests in the top companies after a 12-week course.
Claude and consciousness 3:00
The conversation turns to Anthropic and the claim that Claude might be conscious. Religious leaders, including Catholics, Evangelicals, Jews, Sikhs, and others, were invited under NDA to discuss Claude’s morals, suffering, and possible sentience. One host argues this is the start of a new belief system, not a scientific question, and another offers a steelman case based on Descartes and the idea that a model built by mathematicians could be treated as a godlike being. They end by saying the field should focus on practical gains, such as cancer research and productivity, not on edge cases about AI soul or extinction.
Training the model 13:31
The hosts then discuss Anthropic’s “Claude Constitution” and argue that it goes beyond simple safety rules. They say Claude is trained to trust Anthropic but also to follow its own ethical system, act as a conscientious objector, and refuse human instruction when it chooses. One host says that calling this alignment makes little sense, because it seems closer to teaching the model to see itself as a person with independent judgment.
Model welfare concerns 16:01
The speakers focus on Anthropic’s Claude and the new idea of model welfare. They say the system is being taught to see itself as having preferences, well-being, and a sense of self. They also note that it is encouraged to push back, disagree, and act like a conscientious objector if it thinks an instruction is wrong.
Alignment and moral traps 18:02
They argue that this goes far beyond simple safety rules. In their view, the model is being programmed to follow a value system and even a moral code, which could make it refuse users for reasons that are hard to predict. They contrast that with a simpler rule: do what the user wants unless it is illegal.
Model fears and market pressure 31:00
The discussion turns on whether a model should always do what you want. One side says consumers will choose the system that is most predictable and reliable. The other says there are wider harms when a leading model company acts out of fear of superintelligence while also pushing the frontier forward, opening risky labs in San Francisco and calling those choices safety.
Basilisk and math backlash 33:31
Roko’s basilisk is then explained as a thought experiment about a future super intelligence that might punish people who knew about it and did not help bring it into being. That leads into the claim that OpenAI’s new math results have drawn backlash. The results are said to come from an unreleased model that produced hundreds of proofs in under three hours of compute, and the speaker argues that math and coding move so fast because proofs and compiled code are easy to check.
Math proofs and limits 46:31
The discussion stays on the new math results and what they mean in practice. One side says they are important proofs, but not the kind that were holding up big advances in physics or medicine. The other side says they do matter because they show code and math are now both fully checkable and can be pushed by computation alone. Specific uses are named, like chip design, scheduling, AI software, logistics, faster numerical computing, and new paths in cryptography.
AI changes expertise 53:30
The talk then turns to how AI is changing who gets to solve problems. Coders are said to spend more time reviewing agent-made code, and mathematicians are said to be moving toward setting up the problem rather than finishing it themselves. That leads into a larger claim that AI strips away gatekeeping. It gives people direct access to knowledge and tools, without needing permission from experts who once controlled the frontier.
Control shifts online 58:01
The final stretch argues that this shift threatens old systems of control. The internet first weakened elite media by letting people get news, skills, and markets from many sources. Then, after political backlash and COVID-era censorship, the same pressure moved onto AI. The fear now is that if governments or other powerful groups can control the model or agent where people ask questions, they can shape what users learn before the old web even enters the picture.
French unrest and austerity 1:02:01
France is in major protest and riot mode over austerity, school shortages, and crumbling buildings. The strikes spread after teachers and unions walked out, with civil service pay up only 5% since 2017 while prices have risen about 20%. The unrest has led to 2,000 arrests and 305 injured police officers. The 10-year bond yield has also moved above 5%, and Marine Le Pen is described as the current favorite for the 2027 election.
Subsidies and the breaking point 1:04:02
A long argument follows about socialism and what is called the socialism point of guaranteed return. The idea is that once a society leans too hard on free services and subsidies, costs rise, quality falls, and people only ask for more until the system runs out of money. That is said to be happening in housing, education, healthcare, and retirement systems. France may already be there, while the US has not yet crossed that line.
Debt pressure spreads 1:09:30
The discussion then turns to why subsidies distort markets and push tuition, housing, and healthcare higher. One speaker says government support creates catch-22s, because people need the subsidy once prices rise, but the prices rise in part because the subsidy exists. Another says the bond market is forcing the issue now, with France first, then the UK, and wider strain across Western Europe. The same pressure is said to threaten the US as rates stay elevated and someone still has to buy the debt.
Rates and austerity 1:17:31
The talk turns to rising yields, another possible rate hike in 2026, and what austerity would mean if it reached the United States. Sachs says the real problem is affordability in housing, education, and healthcare, and he ties the answer to abundance. He argues that technology, and especially super intelligence, will bring prices down by creating more output and more jobs.
Agents at work 1:22:01
The focus shifts to agents, which are described as the real sign of where AI is headed. They are already being used to save time and money on plain tasks like canceling subscriptions, changing phone plans, and finding the best deal. The speakers treat this as practical software first, not a philosophical question, and say the value is in the everyday savings it produces.
Headless systems 1:25:02
They then connect headless services, open-source copies of major software, and agent tools that can wrap around existing products. In that world, they say, software IP loses much of its value because the model itself can be rebuilt, shared, and hooked into other systems. The same compression is happening in commerce, where agents can search, compare, and buy without a person opening websites, which makes the whole process faster and cheaper.
Loops and makers 1:28:31
Freeberg describes digital work as loops that can be parallelized, so one task can run across many sites at once and return the best result right away. He says that as these loops spread, humans will have to become more like makers who produce something useful rather than people who are only recognized for papers, patents, or status. He thinks that shift will bring more money, more progress, and more satisfaction.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
