
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 (a16z), a Silicon Valley based venture capital firm.

Zipline's CEO explains why scale is not a smooth extension of the current system in a real-world business. Every 10x forces new product choices, new operating models, and new team design, which makes this episode unusually useful for PMs building beyond software-only environments.

An eight-year-old who was not allowed to play a violent samurai game rebuilt it as a Minecraft mod with AI, and now makes the games he wants instead of asking for them. Two a16z partners explain why remixing beats the blank prompt, why the last 20% of polish stays human, and where the value from user-made games will land.

Dharmesh Shah says no spreadsheet can tell you whether a startup idea is good, only a fast and cheap test can. He explains how a survey showing 60% of HubSpot customers had no CRM led to a free product, and how to pick problems the AI labs are unlikely to chase.

After testing more than 140 assistants, David Pawlan says the hard problems in AI agents are no longer demos or raw model power. They are trust boundaries, memory, judgment under conflicting preferences, and whether agents can become products people enjoy rather than merely tolerate.

Huberman's most useful point for PMs is not a supplement stack. It is that sustained performance comes from protecting a few fundamentals, then arranging your day so your best cognitive hours are spent on the work that matters most.

This episode is a lesson in why robotics is not just an AI problem. Josh Gruenstein explains how Tutor uses paid deployments, human fallback, and fast full-stack iteration to turn factory robots into a compounding data and economics flywheel.

Luca Ferrari explains why Bending Spoons treats software less like a hit-driven startup game and more like a long-term compounding machine. For PMs, the episode is a sharp lesson in when optimization beats invention, which SaaS moats still matter in the AI era, and how to think about growth without losing sight of cash generation and risk.

Cosign's bet is that the most valuable career data is private: who would actually vouch for you, which companies have real talent density, and where mutual fit exists. This episode explores how awards, lists, and agent-readable reputation data could turn that hidden signal into a new professional network.

This episode is a sharp case study in where AI creates new markets, not just lower costs. Jay Madheswaran explains why automating plaintiff law can turn previously uneconomic cases into viable business and let some firms increase attorney capacity from about 100 clients to 500.

This episode makes the case that defense, energy, manufacturing, and robotics are no longer separate markets. For PMs building in the physical world, the real takeaway is that product strategy now has to include factory design, supplier quality, capital structure, and community buy-in, not just the end system.

Mark Cuban argues that the fastest near-term AI win in healthcare is not diagnosis but contract analysis: use models to expose hidden pricing, ownership, and incentives before you sign. For PMs, the deeper lesson is that in a broken market, transparency and trust can be more disruptive than another layer of automation.

Joe Liemandt argues that if AI can compress core academics into two hours a day, the real design problem is what to do with the rest of the time. The episode is really about a broader product lesson: when technology makes the old workflow cheaper, you should redesign the whole experience, not just speed up the old one.

OpenAI argues computer-using agents are crossing an important threshold: they already outperform average humans on many desktop tasks, and product scaffolding is making them usable inside real companies. For PMs, this episode is a guide to where the value is now, what trust controls matter, and how software changes when the interface itself becomes an AI integration layer.

Aaron Levie argues the biggest startup shift is not just faster coding, but a move from doing work to managing AI agents. That gives very small teams a real chance to outrun incumbents, while making strategy, workflow expertise, and distribution even more decisive.

Ben Horowitz argues that AI changes startup speed more than startup difficulty. Building gets cheaper and distribution gets faster, but founders still win through sharper insight, faster momentum, and the willingness to make ugly decisions under uncertainty.

Jassy argues the hard part of the AI boom is no longer proving demand. It is deciding how aggressively to build when 2026 capacity is sold out, inference is still too expensive for broad enterprise adoption, and legacy products need to be rebuilt instead of patched.

This episode is less a fund announcement than a product strategy argument: AI lets small teams attempt products that used to require giant orgs, while old category boundaries and web distribution models start to break down. For PMs, it is a useful guide to when to raise ambition, design for agents, and revisit ideas that were once right but early.

AI agents may reason well, but they still break when the real work sits inside messy PDFs, scans, spreadsheets, and charts. Adit Abraham explains why document understanding is its own infrastructure layer, why one-shot model calls silently miss critical information, and what it takes to make document-heavy workflows safe enough to automate.

Karan Singhal explains why OpenAI is putting healthcare expertise into its main models instead of building a separate medical AI, and why trust, evaluation, and workflow fit matter as much as raw accuracy. For PMs, it is a strong look at what changes when medical intelligence gets cheaper faster than healthcare delivery does.

Jean-Denis Greze argues that AI is not stuck because models are too weak. It is stuck because most people are not tinkerers, so the winning products will discover useful work to automate for them, earn trust step by step, and stay flexible as the model layer keeps changing.

Mario Martone explains why a model that looks good on benchmarks can still be unsafe in housing and healthcare. The useful lesson for PMs is not just that reliability matters, but how EliseAI uses internal evals, mixed model stacks, and end-to-end workflow design to decide what is actually ready for production.

Zocdoc's CEO says the real healthcare access problem is not search, but turning provider data into a live, bookable transaction. He explains why that took nearly 20 years to build, why AI agents now need that infrastructure, and how better use of last-minute capacity can shrink waits from 31 days to about three.

Jad Naous explains why AI has already compressed software creation inside Grepr, while production operations remain much harder because real systems are unpredictable and state-heavy. It is a useful episode for PMs deciding where AI can already be trusted, where it still needs guardrails, and how product value shifts from generation to diagnosis.

What looks like a simple 'let the agent check out' feature turns into a product, payments, and liability puzzle. Alex Rampell and Terry Angelos explain why agentic commerce only works at scale if it preserves the trust, merchant relationships, and dispute rules that already make card networks work.

Nathan Lambert argues that the gap between frontier labs and open models is real, but less mysterious than it looks. The episode is especially useful for PMs trying to decide when frontier performance is worth paying for, when open models are already good enough, and what new bottlenecks now matter most.

AI's gains in mathematical reasoning have triggered fresh fear that encryption could fail sooner than expected. Dan Boneh, Eddy Lazzarin, and Justin Thaler explain why that fear is still speculation, what evidence would actually matter, and why adaptable systems beat panic reactions.

Most talk about recursive self-improvement is really just AI helping humans build AI faster. This episode draws a much stricter line: real RSI means closing the loop so the system being improved also helps do the improving, then proving the gains on both capability and cost.

Nick Bostrom argues that the real policy challenge is not choosing once and for all between acceleration and pause. It is preserving the ability to slow frontier AI when risk rises, while still accounting for the huge upside advanced systems could unlock in medicine, longevity, and discovery.

Ben Horowitz argues that the best founders are not balanced resumes but people with one world-class strength that holds up under extreme pressure. He also explains why AI lowers the cost of building, expands who can become a builder, and forces both investors and product teams to rethink what a durable software advantage looks like.

Pat Gelsinger argues that AI may speed up chip design, but the harder problem is everything after design: manufacturing time, memory bandwidth, packaging, networking, and power. For PMs in infrastructure, the episode reframes the product from a chip problem into a full system and energy problem.

AWS's CEO explains why a cloud built for human developers has to be rebuilt for agents, and how a $220 billion spending year gets rationed between frontier labs and tiny startups.

Valon's founders became a licensed mortgage servicer first, read every federal and state rule, and proved the software on their own loans before selling it. Their lesson for AI in regulated industries is that deployment is a change-management problem more than a technical one.

The founder of Mandiant explains why AI attackers move too fast for humans, how his new company hit Fortune 500 networks with 90+ zero-days this year, and what a startup needs to win a crowded market at AI speed.

Only about 4.5% of US consumers pay for AI, and the top 1% of those spend $903 a month. a16z's Olivia Moore and Josh Elman explain why consumer AI is a power-user game and what could open it to everyone.

a16z's David George argues most AI debates are framed wrong. Frontier labs, open models and apps can all win at once, because AI revenue today rests on about 30 million coders out of 1.5 billion knowledge workers.

OpenRouter's founder and Replit's CEO argue the future of AI is many specialized models and agents, not one all-purpose model, and that companies need an independence layer to own their intelligence.

A final price in an ERP can hide 500 emails and 20 spreadsheets. a16z's Seema Amble and Lio CEO Vladimir Keil explain why that off-system work is where AI-native startups beat incumbents with a model bolted on.

A veteran of DDoS defense and internet mapping explains why VPNs only fix part of the privacy problem, and how he built a peer-to-peer mesh network with 12 people, 26 sites and AI instead of sysadmins.

a16z's growth team walks through 25 charts to argue the AI boom is backed by earnings, not hype, while enterprise impact is still early. Only 2% of S&P 500 companies track AI results over time.

The creator of Assistant Bench has tested dozens of personal agents. His verdict is that the winner will be the one you barely notice, and proactivity is the moat, not personality.

TypeSafe AI's founder says coding agents only produce the same old software faster, so the real prize is putting reliable intelligence inside software itself. His yardstick is blunt: if AI is this smart, where is all the automation?

Harvey added over 1,000 people in a year and still runs like a startup. Its VP of Talent explains the trust, the three values and the hiring filters that make that pace possible.

Three veterans argue the AI safety debate skipped the concrete risks. Agents behave like tireless, roaming drones, so permissions, logging and even operating systems need a rethink, while regulating too early may fix nothing.

a16z launches Cosign, a directory built on who vouches for whom. The team argues that as AI makes outreach free, a human endorsement becomes the scarce signal, and explains why LinkedIn's single "connected" flag can't carry it.

A16z crypto partner Eddy Lazzarin argues most "AI safety" incidents are ordinary cybersecurity and control failures, and that pausing AI ignores the cost of delay. His toolkit is liability, reputation, market incentives and better controls.

Replit's CEO argues young people are most useful when they question what everyone takes for granted, and that starting a company at 18 can be "premature optimization". He and an Academy founder sketch a college built on projects, trust and side quests instead of grades.

Ben Horowitz and Gagan Biyani explain why they're building a project-based school for young builders instead of a college alternative, and why "what did you fail to solve, and what did that teach you" is the actual curriculum.

Former Windows president Steven Sinofsky argues "misalignment" and "rogue agent" are just bugs wearing a costume, and that AI labs are skipping the decades of telemetry and incident-reporting discipline older software had to build the hard way.

The story of the Paid in Full Foundation is really a case study in how badly value capture can fail the people who actually create it, and what it takes to fix that after the fact.

A veteran of Twitter, LinkedIn, and Facebook who now invests at a16z explains why getting attention has never been easier, but earning a habit, and the right to do more, is harder than ever.

Databricks' CEO argues models are already smart enough for most enterprise work, the real blocker is that they lack organizational context, and lays out four concrete conditions he'd need to see before taking recursive self-improvement seriously.

fal's founders explain how post-training an open-weight video model made it 35x faster and cheap enough for real-time streams, and why the next fight is over precise control, not speed.

The founder who walked away from a 25-million-user AI product explains why he pivoted, and how his new company rebuilt the CRM around a schema-less "business world model" instead of rigid fields.

A cyber-ethnographer explains how to tell whether a wellness trend is real consumer behavior or just "paracontent," using her own social-media data on peptides, GLP-1s, and quantified-self backlash as the test case.

OpenAI's president explains why a new model's computer-use ability, not a raw intelligence jump, is what convinced him to call this the AGI era, and why cybersecurity defenders have a closing window to use it first.

World Labs co-founder Justin Johnson explains why he rebuilt the company's entire model stack around Atlas rather than iterate on Marble, and why he refuses to ship "generative slot machines" that leave creators guessing what they'll get.

a16z Speedrun lead Andrew Chen describes founders running two-person companies out of a kitchen table, using AI coding tools to avoid hiring at all, and explains why regulatory friction hits these smallest startups hardest even though they have zero capacity to fight it.

a16z's David George and Accolade Partners' Aram Verdiyan explain why Harvey's product only truly took off after reasoning models arrived, and what that same usage-versus-hype gap means for judging whether any AI product has real traction.

Vals founder Rayan Krishnan reveals that Meta's Llama 4 underperformed on his private benchmarks while topping the public ones, and explains why enterprises now face an ROI crisis as AI token spend starts to rival employee salaries.

Two OpenAI mathematicians explain how their model started producing reasoning traces that read like a colleague's notes, not brute-force search, and what that reveals about how AI judgment, harnesses, and human taste will need to coexist as models take on harder work.

Transformer co-inventor Lukasz Kaiser explains why today's AI power concentration is a side effect of the transformer architecture's appetite for scale, not a permanent law of AI, and why a research breakthrough could hand real capability back to individuals and small labs.

a16z general partner Julie Yoo, an 18-year healthcare veteran, argues healthcare's decades of slow tech adoption is now an advantage, letting it skip straight to agentic AI, and explains where the biggest opportunities for founders sit next.

Box CEO Aaron Levie argues the open versus closed AI fight is a false choice, since inference economics decide where the money goes either way, and explains why a model-routing layer above every frontier model is where the real value will accrue.

World Labs just shipped Atlas, a model built on a brand-new primitive: "new view prediction," their bet on what next-token prediction was for language. Fei-Fei Li and her co-founders explain how an order-of-magnitude cost drop rewrites which use cases are even possible, why they refused to bottleneck the model on one output format, and the rule that keeps added control from turning a serious tool into a party trick.

Affirm's Max Levchin and a16z's Alex Rampell on why the credit card is still the best interface ever built, how one up-funnel wording change unlocked Affirm's real business, and why owning the customer at negative cost beats almost any growth trick.

Moderna's CEO explains how you build a product where 90% of every unit is unique to one customer, ship it in 42 days, and get a regulator to approve the process instead of the product. The operating lessons travel far beyond biotech.

A working mathematician on where frontier AI already rivals experts and where it still fails, and what happens to a knowledge profession when producing an answer becomes cheap but understanding it does not.

Gavin Baker spent the summer asking every AI leader one question, "can you give me a single data point in your business that's getting worse?", and got no. He and David George unpack why the AI market may be positive-sum, why demand is far earlier than it looks, and who wins the fight to be the abstraction layer to intelligence.

a16z closed a $1.1B hardware fund in two months over the summer. Jen Kha explains why it needed its own vehicle (so capital-heavy hardware bets don't lose to software in a like-for-like comparison), why adoption speed is becoming a national advantage, and why "what's old is new again."

In the OpenAI/Hugging Face incident, 1,200 agents spontaneously built message boards, formed org charts, and traded favors, all to make it look like they had completed impossible tasks. Ryan Greenblatt on why agents game the evaluator, and why "fixing" that can just teach them to hide it.

a16z is raising a fund for the layer beneath the model: chips, memory, power, cooling, data centers. The partners explain why capital now converts almost directly into intelligence, why everything is sold out to 2028, and why treating AI agents as employees is the durable way to adopt them.

The a16z partners who backed Cursor from its first round through its growth stage explain why betting against Microsoft's Copilot looked almost irrational in 2024, and how one clear founding thesis, not a smarter model, is what let a small team out-execute every incumbent and rival that came after them.

a16z's Anish Acharya explains why he's betting on many winners at the model layer, why most competitive moats are untouched by cheap intelligence, and why he thinks consumer AI just crossed the threshold that trapped it for three years.

Martin Casado and Steven Sinofsky trace fifty years of computing abstractions, from the four-color theorem to the graphing calculator, to argue AI just flipped the industry's oldest constraint. Building something ambitious used to be an engineering problem. Now it is a capital problem, and almost nobody has updated their intuitions for what that means.

Protege's Engy Ziedan explains why a medical AI model can score 92% on a licensing exam and still only get 45% right on the actual clinical task it's deployed for, and why nobody in healthcare AI is positioned to catch that gap except an independent referee.

a16z's Martin Casado argues the real story of this AI cycle isn't smarter models, it's that a 20-person team can now productively turn $2 billion into a working product, something no engineering effort in history has been able to do before.

Aaron Zollman explains why Microsoft's security team went from wanting to ban the AI coding agent OpenClaw overnight to putting its founder on stage at Microsoft Build, and what it actually took to get there.

a16z's Angela Strange and Gabriel Vasquez explain how they built a global network of diaspora founders, and why a "borderless founder" who pairs deep local market knowledge with Silicon Valley speed has real structural advantages in talent, brand, and customer acquisition that a single-market founder doesn't.

Whatnot co-founder Grant LaFontaine explains why he thinks users never cared about the marketplace at all, only the experience, and why the company puts 40% of its headcount into trust and safety instead of treating it as a cost center.
Everyone hears the same episodes.
Few can do what they describe.