Podchemy

Podchemy is now open source! I don’t listen to podcasts. I prefer reading, but most notes/summary tools serve a few bullet points and quotes which are too shallow for me. This felt like a good problem to solve when I was looking for ideas to build, and learn how to code with AI in the process. A year later, @podchemy gets thousands of visits every month. Many podcast creators and guests whom I admire have praised these notes. Balaji invited me to his Network School, Sajith Pai called it “likely my favourite new podcast tool / offering of 2025”, and my favorite moment was getting an email from David Deutsch with some corrections to the notes on his podcast appearance. Building Podchemy has been a rewarding experience. I’ve learned a lot and strengthened my AI muscles. There are many directions it could go from here, and I hope this decision to open-source will help with that. I am also taking an indefinite break from Podchemy’s active development and maintenance. Given how rapidly AI tools have been progressing, other higher-impact ideas floating in my head, and life getting more interesting but also demanding at both work and home, Podchemy is no longer on my list of priorities. I may come back to it, or not, I don’t know. Open-sourcing feels like the right closure for now. GitHub link here, fork away! github.com/vatsalkaushik/…

@collision @sundarpichai @eladgil "I think you can paralyze yourself thinking 10 years ahead. But we are fortunate to be in a moment where you can think a year ahead and the curve is so steep. It's exciting to just do that year ahead." Full episode notes here if you prefer reading: podchemy.com/notes/the-hist…

.@sundarpichai joined @eladgil and me in the Cheeky Pint pub. I was excited to get into Google in 2026: how AGI-pilled Google is, compute bottlenecks, fast AI products, $180b capex, the intelligence overhang at enterprises, and deciding capital allocation at a company overflowing with ideas.

Highlights from @balajis recent podcast with @a16z on Why AI Raises the Cost of Verification: > Every tool that makes creation cheaper makes verification more expensive. While AI collapses the cost of generating content, the effort required to confirm what is real rises. > AI transforms the individual into a CEO. Humans act as the sensors providing taste and agency, while the AI serves as the actuator that executes the work. > Distillation allows smaller players to replicate the intelligence of large AI models at a fraction of the cost, making it difficult for big labs to maintain a closed ecosystem. > AI is a shortcut that is only safe for those who understand first principles well enough to debug the results. > The friction of verbal prompting often makes AI slower than manual work, leading some users to reject the technology entirely. > BioAI uses the body's internal telemetry as a non-verbal prompt, allowing machines to detect needs or illnesses before a person is even aware of them. > In adversarial environments like markets, using generic AI models provides no edge because others can easily predict and counter those moves. > Much of the fear surrounding AI is self-manufactured by users who prompt systems to mimic dangerous science fiction characters and then fear the results. > True AI autonomy is limited by the need for a physical supply chain to reproduce, which provides natural frictional breaks against a Skynet scenario. > The digital divide is flipping. Digital products and AI services are becoming cheap commodities, while human interaction is becoming the luxury premium. > A job only truly changes when it reaches 100 percent automation. At 99 percent automation, the human workload often increases because the worker must still supervise and verify the machine. > When AI provides a baseline of high intelligence for everyone, human taste and agency become the most important factors for success. > People undervalue CEOs because management is expensive to test, unlike sports or math where individuals quickly learn their own limitations. > Instead of replacing human workers, newer AI models replace older ones, effectively acting as digital employees that a manager hires based on performance. > AI enables people to become high-level generalists, allowing them to perform competently across many disciplines before needing a specialist for final polish. > Distribution is the primary moat that AI cannot easily replicate. Existing companies with large user bases can often ship AI features to their customers faster than a disruptor can build a new network. > AI companies often model technological disruption while ignoring political singularities like shifts in the reserve currency or internal national instability. > Decentralized AI might eventually outperform corporate models because it is less constrained by copyright laws and political backlash. > Bitcoin provides a superior form of collateral because ownership can be verified instantly and cheaply on-chain, whereas physical assets like gold are increasingly vulnerable to AI-faked audits. > The inherent transparency of public blockchains makes Bitcoin an ideal institutional asset, as institutions are structured to handle the tracking and de-anonymization that AI analytics tools now facilitate. Read full episode notes here: podchemy.com/notes/balaji-o…

"Lots of people knew that those little O-rings were unreliable. But every single time you get away with launching a space shuttle without the O-rings failing, you institutionally feel more confident in what you're doing. We've been using these systems in increasingly unsafe ways. This is going to catch up with us." Episode notes here if you prefer reading: podchemy.com/notes/an-ai-st…

"Using coding agents well is taking every inch of my 25 years of experience as a software engineer." Simon Willison (@simonw) is one of the most prolific independent software engineers and most trusted voices on how AI is changing the craft of building software. He co-created Django, coined the term "prompt injection," and popularized the terms "agentic engineering" and "AI slop." In our in-depth conversation, we discuss: 🔸 Why November 2025 was an inflection point 🔸 The "dark factory" pattern 🔸 Why mid-career engineers (not juniors) are the most at risk right now 🔸 Three agentic engineering patterns he uses daily: red/green TDD, thin templates, hoarding 🔸 Why he writes 95% of his code from his phone while walking the dog 🔸 Why he thinks we're headed for an AI Challenger disaster 🔸 How a pelican riding a bicycle became the unofficial benchmark for AI model quality Listen now 👇 youtu.be/wc8FBhQtdsA

Some highlights from this episode: > Autonomous vehicles are becoming demonstrably safer than human drivers, which will force regulators to rethink the purpose and necessity of a driver's license. > Human driving may eventually become a niche skill similar to horseback riding as autonomous systems prove to be demonstrably safer. > AI safety overlays could allow humans to continue driving for pleasure by acting as digital bumpers that prevent accidents while maintaining the feeling of control. > Marketplaces are often more successful when they are supply-led. If you build liquid supply and have product-market fit, the demand tends to follow naturally. > To grow a marketplace, focus on creating the easiest tools for suppliers to use. High liquidity on the supply side is a key driver for overall platform success. > Automation could significantly reduce the legal burden on the United States, where car accidents currently account for nearly half of all court cases. > Managing autonomous robot fleets will likely be easier than managing humans because machines are more predictable and do not have the option to refuse dispatches. > Uber's coordination model focuses on the efficiency of the entire network rather than just connecting a user to the closest vehicle, using real-time predictions to optimize the fleet. > Universal basic services for housing and food can act as a stability lever to prevent social chaos during major economic transitions. > Automation typically augments work rather than replacing it, shifting human roles toward overseeing and managing technology. > As companies grow larger and more profitable, they should become less conservative and take bigger risks since they have the financial stability to withstand failures. > Defining a core value like doing the right thing without complex descriptions forces every individual to take personal responsibility for their decisions and impact. > Large companies should prioritize new ventures that rhyme with their core strengths, ensuring they have a distinct advantage or a right to win in that market. > Uber defines its core competency as being a platform for flexible work, which allows them to expand into diverse fields like AI data labeling. > Ridesharing services are already making driver's licenses unnecessary for some young people who prefer the convenience of being driven over the responsibility of driving. > Autonomous vehicle providers will likely carry product liability insurance, shifting the responsibility for accidents from human drivers to software manufacturers. > Vertiports represent a significant real estate opportunity and must be designed for mass market volume with multiple landing and takeoff points. > The transition to electric vehicles is limited by physical infrastructure, but the rise of autonomous cars will naturally speed up the shift away from combustion engines. > The shift in the labor market provides a significant opportunity for startups to create solutions for the displaced human workforce. > Autonomous vehicle mass production and high costs mean driver's licenses will remain relevant for the next few years, but they will likely become optional within a decade. Link to full episode notes below!

Knowing a terminal diagnosis early can lead to a more meaningful end of life than finding out at the last minute. People tend to reach the acceptance stage of grief quite fast and often report higher happiness levels once they stop resisting their mortality. Arthur Brooks views this as a way to reduce the "resistance" that multiplies pain into suffering, which is a concept he draws from various religious traditions. Approaching death with courage can even be framed as a final act of service that helps others live more fully. There are also specific patterns in how people use the lifespan of their same-gender parent to benchmark their own timeline for work. Read full episode notes here: podchemy.com/notes/arthur-b…

My very interesting Conversation with Arthur Brooks, conversationswithtyler.com/episodes/arthu…, @arthurbrooks, the meaning of life etc.

The foundational principle at Alpha is that children should actually love being there, ideally more than they love being on vacation. While adults often prioritize building enjoyable office cultures, they frequently expect kids to treat school as a miserable chore. Joe Liemandt (@jliemandt) uses AI to remove the social pressure of being critiqued by an adult, which allows students to condense an entire year of curriculum into roughly 30 hours of focused work. This shift in the learning environment is often supported by real-world incentives, such as teaching financial literacy using actual earned money. Link to full episode notes below!

Some highlights from this episode: > The current school system is coded to reward only two traits: high IQ and high conscientiousness. If a child does not naturally possess these, the system fails to adapt to their needs. > The US education system often prioritizes moving through the calendar over student mastery, which causes foundational knowledge gaps to compound until learning stops. > Students can achieve top 1% academic results by spending just two hours a day on core subjects if the learning environment is optimized for efficiency. > Repositioning an educational product from learning more to learning faster can solve product-market fit issues by appealing to the high value parents place on free time. > Mastery-based AI tutors shift the focus of education from IQ to effort by requiring students to fully master basic concepts before they can advance. > Most academic struggles stem from cumulative holes in knowledge, where a failure to master fifth-grade concepts leads to a total collapse in high school performance. > Effective learning occurs in the zone of proximal development, where students succeed about 80 to 85% of the time to maintain engagement without becoming frustrated. > AI-driven learning can condense a full year of subject material into 20 to 30 hours of focused work, allowing students to learn ten times faster than in traditional classrooms. > Student disengagement often stems from low standards rather than high ones, as clear and difficult expectations provide students with a sense of purpose and direction. > Effective education should mirror fixing an engine: present the problem first to create a genuine need for the tools and knowledge required to solve it. > Traditional teaching roles fail because they require one person to be a domain expert, a pedagogue, a motivator, a parent liaison, and an administrator simultaneously. > Students prefer AI feedback over human instruction because AI is non-judgmental and reduces the social pressure of being critiqued by an adult. > Adolescents often resist academic pressure from parents, making external mentors or guides essential for maintaining high standards without damaging the parent-child bond. > Teaching financial literacy with real earned money is more effective than simulations because students value what they have worked for and learn from actual losses. > Ambiguity is a major hurdle in education. When students know exactly how many hours of work are required to reach a goal, the task becomes manageable. > Leadership is most effective when it is binary. You should be 100% in charge or 100% hands-off to avoid the compromises and stagnation of consensus. > Educational systems should take full responsibility for student outcomes. If a student is not learning, the system must adapt its methods rather than blaming the child. > Child development happens through a cycle of struggle and failure supported by a caring adult. This process builds the self-confidence and resilience that children need to succeed. > Tracking every hour of a week helps students see the gap between their ambitions and their actions. It forces them to realize that they are defined by how they spend their time. > High standards are the primary driver of a child's happiness in school because mastery and achievement create genuine engagement. Full episode notes here: podchemy.com/notes/joe-liem…

My conversation with @jliemandt on why the future of education is better than you think. 0:00 The current education system 7:01 What makes Alpha School different 11:01 What are the results 23:20 Current classroom struggles 26:40 What does mastery mean? 35:37 Changing the education system 39:19 Teaching through AI 44:27 How do you solve motivation? 57:01 What makes a good teacher? 1:01:04 Coaching 1:05:17 What life skills matter? 1:08:18 Doing hard things 1:13:25 AI Monitoring 1:21:08 Effort vs. IQ 1:24:40 What happens after Alpha School? 1:38:21 The Genius of Jack Welch 1:45:49 Trilogy IPO: the choice to not go public 1:51:40 Physical vs. virtual learning 2:03:18 Does Paying Kids To Learn work? 2:11:01 What Is Success For You? (Includes paid partnerships)

Some highlights from this episode: > Developing general robotic foundation models may be easier in the long run than building narrow, specialized systems for specific tasks. > True robotic generalization often looks mundane, such as picking up plates in an unfamiliar kitchen, but it is far more difficult than creating specialized demos in controlled environments. > A foundation model for physical intelligence could trigger a Cambrian explosion in robotics by allowing people to build applications without having to solve the core intelligence problem themselves. > The future of robotic surgery involves moving beyond teleoperation so that machines are no longer limited by the speed or dexterity of a human controller. > The primary challenge in robotics is creating cost-effective systems that can handle rare long-tail scenarios without needing massive new datasets for every task. > Multimodal language models provide a path to giving robots common sense by allowing them to leverage general knowledge to navigate situations they have never physically experienced. > Sophisticated software can overcome basic hardware limitations. For example, simple cameras can function as touch sensors by visually tracking how objects deform. > Robotics is moving from a physical bottleneck to a reasoning bottleneck, where the challenge is no longer how the robot moves, but how it interprets the scene to choose the next step. > Common sense in robotics is the opposite of muscle memory. It is the ability to apply abstract knowledge or facts to a specific physical situation to make a correct decision. > True generality in robotics comes from systems that can improve autonomously through their own experience rather than relying on human engineers or manual data labeling. > Tasks like making espresso or folding laundry serve as difficult challenges to push the limits of general-purpose robots rather than being the end goal itself. > The true test of robotic intelligence is performing mundane human tasks like washing a greasy pan or using a plastic bag, which are paradoxically difficult for machines. > Robots can surpass human speed by using reinforcement learning to identify and remove the mental processing pauses that humans naturally take during complex tasks. > True physical intelligence is agnostic to the body. A single foundation model should be able to control any form factor by treating every machine as part of the same general problem. > The bitter lesson suggests that AI reaches its greatest potential when researchers stop trying to program human logic into the machine and instead allow it to learn entirely from data. > Moravec's paradox shows that tasks humans find most natural, like physical caregiving, are the most difficult for robots because we are highly evolved for physical intelligence. > Robotic hardware costs have plummeted from $400,000 to roughly $3,000 per arm in just one decade. > Robotics faces an activation energy problem where robots must be useful enough to deploy before they can gather the real world data required for large scale improvement. Full episode notes here: podchemy.com/notes/sergey-l…

My conversation with Sergey Levine (@svlevine). Sergey is the co-founder of @physical_int -- a company building foundation models that can control any robot to do any task in any environment. The company's thesis is that generality is more scalable than specialization, meaning that a model trained across many different robots and tasks will ultimately outperform any system built to do one thing well (eg, just wash dishes). Sergey is a researcher by background, but I think you will appreciate how practical and commercially grounded this conversation is. We discuss: - Why changing a diaper will be the last task a robot masters - The simulation v. real-world data debate - How multimodal LLMs give robots common sense - Moravec's Paradox + Robot Olympics - Why robots can do long-horizon tasks now - A realistic timeline for robots in our homes I should note that I am an investor in Physical Intelligence -- I made the investment because I believe it is one of the most important companies tackling the problem of robotics. Enjoy! Timestamps: 0:00 Intro 2:39 Defining Physical Intelligence 5:19 The Challenge of Building General Models 6:34 The Stakes and Future of General Purpose Robotics 8:15 Pros and Cons of Humanoid Robots 10:12 Historical Milestones in Robotics Research 15:31 Combining Generative AI and Deep RL 21:24 Moravec's Paradox 25:33 Kitchen Robots 29:30 Simulation vs. Real-World Data 30:48 The Robot Olympics 36:31 The Physiological Reality of Embodiment 38:56 Controversies in the Robotics Community 44:18 What Makes a Great Researcher 48:27 How Businesses Should Prepare for Robotics 54:09 Tracking Progress Through Research Papers 57:02 The Next Step: Mid-Level Reasoning 1:02:00 The Kindest Thing

Some highlights from this episode: > AI is already increasing leisure time by allowing workers to finish tasks more quickly, often without their employers realizing the increase in efficiency. > Open regulatory processes face the risk of being overwhelmed by high quality but pointless AI generated spam. > The primary risk of AI is its unpredictable impact on governance rather than direct economic collapse. > Automation in industries like trucking may happen slower than expected because human jobs involve complex tasks beyond the primary function, such as managing cargo and logistics. > The greatest political risk from AI may be the displacement of the upper middle class, as influential professionals face significant salary reductions and career shifts. > The true value of AI in education is using it to discover the right questions to ask for different contexts. > A radical new model for grading involves having one AI evaluate a semester-long chat transcript between a student and a different AI tutor to measure learning progress. > The traditional fifteen week semester is an artificial constraint that AI-driven learning can break by allowing students to move at their own pace. > Writing specifically for AI allows individuals to build a digital model of their own thinking that can be used by others in the future. > To manage AI cheating, schools can use occasional proctored sessions to establish a performance baseline for each student. > AI is likely already better than the average human at conducting interviews, though it has yet to surpass the very best human evaluators. > Many people underestimate AI capabilities because they only use free versions, which are significantly less powerful than high-end models. > Focus on messy jobs that require face-to-face interaction and non-routine problem solving to stay valuable in an AI-driven world. > Just as people at the start of the Industrial Revolution could not have predicted the job of a podcaster, we cannot yet see the unique roles that will emerge from the AI transition. > Writing instruction should include assignments that require AI to push for higher quality and assignments that ban AI to develop independent thinking. > Colleges can use AI to offer niche subjects that lack dedicated faculty, allowing students to explore specific interests at zero marginal cost. > AI will likely improve education by optimizing existing tutoring methods through data analysis rather than through specialized ed-tech software. > Professionals can adapt to AI by shifting their focus to activities that require a human presence, such as live events and podcasts. > Colleges should devote a third of their curriculum to AI because nearly every future job will require AI literacy. > AI will likely increase the number of billionaires, but the formation of new companies will create enough projects to prevent mass unemployment. Link to full episode notes below!

@HooverInst @tylercowen @EconTalker @Liberty_Fund @Econlib Full episode notes here if you prefer reading! podchemy.com/notes/ai-emplo…

On a new EconTalk, @TylerCowen makes the case that AI won't break education or work—it'll redefine both, and the smartest students will learn to use it, not fear it. Watch the full conversation with Russ Roberts (@EconTalker) from the @HooverInst, @Liberty_Fund, and @Econlib:

"If you want to preserve your ability to synthesize, it makes sense to exercise caution. There are researchers looking at the negative cognitive impacts of depending on AI. Much like your ability to navigate has probably deteriorated since using Google Maps, you want to keep certain muscles strong and able." Full episode notes here! podchemy.com/notes/859-q-a-…

From my brand-new Q&A: youtube.com/watch?v=Vw1-tG…

Courage is learnable.

Highlights from this episode with @polinapompliano x @jposhaughnessy : > Identity is defined by how we move through the world and what we do, rather than by how we describe ourselves. > To understand someone's true character, look for the small moments when their public mask drops, such as through their level of fatigue or the stories they choose to tell. > Group people by their shared mental models rather than their professions to reveal deeper universal insights. > Viewing the world through a specific lens allows you to find inspiration in unexpected places, such as using music structures to design a meal. > Emotional sobriety requires separating your identity from your beliefs so you can critique ideas without attacking people. > To maintain high-level creativity, you must be willing to destroy your best work and start from scratch to avoid the trap of complacency. > Original ideas are often too complex for a thirty-second elevator pitch because they require nuance and unique execution to succeed. > Great creative work is the result of a methodical process that turns a bad initial idea into the least bad version possible. > We are often the most unreliable narrators of our own lives, creating excuses like a lack of time or access to avoid doing the work we are most capable of performing. > For many high-profile figures, authenticity is a manufactured deliverable designed through careful practice and performance art. > True freedom is the ability to criticize the government in a casual setting and immediately forget the conversation because there are no consequences. > In oppressive regimes, even small symbols of individual expression, like writing a country's name on a backpack, can result in severe punishment like expulsion. > Ideological capture occurs when your entire worldview can be predicted from a single opinion, which effectively stops independent thought. > Horizontal tribalism often distracts people from questioning the power structures that actually run the world. > The true value of a societal system is measured by how much it increases individual freedom and the ability to build a better life within a single generation. > High achievement is frequently driven by the desire to prove critics wrong, which can be a more powerful motivator than personal validation. > Motivation fueled by revenge and adrenaline is like fire. It can build a career, but without boundaries, it has the power to destroy everything you have created. > Managing a large family alongside a career requires extreme optimization of every second, often revealing a hidden preference for chaos over quiet. > Passion should precede the business. Polina built a following for years based on her genuine interest in stories before turning it into a full-time career. > Mental models are rarely stated directly. They are often inferred by looking for patterns across many different interviews and research sources. Link below for full episode notes!

Some highlights from this episode: > The survivorship bias argument for architecture is flawed because modern replacements are almost never more beautiful than the pre-1930s buildings they destroy. > Economic decline often serves as a preservation tool, as seen in Dublin where a lack of redevelopment kept its beautiful Georgian architecture intact for centuries. > Historical buildings often avoid being ugly because even their most utilitarian versions, like mills or prisons, used high-quality materials and consistent forms that we now view as aesthetic. > People are less likely to follow elite trends when those trends have real-world consequences for their comfort or enjoyment. > Architectural pastiche often triggers a fancy dress effect where a building looks like a costume because it breaks the current social language of fashion. > The disappearance of architectural ornament was driven more by a shift in elite taste and cost cutting justifications than by an actual increase in production costs. > Mechanization in the late 19th century actually made architectural decoration cheaper and more prevalent until modernism turned the lack of ornament into a status symbol. > One theory suggests modernism emerged because mass-produced ornament became too cheap for elites to use as a status symbol. > Status markers like modernism are adaptive rather than planned. They persist because they are difficult for the general public to enjoy or mimic, creating a permanent barrier to entry. > Objectively bad art can serve as a powerful social signal because it creates a barrier that most people will never naturally cross. > In egalitarian societies, elite barriers must become hidden because explicit social exclusion is no longer socially acceptable. > Architectural beauty is often tied to felt tectonics, which is the intuitive need to see how a building's weight is being supported. > Modern materials like reinforced concrete can create a visual mismatch where a structure is scientifically strong but appears dangerously spindly to the human eye. > Human preference for symmetrical architecture may be an evolutionary byproduct of our biological focus on recognizing faces. > Visual hierarchy through multiple scales of detail makes buildings more legible and sympathetic to the human eye. > High-quality materials like stone and timber provide a safety net for architecture, making it difficult to design a truly ugly building. > We often prefer handmade bricks over industrial versions because they age and break in ways that mimic natural patterns rather than looking like chemical erosion. > Architects often fail by designing for a plan view that looks good on paper but ignores how humans actually experience buildings through perspective. > Geometric perfection in urban planning is often illegible at ground level. Irregular or lumpy spaces often feel more natural and orderly to pedestrians than mathematically perfect shapes. > Aerial perspectives from social media provide a view of urban design that original architects never intended. This shift prioritizes how a city looks on a map over how it feels to walk through. Link to full episode notes below!

@SCP_Hughes "We often prefer handmade bricks over industrial versions because they age and break in ways that mimic natural patterns rather than looking like chemical erosion." Episode notes here if you prefer reading! podchemy.com/notes/did-stat…

Why do new buildings seem, on average, uglier than old buildings? We discuss some options: - Survivorship bias: only the beautiful old buildings have survived (we reject this option); - Cycles of taste: everyone always finds new buildings uglier (we mostly reject this too); - Ornament became too expensive because of rising labour costs (we reject this); - Ornament became too cheap because of mechanisation and then became low status (we reject this); - Some sort of Protestant or Puritan anti-beauty inheritance (we are doubtful); - Some kind of elite status game, perhaps a response to democratisation or elite overproduction (we think there is promise here, but serious work is needed on the details). I discuss this and more with @Aria_Babu and @bswud. Apple podcasts: podcasts.apple.com/gb/podcast/did… Spotify: open.spotify.com/episode/2pIka6… Youtube: youtube.com/watch?v=qvueKt…