tech
The a16z Show discusses tech and culture trends, news, and the future – especially as ‘software eats the world’. It features industry experts, business leaders, and other interesting thinkers and voices from around the world. This show is produced by Andreessen Horowitz (aka “a16z”), a Silicon Valley-based venture capital firm. Multiple episodes are released every week; visit a16z.com for more details and to sign up for our newsletters and other content as well!
Andrew Huberman outlines his three primary evidence-based habits for a healthier life: getting morning bright light ideally from sunlight, following a specific weight training protocol, and optimizing sleep quality.
Andrew Huberman describes his long-standing, time-efficient exercise program consisting of three days of weight training per week—focusing on legs, torso push-pull, and smaller muscle groups—using low volume and momentary muscular failure.
Discussing the science of meditation with University of Wisconsin researchers, Andrew Huberman explains that the true benefit of meditation is not achieving a peaceful state, but learning to observe mental chaos without reacting.
Andrew Huberman discusses a clinical study showing that taking N-acetylcysteine (NAC) at 600 to 900 milligrams three times per day can significantly reduce the likelihood of catching the flu after exposure.
Andrew Huberman predicts that the next wave of neurotechnology will feature sleep masks that control eye movements and use electrical stimulation to help users fall asleep and wake up rapidly.
Josh Gruenstein, co-founder and CEO of Tutor Intelligence, discusses the company's five-and-a-half-year history, noting that they employ about 160 people in Boston and have helped manufacture and ship over a quarter of a billion products. He explains that 99% of robot pitches focus solely on models, while Tudor focuses on full-stack deployment, manufacturing robots at scale, and entering greenfield markets where 90% of US factories and warehouses have no robots.
Josh Gruenstein outlines the morphology of Tudor's two primary robots: Cassie, a heavy-duty single-arm robot designed for pallet and box transport, and Sunny, an industrial semi-humanoid with wheels and higher payload capacity built for collaborative safety standards. He notes that Cassie and Sunny can address 80% to 90% of work inside manufacturing and logistics.
Josh Gruenstein states that Tudor achieves an average time of 40 to 60 days from a first sales call to a signed contract, and about 40 days from contract to robots working on-site. He adds that they frequently stand up robots in production within a day or even hours of arriving at a customer facility.
Josh Gruenstein shares that Tudor has built a 100-robot fleet called Data Factory One in Boston to serve as a research test bed for real-world reinforcement learning at scale. He notes this allows them to run methods that are typically inaccessible in traditional academia or industry due to batch size requirements.
Luca Ferrari explains that Bending Spoons buys technology companies where they see opportunities to make them more successful, integrating them onto a single platform and core team. He notes that the company has completed over 50 acquisitions, including Airtable, AOL, Vimeo, Eventbrite, Komoot, and WeTransfer.
Luca Ferrari discusses Bending Spoons' approach to debt and risk, explaining that the company maintains self-imposed leverage limits well below 3x, currently operating around 2.4 to 2.5. He states that the company aims to ensure it can pay back 100% of its debt before maturity even in severely adverse scenarios.
Luca Ferrari reveals that coding inside Bending Spoons is now heavily driven by artificial intelligence. He notes that while coding written by AI was at 0% two years ago, it recently reached 94%.
David Booth and Dhruv Gupta from a16z's new media team discussed the launch of the Cosign Awards, designed to shine a light on top-performing operators, companies with high talent density, and products. According to Booth, the awards feature categories spanning people, companies, and products, including role-based categories and popular selections like rising stars.
Dhruv Gupta and David Booth detailed how Cosign is building an agent-native professional networking platform that leverages a relationship graph rather than traditional job boards. They explained that users can connect personal AI agents and Model Context Protocols (MCPs) to search the platform and negotiate directly with company agents on job opportunities.
David Booth outlined the structural layers of Cosign, noting that the platform features Wikipedia-style scaffolded profiles for 40,000 to 50,000 individuals alongside lightweight curation lists created by investors and industry operators to help surface hidden talent networks.
Jay Madeswaran, founder and CEO of legal AI startup EVE, shared that law firms utilizing EVE's end-to-end platform have seen significant operational scaling, with one Los Angeles firm increasing its client capacity from 100 clients per attorney to 500. Madeswaran and a16z general partner Alex Rampell discussed how AI can overcome the labor constraints of plaintiff law to expand access to representation.
Jay Madeswaran explained that 96% of labor and employment cases are rejected because the manual work required to research case law and handle intake makes small claims economically unviable for lawyers. EVE aims to automate these processes to make such cases profitable.
Discussing the technical progress of EVE, Jay Madeswaran noted that a single legal case involves roughly 10,000 back-and-forth actions, including outbound calls to healthcare companies and insurance providers. While human touch is still required for some tasks, Madeswaran stated that the company is getting close to achieving fully autonomous end-to-end case resolution.
Mark Cuban shares that his company, Cost Plus Drugs, has expanded its offerings from 111 medications at launch to approximately 4,000 SKUs. He notes that the business is growing steadily and facing the primary challenge of keeping up with demand.
Mark Cuban explains that Cost Plus Drugs developed robotic and AI-powered manufacturing facilities shrunk down into tractor-trailer-sized pods. He notes that the government was their first customer, purchasing pods that can be configured to produce various specific injectable or tablet medications.
Mark Cuban points out that UnitedHealth operates over 3,000 subsidiaries with $150 billion in intercompany transfers annually. He argues that such vertical integration and lack of transparency make healthcare pricing opaque and unfair to consumers.
Mark Cuban argues that the best short-term use of artificial intelligence for employers is running complex PBM and health benefits contracts through AI models. He suggests asking a single prompt to instantly reveal where organizations are getting financially exploited.
Joe Liemandt explains that Alpha School has scaled rapidly, adding 12 new campuses last year to reach a total of 50. He notes that student demand is extremely high, driven by a model where core academics take only two hours a day.
Joe Liemandt describes how Alpha School teaches teamwork by having third graders go through escape rooms while mic'd up, using AI to analyze transcripts of their communication and requiring an 85% team-behavior score to pass.
Joe Liemandt shares that over 70 students from MIT, Stanford, and UT interned at Alpha over the summer, spending 80 hours a week researching learning science and building a small language model designed for education.
Joe Liemandt discusses Founders School, a specialized high school designed for students who want to build businesses, noting that students can earn their tuition back if they achieve specific milestones like a million dollars in profit.
Ari Weinstein and James Sun from OpenAI discussed their latest computer-using model, Astra, noting that it is faster than the average human at computer tasks and can defeat prior models in video games like Fallout. They highlighted that Astra uses fewer thinking tokens and offers greater reliability around the edges compared to previous models.
James Sun and Ari Weinstein explained how OpenAI's Astra model incorporates enterprise safety features, including the Guardian system to review every action and check against internal intent. Enterprises can also establish deterministic controls, like sandboxes and human-in-the-loop triggers for sensitive tasks such as uploads and downloads.
James Sun outlined practical enterprise applications for computer-using agents, such as automating procurement pipelines, month-end financial ledger reconciliation, and market research. He noted these tools can navigate older legacy enterprise software that lacks modern APIs.
Ari Weinstein and James Sun discussed how standards like WebMCP enable computer-using agents to collaborate with humans on websites without entirely replacing the traditional browsing experience. They noted this allows agents to handle tedious work while letting users complete transactions themselves.
James Sun revealed that OpenAI has witnessed week-over-week growth comparable to Y Combinator levels after integrating browser use into existing agentic products. He mentioned that the next big inflection point for the company will be scaling cloud-based computing capabilities for mobile users.
Box CEO Aaron Levie explains that unlike past technology waves where startups ran at the speed of their keyboard input, today's AI-first startups can deploy background agents to act as engineers, allowing small teams to produce the output of organizations many times their size.
Aaron Levie predicts that hundreds of companies valued between $1 billion and $10 billion and beyond will emerge by building specialized AI agents to automate distinct job functions across various industries, effectively outrunning both traditional software incumbents and services firms.
Aaron Levie argues that further advances in AI model capability primarily benefit enterprise and specialized industry applications like life sciences and law rather than general consumer use cases, which are already largely satisfied.
Aaron Levie stresses that despite the rise of AI tools, fundamental tech distribution and go-to-market strategies remain unchanged, emphasizing that simply building a great product does not guarantee user acquisition.
Ben Horowitz discusses how past technology waves like the internet were networks, whereas AI represents a fundamentally new type of computer—a probabilistic computer. He notes that AI applications can instantly reach massive audiences because they do not rely on building up a new network of users from scratch.
Ben Horowitz argues that dismissing AI applications as mere 'GPT wrappers' is mistaken, comparing it to early criticisms of relational databases as just 'RDBMS wrappers.' He explains that applications like Cursor use multiple sophisticated models and incorporate unique data to create deep differentiation.
Ben Horowitz shares his conviction that AI-generated video, entertainment, and storytelling will evolve into an entirely new medium rather than just automatically generating traditional movies or TV shows, comparing the shift to how early cinema initially just filmed theatrical plays.
Ben Horowitz highlights a high-stakes decision from his entrepreneurial past, noting that his company went public on only $2 million in trailing 12-month revenue with a forecast of reaching $75 million because the alternative was bankruptcy.
Vlad Kyle, CEO of Leo, argues that AI-native startups have an advantage over established software vendors in enterprise procurement because incumbents are limited to their system of record and don't handle the end-to-end job. He explains that much of the actual work, such as negotiations and exception handling, happens outside these systems, across emails, spreadsheets, and contracts.
Seema Amble outlines four types of AI agents: retrieval agents (like chatbots) that fetch information, process agents that execute tasks based on policies, policy agents that apply judgment to complex situations, and principal agents that can negotiate or make strategic decisions. She notes that most incumbents started with retrieval agents and are slowly moving towards process and policy capabilities.
Seema Amble explains that while incumbents have distribution advantages and customer trust, internal incentive issues prevent them from developing more advanced AI agents. She notes that deeper AI capabilities can conflict with existing product offerings, creating internal competition and making it difficult to align different VPs and sales teams on a unified strategy.
Vlad Kyle of Leo discusses how they build trust for their AI agents, especially for autonomous negotiations. He explains that no enterprise starts with fully autonomous agents from day one due to a lack of trust in both the company and the technology. Leo uses a 'human-in-the-loop' approach, feeding the agent feedback and learnings to gradually increase trust for larger negotiations.
Vlad Kyle highlights that procurement involves more than just negotiating prices; it's an intense process involving multiple stakeholders across legal, cost engineering, finance, and procurement. He explains that while ERP systems show a final price, they don't capture the extensive work, emails, and analysis that happen behind the scenes, which is where AI agents can add significant value.
Vlad Kyle describes Leo's approach as a multi-agent system designed to handle the end-to-end procurement process. He explains that agents across retrieval, process, policy, and principal categories must share information and work in a specific order to solve complex tasks that often involve multiple stakeholders and software tools.
Vlad Kyle explains how AI agents can intervene in physical supply chain problems by leveraging context to predict and mitigate disruptions. By analyzing supplier reliability, potential delays, and even external news, agents can help protect goods and ensure more reliable deliveries, even when choosing between cheaper but less reliable suppliers versus more expensive, dependable ones.
Vlad Kyle and Seema Amble discuss why companies often fail when trying to build AI solutions internally. While simple use cases like data retrieval into an ERP might be built quickly, achieving high performance (over 70-80%) and integrating complex workflows, exception handling, and diverse systems requires specialized expertise and infrastructure that dedicated companies like Leo provide.
Vlad Kyle predicts that in the future, both buyers and suppliers will utilize AI agents, fundamentally changing business transactions. While sales teams have historically been ahead in AI adoption, he notes that suppliers in Fortune 100 companies are now exploring AI. Leo aims to influence suppliers to adopt their agents, enabling buyers to dictate the AI tools used, thereby owning both sides of the transaction.
Vlad Kyle believes AI agents represent a trillion-dollar business opportunity in procurement by addressing its historically frustrating and manual nature. He argues that unlike previous tools that only improved efficiency, AI agents can fundamentally change how people work, driving significant business impact and savings, potentially increasing profits by 1% which is equivalent to a 10% increase in sales.
Barrett Lyon has launched DocsNet, a new infrastructure aimed at rebuilding the internet with a focus on privacy and peer-to-peer communication. The platform is designed to protect users from advertisers, trackers, malware, and foreign adversaries by eliminating central servers and enabling encrypted, direct connections.
Internet infrastructure pioneer Barrett Lyon recounted a harrowing experience where his plane lost power shortly after takeoff, forcing an emergency landing. This incident, combined with his broader concerns about online privacy and censorship, significantly influenced his drive to develop new technologies like DocsNet.
Barrett Lyon explained that DocsNet is an attempt to address the stagnation in internet protocol design and the increasing erosion of user privacy. He noted that much of the internet's current infrastructure is built on top of older protocols, and DocsNet seeks to create a more open and experimental environment.
Barrett Lyon argued that existing VPN services are not sufficient for true online privacy, describing them as mere entry and exit points that lease servers from third parties. He proposes DocsNet as a more robust alternative that offers greater control and security.
DocsNet enables direct, encrypted peer-to-peer communication, eliminating the need for central servers and personal identifiers like phone numbers. Lyon highlighted that this approach allows for features like large file transfers with no size limits, contrasting it with the server-dependent models of apps like WhatsApp and Signal.
Barrett Lyon expressed distrust in many VPN services, calling them 'creepy' and highlighting the risks associated with leased infrastructure and potential insider threats. DocsNet addresses this by owning its infrastructure end-to-end, utilizing AI for network administration to minimize human access and potential vulnerabilities.
Barrett Lyon discussed how his early experiences with cyberattacks on his own servers in high school led him to develop DDoS mitigation techniques. This expertise eventually formed the basis for companies like Prolexic (now Akamai Prolexic) and DefenseNet (which became F5 Silverline), pioneering early cloud-based security services.
Barrett Lyon discussed his art project, Opti, which mapped the entire internet using techniques similar to penetration testing. This project, born from a $50 bet and developed during a period of recovery from a bicycle accident, was eventually acquired by the Museum of Modern Art (MoMA) and included in its permanent collection.
Barrett Lyon mentioned a book about his early work, which details his efforts to track down and pursue aggressive Russian cybercriminals involved in extortion and client-side attacks. The book covers incidents where businesses suffered significant losses, particularly during peak sales seasons, and Lyon's team's attempts to bring these criminals to justice, including a trip to Moscow.
The current investment cycle, driven by AI, is one of the largest in modern history, with high-tech equipment, software, and R&D accounting for 55% of US capital spending. This buildout has surpassed railroads as a percentage of GDP, and projections suggest a 20x increase in cumulative investment over the next decade.
The five largest hyperscalers (Alphabet, Amazon, Meta, Microsoft, and Oracle) are projected to spend over $1 trillion annually on capital expenditures starting in 2027. This massive investment is driven by the continuously increasing demand for compute power, which still outstrips supply.
Despite market highs, trading multiples for the S&P 500 are down, indicating that performance is driven by fundamental earnings rather than inflated valuations. This contrasts with the dot-com boom, where companies traded at extremely high P/E ratios, unlike today's market where earnings multiples are below 20.
The increasing prevalence of AI agents, capable of parallel and long-running tasks, is significantly boosting compute requirements. These agents can autonomously write code, search, and perform tasks, leading to a much larger demand for computational resources than previously anticipated.
The current demand for compute power is described as a 'model buster,' indicating it's pushing the boundaries of what's possible. This insatiable demand led to OpenAI pausing new subscriptions for its pro plans, highlighting the immense pressure on resources and the prescience of early, large compute commitments.
While overall AI adoption in enterprises is still in its early stages, a segment of 'power users' is emerging, demonstrating significant impact. These advanced users are leveraging AI improvements to drive both growth and better profit margins.
The personal AI agent space has exploded in recent weeks, moving from experiment to one of the fastest-moving areas in consumer AI. David Pawlan notes that while it's the 'wild wild west,' there's intense interest in what these agents can do day-to-day.
David Pawlan created 'Assistant Benchmark,' a consumer-facing platform that compares AI assistants based on their performance across various use cases. The site has seen over 100,000 visitors and significant outreach from founders, indicating strong industry curiosity.
The top use cases for AI assistants are daily administrative tasks like inbox management, agent orchestration, and finance-related applications. While travel is frequently discussed, it's considered less of a day-to-day utility and more of an attention grab.
The most successful AI agents are those that function invisibly as 'cost savers,' rather than just offering marginal efficiency gains. Examples include automatically filing reimbursements, securing flight credits when prices drop, and optimizing home sprinkler systems to reduce water bills.
The AI assistant landscape is vast, with over 122 identified agents categorized into B2C generalized agents like Instinct and Muse, and more specialized ones for travel (Sora, Meso) or email. B2B workflow assistants also exist, functioning as executive assistants.
Harvey, a legal tech company, has experienced explosive growth, adding over 1,000 employees since Maggie Landers joined a year ago. The company is on track to reach 2,000 employees by the end of the year, with over 70% of its workforce having joined in 2026.
Harvey has seen a dramatic increase in its valuation over the past year, growing from $3 billion in February 2025 to $11 billion by March 2026. This financial surge coincides with significant company and customer growth.
Harvey fosters a culture of 'progress over perfection,' encouraging employees to experiment, fail fast, and learn. The company emphasizes giving employees autonomy to move quickly and build novel solutions in the legal industry.
Harvey seeks candidates who are comfortable with ambiguity and possess strong judgment, essential traits for succeeding in a fast-moving environment. The company has a self-selection process that attracts intrinsically motivated individuals aiming to build something significant.
Harvey operates with a 'jobs not finished' culture, a core value that balances celebrating accomplishments with a constant focus on future goals. This mentality drives the company's rapid progress and ambition.
Despite celebrating its fourth anniversary and significant growth, Harvey maintains a startup mentality, with no one internally feeling they have reached their full potential. The company celebrates wins but remains focused on future aspirations.
Showing the latest 15 of 54 covered episodes.