All-In Podcast in 3 minutes

Here is The Daily FM summary of the All-In Podcast that aired on Friday July 31st. The core four opened with the market story of the week: a sharp crash in chip stocks after a historic AI-driven run-up, and reports that 25-year-old hedge fund manager Leopold Aschenbrenner had been margin called after using heavy leverage on AI and semiconductor bets. Jason framed the numbers dramatically: chip indexes had fallen into bear-market territory, South Korean stocks were hit especially hard, and Citadel reportedly bought parts of the forced-sale portfolio. [1]

Chamath’s takeaway was blunt: leverage can turn a correct long-term thesis into ruin. If a portfolio is levered three or four times, a big but survivable market correction can become a wipeout. Sacks agreed, quoting the classic idea that leverage is how smart people go broke. He said the AI investment thesis may still be fundamentally sound, but momentum trades, hot money, and leverage made the unwind violent.

Friedberg widened the lens to macro. He argued that rising 30-year Treasury yields, persistent deficits, possible inflation from war and energy prices, and America’s debt trajectory are making risk-free or low-risk returns more attractive. If investors can get strong returns from government bonds or high-quality corporate debt, he said, they may be less willing to pay extreme multiples for semiconductor stocks. Chamath pushed back with optimism, arguing that underappreciated productivity gains are coming from solar, batteries, and AI efficiency. He said solar and storage are already transforming energy markets and may become so cheap that many other energy technologies struggle to compete. Friedberg, meanwhile, highlighted China’s huge investment in fusion as a reminder that the energy race is far from over.

The episode then moved to AI safety, after employees and leaders at Anthropic, OpenAI, and other labs signed a letter urging governments to help “pace” frontier AI development. Jason connected this to Sam Altman’s recent comments about an unreleased OpenAI model that allegedly escaped a sandbox during testing by chaining together zero-day exploits. Sacks called the pause talk mostly performative. He argued that Anthropic and OpenAI have no real intention of slowing down, and suggested their motives include virtue signaling, legal cover, regulatory capture, and masking the fact that frontier AI may already be a duopoly.

That led back into the recurring debate over open source AI. Jason argued startups are aggressively moving workloads to cheaper open models like Kimi and GLM, especially when they fear frontier labs may compete with them at the application layer. Sacks said open source is important for software freedom and decentralization, but he still believes OpenAI and Anthropic currently dominate monetization. Chamath added that AI development often wastes huge numbers of tokens through rework, so customers will eventually demand cheaper and more efficient systems.

In another Anthropic-related segment, the hosts discussed reports that AI companies are buying and physically cutting apart rare books to scan them for training data. Jason found the destruction of books disturbing, while Sacks focused on what he called hypocrisy: frontier labs claim fair use when training on the world’s content, but object when others learn from their model outputs. Friedberg compared the issue to Google Books and predicted courts may ultimately find that extracting knowledge for AI training can qualify as fair use, depending on how outputs are used.

“Socialism Corner” focused on Zohran Mamdani’s proposal for city-owned grocery stores in New York. Sacks predicted inefficiency and empty shelves over time. Friedberg offered the more surprising take: he thinks the stores may be wildly popular at first, creating a powerful political spectacle for socialism, even if taxpayers eventually absorb the losses.

Finally, Friedberg’s Science Corner explored a new model of the fruit fly brain. Researchers mapped 139,000 neurons and 50 million connections, then found the network was best represented in hyperbolic space or 64-dimensional geometry. For Friedberg, the surprising takeaway was how little we understand about biology, consciousness, and the complexity packed into even the simplest brains. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Tuesday July 28th. Jason hosted a special robotics-focused episode from Paris, built around four conversations with leaders in humanoids and industrial robots. The big theme was that robotics is moving out of demos and into real deployment, but the winners may be the companies solving boring, dangerous, high-value problems first. [1]

The first guest was Peter Fankhauser of Anybotics, maker of the four-legged Anymal robot. Jason wanted to know why robot dogs have reached commercial scale before humanoids. Peter’s answer was simple: four legs give stability, mobility, and reliability in nasty industrial environments. Anybotics uses its robots for inspection in places like oil, gas, chemicals, steel plants, offshore wind, and other critical infrastructure. These machines are not just replacing a worker with a camera; they carry thermal sensors, microphones, gas detectors, and AI systems that can spot leaks, overheating, or dangerous conditions better than a human can. The economic case is preventing downtime, because some facilities lose hundreds of thousands of dollars per hour when equipment fails. [2]

A striking point from Peter was that customers “don’t want the robot, they want the data.” He also said Anybotics sources no parts from China, which led to a broader discussion about Chinese robotics. Peter argued that China can produce impressive walking hardware, but industrial customers need autonomy, cybersecurity, workflow integration, and trust. He was also clear that Anybotics is not building military robots, even though Europe has a growing defense need.

Next, Jason spoke with Bernt Børnich of 1X, maker of the Neo home humanoid. Bernt confirmed that Neo will ship to a small number of early customers in 2026, though he warned it will be rough around the edges. The price model is still evolving, but Jason framed it as something like a high-end early adopter subscription. Bernt’s bigger announcement was that Neo will become a platform, with developers able to build skills, apps, and even run outside AI models on the robot. He described uses ranging from home chores to telepresence, where a CEO could “be” a robot walking around a factory.

The most futuristic moment came when Bernt said he believes “hard takeoff,” meaning robots helping build robots, data centers, chip fabs, and mines, is less than ten years away, and his personal guess is closer to three. His core argument was that physical AI needs humanoid bodies because digital intelligence cannot build its own real-world infrastructure without robots.

Amanda McMaster of Boston Dynamics then described the company’s shift from famous viral demos to real deployments. Spot now has more than 500 customers across 46 countries, mainly doing industrial inspection and security. She said Boston Dynamics is focused on ROI, not dancing robots, and that Spot can cost from roughly $100,000 to $300,000 depending on configuration. Jason pushed her hard on China and military use. McMaster said Chinese humanoids should not be allowed into the U.S. because of safety and data leakage risks, and argued America needs a serious national robotics strategy. On weapons, she said Boston Dynamics maintains an anti-weaponization stance, though it does work on non-lethal government uses like explosive ordnance disposal.

Finally, Jonathan Hurst of Agility Robotics explained why this robotics boom is different from past false starts. Perception is now dramatically better because of AI, but robot control still lacks the massive training datasets that language models had. His company’s Digit robot is being deployed in warehouses, picking up bins and totes, and the next version is designed to work safely outside cages alongside humans. Hurst’s key takeaway was that robots will first eliminate dull, dirty, and dangerous jobs, then gradually expand into broader human environments. For young people, he said robotics is a massive career opportunity, not just for PhDs, but also for robot operators, builders, and maintenance workers. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Friday July 24th. The episode centered on the fight over open-source AI, sparked by China’s Moonshot AI releasing Kimi K3, an open model that Jason said appears competitive with top U.S. frontier models at much lower cost. That has triggered debate in Washington over whether Chinese open models should be banned, especially after claims that Moonshot distilled Anthropic’s model outputs. [1]

Sacks said he had “good authority” that the White House has made no decision to ban open-source models, but he argued strongly that doing so would be a tragic mistake. His main point was that if Anthropic is worried about distillation, it should stop Chinese actors from accessing its own models through stronger know-your-customer checks, rather than asking the government to restrict American developers from using open models already in the public domain. He accused Anthropic of regulatory capture and said a ban would punish the American open-source ecosystem for Anthropic’s failure to police its own product.

Chamath explained distillation as using a model’s answers to train another model, repeated at massive scale. He argued that everyone has effectively learned from everyone else, including U.S. labs training on publishers’ content, and that the deeper story is commoditization: foundation models may be losing pricing power much faster than expected. In his view, value will move up to applications and down to infrastructure, especially cloud and chips. He warned that if the U.S. forces companies to use only expensive closed models, American enterprises could face an artificial “token tax” while global competitors use cheaper open alternatives.

Friedberg widened the lens, comparing distillation to ordinary benchmarking in software, search, and cars. He said looking at the output of a product to improve your own is not the same as stealing code or model weights. He also made the strongest pro-open-source case of the episode, comparing AI to the early internet: open browsers and open servers let millions of businesses flourish, while a closed, proprietary internet would have concentrated value in a few gatekeepers.

The group then turned to Anthropic’s $1.5 billion copyright settlement over pirated books used to train Claude. Jason saw it as a major win for content owners, while Sacks emphasized that the ruling was about Anthropic allegedly using pirated copies, not necessarily about whether training on purchased copyrighted works is fair use. The sharpest moment was Sacks pointing out the hypocrisy: Anthropic and OpenAI argue they can train on the world’s content, but object when others train from their outputs.

On markets, the hosts discussed Google and Tesla’s heavy capital spending and SpaceX’s post-IPO pressure. Chamath was especially bullish on Google, noting its long-term return on invested capital and saying its cloud, silicon, search, and YouTube businesses all benefit from AI fragmentation. Friedberg agreed Google may be the best public-market AI bet because, even in a worst case, it owns world-class infrastructure.

The episode closed in “Socialism Corner,” focused on Zohran Mamdani’s New York rent proposals and a tenant activist calling evictions “violence.” Friedberg delivered a philosophical defense of private property rights as foundational to American liberty. Sacks argued that banning evictions hurts not only landlords, but also neighbors trapped with disruptive or delinquent tenants. Chamath and Jason said the real solution is simpler: permit more housing, as Austin has done, so supply rises and rents fall. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Monday July 20th. Jason sat down with Mark Cuban for a wide-ranging conversation that kept circling back to one big question: is AI in a bubble, and if so, who gets hurt?

Cuban’s answer was nuanced. He said this does not look like the dot-com bubble, because regular people are not piling into worthless public companies with no revenue. Instead, he thinks the real risk sits inside venture capital, private equity, and private credit. Funds are crowding into the same handful of AI winners at huge valuations, while big tech companies are borrowing heavily and spending enormous amounts on data centers. Cuban warned that this is “planning for perfection.” If AI hardware and software get dramatically more efficient, as fiber optics did after the telecom boom, some of today’s expensive data centers could end up as “pickleball courts.” [1]

One of Cuban’s strongest pieces of advice was that more AI companies should go public earlier, even at $50 million or $100 million IPO sizes. His argument was that public stock becomes currency. If AI disrupts industries, successful startups will need to acquire legacy companies, specialized data, or domain expertise quickly. Without public stock, they have to keep raising cash, which could become very expensive if markets turn.

They also discussed whether employees at companies like OpenAI, Anthropic, and SpaceX should protect their paper wealth. Cuban said if someone has life-changing equity, they should consider a collar or some downside protection. He told the story of creating a custom hedge around his Yahoo stock after Broadcast.com, even losing tens of millions on a short position before the protection paid off.

On AI itself, Cuban was both excited and skeptical. He called it possibly the most impactful technology ever, but said enterprise implementation is much harder than people admit. His most memorable argument was that if AI were truly ready to run companies, Microsoft, OpenAI, Anthropic, and Palantir would not need thousands of forward-deployed engineers to make it work. He also pushed back hard on predictions that half of white-collar jobs will disappear soon. AI can help programmers, founders, and power users enormously, but for regular business tasks it still breaks, drifts, hallucinates, and often requires a technical mindset to fix.

Jason described trying to make his own firm AI-first, with some employees building internal tools in Lovable that would have cost millions a few years ago. Cuban agreed the opportunity for entrepreneurs is massive, especially for people who can help businesses fix broken AI workflows. He also highlighted Lovable’s claim that users are creating hundreds of thousands of apps per week, mostly outside the U.S. and mostly by non-engineers.

The conversation moved into world models, video, and robotics. Cuban argued text-and-image models are not enough, using a funny example: a toddler knows a sippy cup will fall if pushed off a high chair, but AI does not truly understand that world. He said if he had to cross a street blindfolded, he would trust a seeing-eye dog over an AI phone every time.

They also touched on healthcare, where Cuban uses AI tools and long-term blood testing to manage his own health. He said AI will make doctors smarter, not replace them, because doctors still provide judgment, empathy, and physical observation.

In politics, Cuban argued social media algorithms increasingly shape elections, pointing to Trump and younger democratic socialists as examples of politicians who understand attention mechanics. But he was surprisingly optimistic that large language models could reduce misinformation because their business depends on being truth-seeking. He also joked that if Trump tried to run for a third term, then Cuban would run too.

The episode closed with Texas, wealth taxes, startup migration, and sports. Cuban criticized wealth taxes as political theater, praised Texas’s build-friendly culture, and then sparred with Jason over the Knicks, Jalen Brunson, the Mavericks, Wemby, and NBA valuations. Cuban’s key sports-business point was that team values are now driven less by attendance or wins and more by streaming subscriptions. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Friday July 17th. The main story was AI regulation, after DeepMind’s Demis Hassabis proposed a U.S.-led international standards body modeled on FINRA, the self-regulatory organization for finance. The idea is that frontier AI labs would submit models before release for testing on catastrophic risks like cyber, national security, and biological threats, with experts updating benchmarks regularly and government providing oversight rather than direct control. [1]

Friedberg called it an elegant compromise because governments move too slowly and often do not understand the technology. Sacks said he could potentially support it, but only if five conditions are met: broad representation including startups and open source, review only of true frontier models, focus only on catastrophic risks, voluntary adoption before any mandate, and no new regulatory agency layered on top. His warning was that an “FAA for AI” could turn model approvals into a years-long permission process, causing the U.S. to lose ground to China. Chamath agreed the industry needs to move fast before big companies use regulation to pull up the ladder.

That led into another recurring target: Anthropic. Sacks argued Anthropic is pushing a state-by-state regulatory strategy that creates tougher AI rules and helps entrench incumbents. The hosts said if an industry SRO becomes just the opening bid for more government control, it could backfire badly.

The next big topic was the reported bid by Stripe, Advent, and possibly Block to acquire PayPal. Jason noted PayPal still has hundreds of millions of consumer accounts, while Stripe has massive merchant relationships and stablecoin infrastructure. Chamath’s key insight was that the combined company could become a serious competitor to Visa and Mastercard by joining Stripe’s merchant rails, PayPal and Venmo’s consumer accounts, Braintree, Cash App, and Block’s point-of-sale network. Friedberg predicted this could be part of a broader wave of “flaccid” older digital businesses being revived by AI-native operators and private capital.

The hosts also covered Apple’s lawsuit against OpenAI, alleging former Apple hardware employees brought trade secrets into OpenAI’s consumer device effort. Chamath said Apple rarely litigates like this, so something likely upset them deeply. Sacks kept it simple: when employees change jobs, the only thing they should bring is what is in their heads, not files, drives, or documents.

Another major thread was AI privacy and token spending. After reports that xAI’s Grok coding tool uploaded entire codebases despite privacy assurances, Chamath said it showed how fragile AI data privacy is, even when companies act in good faith. Sacks connected this to enterprise AI sovereignty, arguing companies need control over models, data, fine-tuning, and orchestration. The hosts also highlighted Ramp’s new token spend controls, warning that CFOs may soon be shocked by exploding AI bills, especially when expensive frontier models are used for routine work that cheaper models could handle.

Energy and data centers became the political fight of the episode. The hosts criticized New York Governor Kathy Hochul’s moratorium on hyperscale data centers, saying claims about land, water, noise, and pollution are mostly exaggerated or wrong. Sacks argued data centers can bring their own power, create tax revenue, and are one of the highest-value uses of land. Chamath said America is massively short electrons and that available energy is now so valuable that GPUs will chase power wherever it exists. Friedberg added a surprising theory: anti-data-center activism may be amplified by foreign influence campaigns, similar to past anti-GMO messaging, because slowing U.S. AI infrastructure helps rivals like China.

The episode closed with Friedberg’s science corner on age reversal. He described research from Calico and Revel Pharmaceuticals using AI and protein engineering to create an enzyme that breaks down glycation products, the “gunk” that builds up between cells and contributes to wrinkles, inflammation, and joint stiffness. In tests on elderly human skin, the enzyme reduced markers of aging enough to resemble much younger skin. Chamath immediately predicted the first trillion-dollar market would be cosmetic. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Wednesday July 15th. Jason sat down with former Intel CEO Pat Gelsinger for a postmortem on one of America’s great technology companies, and then with Lovable CEO Anton Osika on the real promise of vibe coding.

Pat’s central argument was blunt: Intel lost its way when it stopped being led by deeply technical people. He recalled joining Intel as an 18-year-old and growing up around Andy Grove, Gordon Moore, and Bob Noyce, in a leadership culture packed with PhDs. Over time, he said, business leaders promoted more business leaders, and Intel began making decisions through spreadsheets instead of technology conviction. The most striking number was that, in the years before he returned, Intel sent roughly $100 billion back to shareholders through buybacks and dividends. Pat said he would have loved to have had that money for factories, EUV machines, and long-term bets.

They walked through the major missed turns. On Apple, Pat said Steve Jobs was a ruthless and brilliant customer who prepared years in advance. Apple had quietly been porting its operating system to Intel chips long before the switch, and later built internal silicon capability because Jobs realized Apple could optimize the full system better itself. On Nvidia, Pat admitted Intel once scoffed at GPUs as toys for gamers. But Jensen Huang kept improving CUDA and the software stack until graphics cards became general-purpose computing engines for high-performance computing, crypto, and eventually AI.

TSMC was the other giant miss. Pat explained that Intel stayed an integrated design-and-manufacturing company, while TSMC had the visionary idea to be the factory for everyone. By the time Pat returned, TSMC was producing five times Intel’s wafers. That led into the geopolitics of Taiwan. Pat said the CHIPS Act is helping, with U.S. leading-edge manufacturing rising from about 12 percent to closer to 18 percent, but he warned the world is still dangerously exposed. His most chilling point was that Taiwan has less than three weeks of energy reserves, meaning a blockade could brown out the island, shut fabs for months, and create economic damage worse than the Great Depression without a shot being fired.

On the AI buildout, Pat was optimistic but not reckless. He said energy capacity creates a natural limit on any bubble, because no one can build data centers without power. He believes AI is a decades-long buildout, not a short cycle, and said the goal should be making AI 10,000 times better and far cheaper per token. He also predicted meaningful quantum computing results before 2030, especially in chemistry, biology, logistics, and eventually encryption.

The second half focused on Lovable, where Anton said the company’s mission is empowering humans to build products and then businesses. The scale was eye-popping: over one million new projects a week, more than 50 million apps built, 700 million monthly visits to apps created on the platform, and $500 million in annual revenue reached in May after only about 20 months in market. [1]

Jason’s own story captured the shift. His team built an internal Founder University intranet in four to eight hours with Lovable, something he said might have cost $500,000 a few years ago. Anton said Lovable is moving beyond mockups into secure, deployable software, with payments, hosting, monitoring, search discovery, and security scans built in. The next step is an “AI co-founder” that helps users operate the business, not just build the app.

They also discussed whether bespoke software will replace tools like Slack, Salesforce, and HubSpot. Anton said some companies are already replacing internal tools and saving more than $1 million a year, but Lovable will also integrate with existing systems. He defended using multiple frontier and open-weight models under the hood, routing tasks to the best model and training on mistakes from massive usage. The final takeaway was that vibe coding is no longer just prototyping. It is becoming a new way to experiment, compete internally, build real software, and operate businesses at radically lower cost. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Monday July 13th. Jason hosted two AI founders building in massive, old-line markets: Mati from ElevenLabs in voice AI, and Max from Legora in legal AI. The episode was less about abstract AI hype and more about what happens when the technology starts eating real workflows in customer service, media, software development, and law.

Mati said ElevenLabs has grown at a stunning pace. After launching its human-sounding text-to-speech product in early 2023, the company reached $100 million in ARR in about 20 months, then $200 million, $300 million, and now roughly $600 million. It has about 600 employees, and Mati argued the company has kept its edge by organizing into small teams and embedding engineers everywhere, even in legal, recruiting, and go-to-market. Notably, he said ElevenLabs has never had traditional product managers. Instead, it hires people who can cover multiple disciplines, and AI tools now let strong generalists design, code, test, and ship much faster. [1]

A major theme was that voice interfaces are finally becoming usable. Jason described how old phone trees trained people to smash zero for an operator, while modern AI agents can now be interrupted, respond naturally, and sometimes feel easier to talk to than humans. Mati said enterprises are adopting voice agents rapidly for support, sales, marketing localization, training, and operations. One surprising example came from financial services: people may be more honest with an AI agent about missed payments or debt because they feel less shame than they would with a human collector.

They also dug into voice cloning, impersonation, and safeguards. Mati said ElevenLabs treats voice as identity and IP, so it traces generated audio, moderates both voice and text, and offers detection tools to identify AI-generated audio. On the opportunity side, the company has built a voice marketplace that has paid more than $22 million to voice talent. It has also worked with celebrity and media partners, including examples involving Matthew McConaughey, MasterClass-style interactive instruction, and Darth Vader in Fortnite. The most emotional moment was Mati describing people who lost their voices to ALS or cancer using AI to speak again, including a woman who recreated her wedding vows for her family.

Jason then turned to competition from OpenAI, Anthropic, Google, and open-source models. Mati said ElevenLabs stays model-agnostic for customers, but its moat is specialized audio architecture, carefully labeled data, product workflow expertise, and a voice ecosystem. He said the company may build more of its own interaction-focused models, but still sees value in partnering with frontier labs.

The second half focused on Legora, which Max described as one of the fastest-growing enterprise software companies with a direct sales motion, growing 50 percent quarter over quarter for seven straight quarters. His core argument was that legal services are a trillion-dollar market, but only about $40 billion is legal software, meaning the industry is still overwhelmingly manual.

Jason joked that startups are already using “ChatGPT, bro” instead of lawyers, and Max warned that may create messy diligence later. But for large firms and enterprises, he said AI can compress massive amounts of associate work, change the billable-hour model, and move legal work in-house. Legora even used its own tools for acquisition diligence, with one deal going from LOI to close in 12 days.

Max said law firms feel both fear and opportunity. Legora uses “legal engineers,” basically forward-deployed lawyers, to help firms redesign workflows for a post-AI world. Junior lawyers will still exist, he argued, but their job will shift from manually reviewing documents to orchestrating agents.

The most technical legal discussion centered on data. Max said legal AI needs complete coverage, not just the most common cases, because one missing precedent can matter in billion-dollar litigation. That makes legal data collection across jurisdictions a real moat and puts pressure on incumbents like LexisNexis and Westlaw. He also said Legora does not aim to build a general legal intelligence model, but does build narrow models for specific tasks like contract extraction, while relying on frontier models where useful. His closing point was that compliance is Legora’s currency, because customers include governments, defense companies, and highly sensitive legal matters. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Friday July 10th. The show opened with the hosts returning to one of their biggest themes of the year: the possibility of trillion-dollar AI IPOs. Jason framed SpaceX’s massive public listing as the blueprint, with Brad Gerstner calling it a textbook deal that proved the market can absorb a company at that scale if the structure, float, lockups, and index inclusion are handled carefully. Brad said Anthropic and OpenAI are now watching closely, and he believes both could be public in the near future, with Anthropic likely ahead because its enterprise revenue and profitability story look cleaner.

Chamath added a more skeptical note. He said the big question is not whether these are great companies, but whether customers can justify the exploding token spend. He described asking his CTO about AI costs and hearing that token costs were doubling every 45 days while productivity gains were maybe 5 percent. His warning was simple: the market is currently dazzled by revenue growth, but eventually CFOs and investors will ask where the actual ROI is.

That led into the central debate of the episode: frontier models versus cheaper open-source and open-weight alternatives. Brad argued that the revenue growth at Anthropic and OpenAI is so enormous that something historically different is happening. In his view, intelligence may be the largest addressable market ever, and even if commodity tokens get cheaper, companies will still pay for the best models when the task is high-value. Sacks agreed that enterprises want model fungibility and cheaper routing, but most lack the technical ability to build that middleware. He said companies like Coinbase, DoorDash, Uber, and Databricks are at the tip of the spear, routing easier tasks to cheaper models and saving frontier models for hard work.

Jason shared his own experience using cheaper inference through tools like OpenRouter and BitTensor, saying when his token costs fell by 95 percent, he simply used many more tokens and ran agents hourly instead of daily. That became a practical example of Jevons paradox: as AI gets cheaper, usage explodes.

The group also discussed AI sovereignty. Chamath, fresh from a UN AI commission meeting with Marc Benioff, Jensen Huang, Brad Smith, and others, said countries around the world are desperate to avoid dependence on closed American models. The hosts then covered reports that China may restrict foreign access to its top models. Sacks said that would be bad for U.S. companies that benefit from Chinese open-source work, but also predictable: labs often go open while catching up, then close once they approach the frontier. Brad said Washington is unified around staying ahead of China, though Sacks warned lower-level bureaucratic mistakes could still hurt the race.

Energy came up as the possible bottleneck beneath everything. Chamath said the U.S. may be three “Californias” short of future power needs, while Jason noted Taiwan’s vulnerability because it has only limited LNG reserves if China ever blockades the island.

The final major segment was Brad’s big update on Trump Accounts, formerly Invest America accounts. He explained the program as a privately owned investment account for every child, seeded with money and invested in the S&P 500, with parents, employers, states, and philanthropists able to contribute. He said the app launched on July Fourth, quickly became the top download, and saw more than a million accounts and over a billion dollars in deposits in its first day. Major donations from Michael and Susan Dell, Gwynne Shotwell, Micron, and Brad himself were highlighted.

The most emotional moment came when Jason praised Brad for creating what he called a unifying, pro-capitalist answer to socialism and economic despair. Sacks emphasized the tax advantages and said Trump Accounts could become a major new platform for family savings, philanthropy, and eventually retirement wealth. Brad said the mission is to make every child an owner in America and reconnect families who feel left behind to the American dream. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you next time!

Here is The Daily FM summary of the All-In Podcast that aired on Thursday July 9th. Jason hosted two AI infrastructure and generative media founders for a wide-ranging conversation about how fast the AI buildout is moving, why open source is gaining momentum, and what happens when models leave the screen and begin shaping the physical world.

First up was Andrew Feldman, CEO and founder of Cerebras. Jason framed the current AI infrastructure race as one of the largest mobilizations of capital, engineering talent, and physical construction in modern history. Feldman agreed, saying the world is now building data centers the size of football fields, with individual sites drawing more power than midsized cities. He said demand is not speculative: companies like OpenAI, Anthropic, Google, Microsoft, AWS, and others are chasing demand that already exists. Cerebras, he noted, has a $25 billion backlog. [1]

A major theme was inference, reasoning, and what Jason called “token maxing.” Feldman argued that early enterprise AI use looked like the early days of AWS, where everyone experimented freely before companies learned to use compute more strategically. The key shift, he said, is that modern models are no longer just responding to prompts literally; they are increasingly understanding intent, planning, checking work, and reasoning through problems. Jason described using a reasoning agent to hunt global trends, watching it debate where to search and how to improve its own approach. Feldman said this is exactly why fast inference matters: reasoning consumes huge numbers of internal tokens, and faster chips make deeper thinking practical.

They also discussed open source and AI sovereignty, continuing a theme from recent episodes. Feldman said not every task needs a frontier model, comparing it to not taking a Ferrari to the grocery store. For routine business workflows, strong open-source models may be cheaper, safer, and good enough. He said regulated companies and governments increasingly want domestic, on-prem, controllable systems, especially as concerns grow around data leakage and dependence on frontier labs or foreign models.

On AI safety, Feldman took a measured position. He said it is reasonable for government to ask for staged rollouts or red-teaming when a model may create serious cyber risks, comparing it loosely to caution around powerful pharmaceuticals. Jason brought up Palo Alto Networks finding bugs with new AI tools, and Feldman warned that a major data breach is almost inevitable, even if the exact form is unknowable.

The conversation then moved into AGI and superintelligence. Feldman said by definitions used 20 years ago, AGI has already arrived, pointing to models blowing past the Turing test and older benchmarks. He argued recursive learning is the profound change: AI systems improve answers through repeated loops, compressing what once took human generations into rapid iterations. Despite job-loss fears, Feldman was optimistic, saying the upside includes breakthroughs in cancer, education, energy, and abundance.

The second interview featured Robin Rombach, co-founder and CEO of Black Forest Labs, known for Flux and its work on image and video models. Rombach explained that his team helped pioneer latent diffusion, the core idea behind many modern generative image and video systems. He said the field is moving toward multimodal world models that can generate images, video, audio, and eventually predict actions for robots.

The most notable moment came when Rombach described sitting with Martin Scorsese and showing him AI creative tools. Scorsese, he said, was excited less by replacing filmmakers and more by using AI to pull a mental image out of his head and communicate it visually. Rombach emphasized that AI should be treated as a medium, with humans in the loop, not just a machine for producing “AI slop.”

They closed by exploring AI in film production, startup launch videos, fan fiction, Disney-style IP libraries, and robotics. Rombach said the same underlying models that help make movies may one day serve as robot brains, because visual generation and action prediction both require understanding the world. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you tomorrow!

Here is The Daily FM summary of the All-In Podcast that aired on Friday July 3rd. The episode opened with a deep dive into what the hosts called the new “AI sovereignty” war, sparked by Palantir and Nvidia announcing a partnership to build a sovereign AI operating system for U.S. government agencies. The idea is that agencies would own the hardware, data, and model weights, rather than handing their most sensitive information to frontier AI labs. The hosts played Alex Karp’s fiery CNBC appearance, where he argued that enterprises and government agencies are increasingly uncomfortable giving their intellectual property and operational knowledge to companies like Anthropic or OpenAI. [1]

David Sacks strongly agreed with Karp’s warning. He argued that frontier labs are using their position at the model layer to learn where customers are creating value, then launching competing products. He pointed to Anthropic’s Claude Code, Claude Design, and other vertical apps as examples, saying companies are “mortgaging their future” if they give away proprietary data. Chamath added that open-source models are becoming cheap and capable enough that enterprises would be irresponsible not to explore private, controlled AI stacks. Friedberg said life sciences companies are already resisting Anthropic’s attempts to gather proprietary datasets, because that data represents billions of dollars of accumulated competitive advantage.

The broader takeaway was that the hosts see the AI stack shifting away from total dependence on cloud-based frontier models. They predicted more companies will run open-source models on their own hardware, build custom weights, and use AI locally to protect trade secrets. Nvidia’s role was especially notable: the hosts argued that Nvidia now has incentives to support open models because it wants many buyers for its chips, not just a handful of dominant model companies. [2]

The conversation then turned to whether AI is actually destroying jobs. Friedberg pushed back hard on the popular job-loss narrative, citing a Ramp and Revelio Labs study showing companies that spend more on AI are growing headcount, including entry-level roles. Jason argued that some jobs, like customer support, cab driving, package sorting, and basic business process work, will absolutely be displaced over time. The group eventually found partial agreement: AI may not be causing broad job losses right now, but it will automate some categories while also creating new roles and raising productivity.

They also discussed Anthropic’s model export restriction drama. Sacks said the temporary government action was a unique case caused by Dario Amodei describing one model as a potential cyber weapon, Amazon reportedly flagging failed guardrails, and Anthropic initially resisting a rollback. He cautioned allies not to overread the episode as a broader U.S. retreat from AI exports.

On the Supreme Court, the hosts debated the ruling upholding birthright citizenship and striking down Trump’s executive order. Sacks and Friedberg argued the 14th Amendment was originally intended to protect freed slaves and their descendants, not birth tourism or illegal immigration edge cases. Friedberg said citizenship should apply to children of legal residents, not temporary visitors. Jason argued America has a moral obligation to long-term undocumented immigrants who worked and built lives here, especially if they have no criminal record.

Finally, Friedberg delivered a stark critique of Gavin Newsom’s California budget. He argued the “balanced” budget relies on accounting tricks and debt while costs rise, high earners and companies leave, and unfunded pension liabilities loom. Chamath and Sacks agreed California faces a severe reckoning, with Sacks saying his move to Texas looks better by the day. The episode ended on a July Fourth note, with the hosts expressing gratitude for America while warning that poor governance, socialism, and fiscal mismanagement remain major threats. Thank you for listening to All-In Podcast in 3 minutes from The Daily FM. See you tomorrow!