Collection
The wedge, the build, the sell. ~100 conversations with AI founders and domain experts, distilled into searchable insights. The lessons too valuable to leave buried in a 45-minute interview.
Simon Du (Chief Scientist, Apodex) explains why frontier AI labs hire on referral — and why building an open-source agent framework is the best way to break in.
Almost everything digital you 'own' is actually rented. What if software could be inherited? A jeweler in Newfoundland is testing that idea right now.
Simon Du (Chief Scientist, Apodex) explains why AI 'predicting the future' is really deep research — closer to a fundamental analyst than a quant model crunching market data.
It's not that AI writes for you. It's that everyone's AI writes from the same base model. A founder's take on why voice-flattening is the real AI writing problem.
Simon Du (Chief Scientist, Apodex) explains why a separate verification agent — not a better prompt or bigger model — is what keeps research agents from shipping confident, wrong answers.
The management burden shifts, it doesn't disappear. What a founder who went from 15 humans to 9 humans plus N agents learned about the new cognitive load.
Simon Du (Chief Scientist, Apodex) explains why decomposition — not a bigger context window — is what lets agents solve problems that burn 10M+ tokens.
Ali Parandeh, author of Building Generative AI Services with FastAPI (O'Reilly), breaks down the OWASP LLM vulnerability that turns helpful agents into destructive ones — and the guardrails you need before you ship.
Ali Parandeh, author of Building Generative AI Services with FastAPI (O'Reilly), walks through the Role-Context-Task framework — the simplest reliable way to get production-grade output from an LLM.
Most founders building vertical AI are racing for speed. Rec Technologies CEO Birju Kadakia argues the real wedge is something else entirely — and three pillars decide which vertical SaaS survives AI disruption.
Ali Parandeh, author of Building Generative AI Services with FastAPI (O'Reilly), explains why some AI startups are running their entire production stack on free cloud tiers — and the tradeoffs between Go and Python for AI services.
Ali Parandeh, author of Building Generative AI Services with FastAPI (O'Reilly), explains the three behavioral tests every production LLM service needs — Minimum Functional, Invariance, and Directional Expectation.
Kashish Gupta (Co-CEO, Hightouch) explains how a semantic layer turns a data warehouse into a context layer for LLMs — enabling accurate SQL generation and reducing hallucination.
Mike Taylor lays out the permanent underclass theory, then dismantles it with economics. His argument for why human intelligence is still the galaxy's best deal.
Deepak Bapat (CTO, Tabs) shares how AI coding tools eroded his own skills and why he now invests in junior engineers who write code themselves — plus why code review replaced coding exercises in interviews.
Mike Taylor ran a 50-person agency before building AI products. He argues the skills that matter most with AI agents are the same ones that worked on human teams.
Kashish Gupta (Co-CEO, Hightouch) argues that AI has made building UIs trivial — the real competitive advantage in marketing technology is now infrastructure.
Mike Taylor built Ask Rally to create synthetic focus groups from real human data. Here's what works, what breaks, and the 85% ceiling nobody talks about.
Kashish Gupta (Co-CEO, Hightouch) explains why assembling marketing content from existing brand assets beats generating it from scratch — and how it guarantees brand determinism.
Deepak Bapat (CTO, Tabs) predicts AI agents will let a $100M ARR company run its entire finance operation with a single person — here's the architecture that makes it possible.
John Berryman, founder of Arcturus Labs and early GitHub Copilot engineer, predicts that landing pages will be replaced by machine-readable llms.txt files paired with APIs — and explains why the timeline might be shorter than you think.
Mike Taylor, O'Reilly author and CEO of Ask Rally, explains why treating AI as a search engine over existing knowledge — not a creative generator — changes how you build with it.
Deepak Bapat (CTO, Tabs) explains why prompting on top of AI context objects causes more hallucinations — and why background agents beat chat for finance teams.
Kashish Gupta (Co-CEO, Hightouch) explains how his team discovered AI agents fabricating query results and built a smaller LLM to catch them in real time.
John Berryman, early GitHub Copilot engineer and author of Prompt Engineering for LLMs, explains why treating RAG as a black box guarantees failure and shares the decomposition framework he uses to debug production systems.
Deepak Bapat (CTO, Tabs) explains why context engineering matters more than model upgrades for production AI accuracy — and why his team pins to simpler models on purpose.
TextQL's CTO shares a hiring framework that measures what candidates add beyond what Claude or Copilot produces alone — the 'net improvement' approach.
Nicole Königstein (CEO, Quantmate) reframes multi-agent systems as distributed state machines with topologies, changing how builders think about agent communication and failure modes.
VeryAI has 3 engineers, 10 patents, and one of the largest palm datasets in the world. CEO Zach Meltzer explains why the Swiss army knife hiring model works for AI-era startups.
TextQL has 40 employees and no dedicated data team. Their CTO explains why data analytics became a shared responsibility — and what that means for hiring.
Nicole Königstein (CEO, Quantmate) shares research showing state-of-the-art AI models burn multiples of tokens compared to smaller open-source models, with worse results on some tasks.
Most people assume deepfake fraud means generating a fake face in real time. The reality is simpler and more common: deepfaking the photo on a passport, then matching it to a real face during verification.
Why dumping your database schema into Claude and letting it write SQL is a better starting point than most enterprise data teams think, according to TextQL's CTO.
Nicole Königstein (CEO, Quantmate) explains why vibe-coded agent systems pass tests but hemorrhage tokens in production — and why you can't outsource the thinking.
The internet was built to block bots. Now we want agents to act on our behalf — and the infrastructure can't tell the difference. Zach Meltzer (CEO, VeryAI) explains the identity gap.
Mark Hay, CTO of TextQL, explains why the biggest mistake in AI data analytics isn't doing too little — it's building too much infrastructure before delivering any value.
Nicole Königstein (CEO, Quantmate) explains why adding agents to a multi-agent system creates the same coordination overhead as adding engineers to a late software project.
Zach Meltzer (CEO, VeryAI) explains why verifying identity once at account creation isn't enough — and why re-verification at the moment of transaction is the real unlock for financial platforms.
Jesse Xu (CTO, Podqi) explains why brand protection's tight-knit industry generated 70%+ referral revenue — and the GTM lesson for any startup selling to professional communities.
Jesse Xu (CTO, Podqi) shares the framework that made his team pivot from patents to trademarks in under 4 months — if you can't validate your own AI's output, you're building in the wrong space.
Jesse Xu (CTO, Podqi) breaks down the 95/5 enforcement split — automate 95% of takedowns, investigate 5% for seven-figure legal settlements — and why clearing fakes grows baseline revenue 1-2%.
Jesse Xu (CTO, Podqi) explains the structural conflict of interest that lets fake products thrive on Amazon, Walmart, and Meta — and why platforms have no legal mandate to stop it.
How a Japanese philosophy concept became the product vision for an AI monitoring platform.
The author of LLM Engineer's Handbook tells engineers to stop specializing in code — and start specializing in a domain.
Understanding the economic layer of web agents and why in-house models became necessary.
Paul Iusztin (Founding AI Engineer, author of LLM Engineer's Handbook) on why engineers should stop reviewing AI-generated code line by line — and what to evaluate instead.
The counterintuitive technical decision that makes general-purpose web automation possible at scale.
Paul Iusztin (Founding AI Engineer, author of LLM Engineer's Handbook) on why MCP is right for dev tools but wrong for production applications you control end-to-end.
Why a single agent with 100 tools doesn't work — and how hierarchical architecture fixes it.
Paul Iusztin (Founding AI Engineer, author of LLM Engineer's Handbook) explains why RAG hurts more than it helps in production AI agents — and the 64,000-token threshold that should change how you think about retrieval.
Devi Parikh explains the hidden web of databases and forms that search engines can't crawl — and why monitoring it requires actual clicking.
Wiley Jones (CEO, Doss) argues hallucination is a system design problem, not a model problem. A statistician pushes back. The resolution changes how you think about building reliable agents.
Jay Hack (Head of AI, ClickUp) shares the Figma-inspired narrative strategy that works better than product pitches for hiring, fundraising, and customers.
Nic Baird (CEO, Koah Labs) explains why traditional publishers are adding AI interfaces to their sites — and what that means for monetization, advertising, and the future of the web.
Wiley Jones (CEO, Doss) explains why creativity, agency, and judgment matter more than framework knowledge when hiring engineers in the AI era — and how his interview process tests for it.
Joris Corthout (CEO, Prismax) on why human imperfection makes live visuals better, and what 20 years of running shows for 400,000 people taught him about attention.
Wiley Jones (CEO, Doss) shares the simple framework his team uses to decide what to build internally vs buy off the shelf — and why 'is it even worth it' matters as much as 'can we do it.'
Jay Hack (Head of AI, ClickUp) explains the Bitcoin-inspired architecture behind ClickUp's organizational memory system for AI agents.
Jay Hack (Head of AI, ClickUp) breaks down why most companies shouldn't build their own AI agent infrastructure — and the few that should.
Wiley Jones (CEO, Doss) explains the architectural requirement for software that can modify itself: fusing the application and its lifecycle into one runtime.
Jasmin Alić explains the counterintuitive reason growing your connections can actually kill your reach.
Wiley Jones (CEO, Doss) argues that product-based companies with no real technology or brand will be destroyed — and explains the durability test every SaaS company should run.
Jay Hack (Head of AI, ClickUp) explains why chat has more staying power than custom UIs — and what that means for how we build software.
Gil Feig watched AI become a better coder than most people he knows. His honest timeline for what comes next.
Wiley Jones (CEO, Doss) explains how control theory reframes AI hallucination as a basic probabilistic mistake — and why that changes everything about how you build agent systems.
Alex Beller (CEO, Postscript) on the hallucination patterns AI discovers when optimizing for conversions — and why supervisor agents are the only defense.
Xuan Zhao explains the strategic difference between designing AI as therapy versus designing it as wellbeing support.
Trip Adler breaks down the real decision: move fast and risk legal grey area, or license early and slow down. Here's how startup founders should think about it.
Ahmed Rashad (CEO, Perle.ai) on the four-pillar framework for scaling expert annotators — and why paying labelers more produces lower TCO than the cheap-labor model.
Jorge Colindres (Radical AI) breaks down how agentic systems translate from code to robotics, and why building agents for atoms is an order of magnitude more complex than agents for bits.
Dara Ladjevardian (CEO, Delphi) on why knowledge networks are fundamentally different from professional networks—and how they enable true expertise scaling.
Steve Ruiz had an insight observing how people actually use AI tools: we've been doing shallow knowledge work forever, accepting it as a ceiling. But that ceiling is about to collapse, and nobody's prepared for what's underneath.
Adam Biddlecombe refused to build on X, YouTube, or TikTok—and explains why obsessing over one platform creates better returns than being everywhere at once.
Ali Parandeh on why heavy industries slow to adopt AI, what it costs them, and how to change it.
Coco Mao (CEO, OpenArt) dismantles the day-one-visionary myth and explains why fast decisions beat right decisions when you're trying to find product-market fit.
Trip Adler shares the founder wisdom that kept him going through Scribd's hardest years — and why leaving after 16 years actually proved it works.
Alex Beller (CEO, Postscript) on the transition from SMS Sales Center in Phoenix — a human-powered operation — to AI agents handling customer conversations at scale.
Xuan Zhao explains the technical architecture that separates emotionally intelligent AI from generic chatbots.
Alex Reichenbach (CEO, Structify) explains why 2% better pipeline accuracy is the moat that justifies enterprise AI premium pricing — and why everything else is commoditized.
Olga Beregovaya (VP AI, Smartling) breaks down the real costs of building in-house translation systems, and when—if ever—it makes sense to build instead of buying.
Jorge Colindres (Radical AI) explains why fusion, batteries, and hypersonic flight all have the same bottleneck: nobody has figured out how to iterate on materials fast enough.
Tom Shapland (PM, LiveKit) explains why proprietary GUI tools feel easier but fail at scale, and why developers migrate to open-source for control.
Magnus Müller explains how autonomous agents flip the relationship between humans and AI—users set high-level goals, agents run continuously and propose actions.
Mercedes Bent breaks down five sources of market tailwinds—technology, regulatory, cultural, geographical, macro—and how founders use them to find uncontested territory.
Ahmed Rashad (CEO, Perle.ai) on the issue tree for diagnosing AI agent failures — and why most teams default to fixing the algorithm when the bottleneck is the training data.
John Kim (CEO, Paraform) on how job seekers, career-changers, and founders can position themselves for the AI economy by understanding what humans are uniquely good at.
Alex Beller (CEO, Postscript) on the $1,500-per-message liability that shapes SMS marketing constraints and why ignoring TCPA is a catastrophic business decision.
Jan Liphardt (CEO, OpenMind) explains why the counter-intuitive architecture choice of routing vision, audio, and sensors through natural language improves reliability and transparency.
Daniel Davis on how Wikipedia's edit structure allows bad actors to inject misinformation into AI training datasets — and what it means for model reliability.
Jorge Colindres (Radical AI) explains why materials companies must own their entire value chain — from discovery through manufacturing — or watch their breakthroughs die at the handoff.
Dara Ladjevardian (CEO, Delphi) on why feature creep kills AI companies faster than it kills traditional software—and how focus became his competitive advantage.
Early data on how PMs, designers, and engineers work together when AI eliminates the build bottleneck — and why the answer surprises everyone.
Coco Mao (CEO, OpenArt) explains why her decision not to train models was the right one — and what defensibility looks like at the AI application layer when foundation models are commoditizing.
Ali Parandeh explains why engineering companies spend millions on AI pilots that never ship — and how product mindset could save them.
Most mental health apps increase screen time. Xuan Zhao explains the specific design pattern that makes people use their phones less.
Adam Biddlecombe explains why monetizing content directly kills growth, and how great creators use content as a distribution engine for something else.
Alex Beller (CEO, Postscript) reveals how infinity testing found a simple variable that doubled conversion rates — without anyone telling the AI what to look for.
Trip Adler breaks down the multi-rights complexity that makes AI licensing books fundamentally different from licensing any other media.
Tom Shapland (PM, LiveKit) breaks down why enterprise voice agents need complex orchestration while consumer apps can ship simpler.
Alex Reichenbach (CEO, Structify) explains why multi-agent loops compound errors exponentially — and why deterministic code-generation is the only architecture that scales.
The MCP hype cycle shipped unreliable tools. Here's what enterprise-grade actually looks like.
Jorge Colindres (Radical AI) explains the data bias in scientific publishing and why capturing failure data is the missing bottleneck in materials discovery.
Everyone asks: you invented sketching-to-code with Make Real in 2024, before Lovable and Bolt existed. Why didn't you become Lovable? Steve's answer reveals why infrastructure is a bigger play than products.
Mercedes Bent explains why everything looking like ChatGPT is a problem of being early, and what new interface patterns will emerge as AI matures.
Olga Beregovaya (VP AI, Smartling) explains why voice-to-voice communication stays human longer than text translation, and which market survives the longest.
Jasmin Alić breaks down the complete GTM playbook: custom landing pages, post-to-conversion funnels, and the VA strategy that scales it.
John Kim (CEO, Paraform) explains why the first 5-10 hires are the most critical decision a founder will make, and why the tools most startups use to hire them are fundamentally broken.
Magnus Müller shares the production patterns for detecting and breaking hallucination loops in browser agents—when an LLM keeps retrying the same failed action.
Ahmed Rashad (CEO, Perle.ai) on why synthetic data can't fix the production gap that kills 95% of AI systems — and what real-world data labeling actually requires.
Dara Ladjevardian (CEO, Delphi) explains the technical architecture that preserves explainability and control—and why it matters for trust.
Daniel Davis explains the architectural difference between Neo4j-style property graphs and RDF models — and why the choice determines whether your AI can reason about trustworthiness.
Jan Liphardt (CEO, OpenMind) explains why hallucination isn't unique to AI — humans do it constantly. The real question is how to build systems that catch errors before they cause harm.
Joris Corthout (CEO, Prismax) explains why AI-generated video falls short for live events — and what it would take to change his mind.
Coco Mao (CEO, OpenArt) explains the filter that turned 350,000 Discord users into product clarity — listening to problems instead of feature requests, and how 'consistent characters' became visual storytelling.
The viable business models emerging — and why agencies are the biggest opportunity right now.
Ali Parandeh's real-world AI tool stack as a bootstrapped solo founder running Build Your AI — and why OpenClaw isn't in it.
Adam Biddlecombe's moment of dark clarity when Mindstream's master spreadsheet showed the business was 'destined to die'—and why the formulas being wrong changed everything.
The intelligence bottleneck everyone's obsessed with isn't actually your bottleneck. Gil Feig of Merge explains what is.
Woebot raised massive funding betting on healthcare integration. Xuan Zhao explains why that strategy failed and what works instead.
Trip Adler explains the framework that turned music piracy into a legal marketplace — and why AI licensing is following the same pattern 25 years later.
Tom Shapland (PM, LiveKit) reveals why overlapping speech breaks most voice agents, and the counterintuitive reason humans solve it naturally.
Steve Ruiz discovered that the same properties that make canvas tools great for human collaboration make them even better for AI. Here's why chat, despite its dominance, is a dead end for multi-agent work.
Olga Beregovaya (VP AI, Smartling) reveals why models fail catastrophically in low-resource languages, and what this means for global content accuracy.
Alex Reichenbach (CEO, Structify) explains why selling AI tools to heads of data is the slowest path to revenue — and which buyer actually closes.
Nic Baird (CEO, Koah Labs) explains the unit economics trap killing consumer AI apps — and why subscriptions alone can't save them.
Joris Corthout (CEO, Prismax) on how losing all revenue overnight led to a pivot from pre-rendered video to real-time 3D — and why they never went back.
Magnus Müller explains the architecture shift that transforms flaky LLM agents into production-ready systems through meta-learning and skill distillation.
Mercedes Bent shares her 7-point founder framework and the 4 categories that separate portfolio companies from ideas. Your TAM slide doesn't matter as much as you think.
John Kim (CEO, Paraform) debunks the founder myth that early-stage is the hardest part. He explains why the real difficulty hits after you've already proven product-market fit.
There's a playbook for moving from startups to Fortune 500s. Gil Feig used it for six years.
Jan Liphardt (CEO, OpenMind) explains why Tesla and BYD will manufacture robots at scale while AI startups become software providers — it's about decades of manufacturing expertise.
Dara Ladjevardian (CEO, Delphi) argues that as AI becomes smarter and cheaper, lived human experience—not intelligence—becomes the competitive advantage.
Ali Parandeh on what MIT research shows about AI overuse, why critical thinking atrophies, and how to avoid it.
Ahmed Rashad (CEO, Perle.ai) explains why $18-per-document labeling can be cheaper than $3-per-document labeling — and how to evaluate AI training data vendors on TCO.
Daniel Davis explains why language models have no temporal reasoning capability — and why that makes them systematically hallucinate stale information.
Jasmin Alić reveals why tech founders lead with features instead of outcomes — and the simple rewrite that fixes it.
It took 12 engineers a full year to build Koah Labs' ad infrastructure. Nic Baird breaks down why the technical and go-to-market challenges make DIY advertising a trap for AI startups.
Adam Biddlecombe on dashboard addiction, the problem with constant metrics checking, and Shawn Puri's 'one important thing' framework that helped him break the habit.
What non-engineers actually need to learn to build production apps with AI — and why it's nothing like traditional programming.
Trip Adler (CEO, Created by Humans) explains why AI companies prefer expensive litigation to cheaper licensing deals — and what changes that.
Harvard's emotional manipulation audit exposed a critical failure in how most AI mental health apps are built. Xuan Zhao explains what separates the outliers.
Sarah Chen (CEO, AgentStack) on why enterprise AI agent adoption is stuck and how to fix it.
Olga Beregovaya (VP AI, Smartling) explains the shift from source-target mapping to direct multilingual content generation, and why this changes everything about global content strategy.
Coco Mao (CEO, OpenArt) explains why launching when you're polished is the wrong move — and how the 'dressed-down' MVP actually tells you whether you have product-market fit.
Tom Shapland (PM, LiveKit) explains why the simpler speech-to-speech approach fails for production voice agents, and why the pipeline architecture wins where it matters.
Steve Ruiz (CEO, tldraw) built the infinite canvas that powers whiteboards for Google and Shopify on standard web technology. Here's why that bet—HTML and CSS instead of GPU rendering—unlocked capabilities that graphics engines can't touch.
Mercedes Bent (Partner, Premise Ventures) explains why consumer products live or die on day-one retention—and why the YC wisdom to avoid consumer AI might actually be right.
Magnus Müller (CEO, Browser Use) explains why the simple screenshot-to-LLM approach fails in production and what actually separates working browser agents from broken ones.
An engineering decision so specific it reveals how seriously enterprise data companies take permissions.
Jan Liphardt (CEO, OpenMind) explains the coming bifurcation in higher education — generic universities are dying, but brand-name institutions will become more essential than ever.
John Kim (CEO, Paraform) explains why the bottleneck in hiring great talent isn't finding candidates — it's knowing which recruiters have access to them.
Dara Ladjevardian (CEO, Delphi) on the moment the company faced existential threat—and what it teaches about anti-fragility in the AI era.
Alex Reichenbach (CEO, Structify) explains why teaching a business operator to data engineer is now easier than teaching a data engineer the business — and what that means for hiring.
Gil Feig (CTO, Merge) explains the infrastructure decisions that separate AI startups that scale from those that don't.
Daniel Davis, co-creator of TrustGraph, explains why the context graph hype cycle is mostly hollow — and how to spot the difference between real infrastructure and rebranded databases.
Gil Feig (CTO, Merge) explains why most RAG implementations break at scale and what to do about it.
Adam Biddlecombe explains how Mindstream ranked for 'top AI newsletters' and flipped the sponsorship game from cold outreach to 30-40% inbound through one deliberate SEO play.
Showing 12 of 164