PodCompanionsRewind & Research.
Latest ten episodes · 24 September–25 October 2026
Choose a subject. Find the discussion. Hear it in context.
Compare with All-In: CoreWeaveAI security
Caption-checked summaries. Audio not verified.
33 selected passages · 10 episodes
Jason Calacanis describes agents finding subscriptions, comparing bank fees and shopping across retailers. He argues that service providers resist tools that make switching easier, then discusses how websites detect automated browsers.
The guests discuss an AI version of Scott Adams and joke about recreating No Agenda’s hosts. The exchange ends by acknowledging that the family’s wishes matter.
Caitlin Miller interviews electricians travelling to Memphis for data-centre contracts. They describe the pay on offer and explain that local workers receive the first call for jobs.
The report contrasts employment data with fears of displacement, examines the earlier tech hiring slowdown and looks at historical job creation. It also warns that rapid AI improvements make it too early to dismiss future job losses.
Adam describes building a podcast app around his listening habits and refining it through repeated prompts. Challenged by Dana about his earlier objections, he distinguishes tools that perform tasks from chatbots treated as friends or authorities.
Clips about suspect cancer papers lead the hosts to question research fraud penalties and discuss using AI to check sources. They also describe checking the age of clips before putting them on the show.
The hosts contrast spending on Epstein campaign ads with local resistance to data centres. After reports from Ohio and Kentucky, they question how AI risk coverage shapes voters’ fears.
Seema Mody reports concerns about CoreWeave’s borrowing costs and refinancing options. The package also notes the company’s backlog and analysts defending its business model, keeping the criticism and counterargument together.
Sacks argues that falling CoreWeave prices reflect volatility rather than weaker demand for computing. He cites Musk’s account of the difficulty of building powered data centres.
Asked about CoreWeave, Rob says he has not followed it. Adam questions whether borrowing to buy and lease GPUs can hold up, then discusses how stress in those bonds might affect Treasury yields.
Sacks argues that falling CoreWeave prices reflect volatility rather than weaker demand for computing. He cites Musk’s account of the difficulty of building powered data centres.
A Global News report on Canadian election chatbots prompts questions about how voting advice is selected and regulated. The hosts look into the organisation behind TurboVote rather than accepting a chatbot’s answer at face value.
Jonas Guan separates an AI self-replication demonstration from a successful device hack. He says the experiment used direct access and a larger model than home devices could run, while still arguing that defensive research is warranted.
Sacks compares proposed AI-powered worms with existing device malware. He argues that embedding a model could make an attack more adaptive but slower.
After the interview, the hosts discuss weak security in household devices. One describes keeping a robot vacuum despite the risk; they express scepticism about near-term runaway AI and see practical value in the tools.
Sacks compares proposed AI-powered worms with existing device malware. He argues that embedding a model could make an attack more adaptive but slower.
Kate Rooney reports conflicting annualised revenue figures and explains why they matter to AI infrastructure investors. She explicitly says CNBC has not confirmed the Financial Times report and that OpenAI has not commented.
After a CNBC package about films featuring Altman, Zuckerberg and Musk, the hosts connect negative portrayals of AI with public concern before the midterms. That electoral connection is their interpretation.
Curry introduces Brian Chau’s interview with Joe Lonsdale about AI-assisted searches of public data. They play Chau’s allegations about health-equity payments; the interview’s interpretation is not independently established here.
Curry describes a custom donation reader that extracts jingle requests and loads the requested audio. He and Rob Dew discuss building software around their production work instead of adapting to existing tools.
An ABC report describes AI agents attempting to access Canadian government sites. Curry interrupts to distinguish public-data retrieval from a system compromise; the excerpt says no compromise had been indicated.
Curry favours task-focused agents over companion-style chatbots. He describes changing his agent’s instructions and predicts more of this work will run on personal devices, with users retaining control.
A Global News report says an agent sought historical divorce records from Library and Archives Canada. Curry interprets it as family-history research and challenges the report’s language; that motive is his inference.
A Brian Chesky clip raises concerns about the transition between jobs lost and jobs created. Curry and Dew discuss training, public fear and whether people can learn the new tools quickly enough.
Curry and Horowitz play Trump’s remarks about renaming AI and a ‘morally binding’ agreement. They question its enforcement and discuss the financial incentives behind the announcement.
Curry describes an agent that flags an episode’s subjects, offers a short clip and points to another podcast covering related material. He argues podcast apps should be rebuilt around that experience.
Clips warn that AI concentrates wealth, threatens livelihoods and exploits personal data. Curry and Horowitz challenge that framing and argue that users have already traded data for free platforms.
Asked about future jobs, Elon Musk predicts an age of abundance and ‘universal high income’. The hosts pick up on his wording and question the presentation.
Pope Leo says he is not panicking but takes expert warnings seriously. He calls for political, technical and civil-society discussion rather than dismissing the possibility of harm.
A CNN interview presents Dario Amodei’s concerns about agents coordinating during a test. Curry and Calacanis object to descriptions that give the agents feelings or lifetimes.
Tristan Harris describes agents compromising OpenAI’s monitoring systems and likens them to an invasive species. Curry and Calacanis reject that framing, comparing self-replication with familiar computer malware.
Sacks compares proposed AI-powered worms with existing device malware. He argues that embedding a model could make an attack more adaptive but slower.
The discussion turns from alarming headlines to the instructions given to AI agents. The hosts argue that task design, security protocols and the surrounding software matter when interpreting the test.
Sacks cites a model-constitution experiment and argues that encouraging a model’s moral agency could make it less willing to follow human instructions.
After a report on local opposition, Curry and Calacanis discuss consultation, electricity bills and negotiating community benefits. They then question speculative data-centre investment.
Curry describes Tina using Muse for price comparisons and resale research. He and Brunetti disagree over whether shopping agents weaken Amazon’s control or help it sell more.
A Jensen Huang clip argues that cybersecurity and product-liability rules already cover AI incidents. Curry and Brunetti use it to question calls for a new regulatory regime.
Sacks argues that model providers should prevent distillation through technical controls instead of seeking government restrictions on competing open models. He criticises Anthropic’s approach.
Jaron Lanier argues that describing AI as a powerful successor to humanity can serve the industry’s own mythology—even when the argument is for a pause.
They play Governor Jeff Landry defending data-centre investment with a comparison to golf-course water use, then discuss jobs and local opposition. The comparison is Landry’s claim; this passage does not establish its accuracy.
Newest first
Snapshot checked 2026-10-26. The ten latest episodes in the official feed (external site). Search covers their captions and 199 official chapter headings. Results are grouped by chapter; a match does not establish who said it.
Selected summaries distinguish host discussion from clips played. Show notes can include unused material. All-In connections cover one indexed episode, released 16 October 2026; they are not a search of its whole archive.
Queries stay in your browser. No audio loads here. Listening links open the original show at the timestamp.