AI Podcasts for Developers

Lex Fridman talks with military analysts Preston Stewart and Michael Kofman about the state of the Russia-Ukraine war, including drone warfare and the kill zone, the Battle of Bakhmut, Prigozhin's mutiny, and paths to peace. Mostly geopolitical and military analysis rather than AI-focused.

A deep dive into the Fable 5.2 report and METR's recursive self-improvement report, plus hands-on testing of Gemini Carbon. Covers how fast AI is accelerating AI development, lesser-discussed benchmarks, Simple Bench saturation, and remote labour automation.

Why this matters: Evidence on AI accelerating its own development and on benchmark saturation affects how developers and teams plan around fast-changing model capabilities.

NLW reviews early reactions to Anthropic's Fable 5.2, which is topping benchmarks and powering ambitious demos, alongside his own experiments. He compares it with Opus 5.5 and explains why it isn't yet an obvious everyday upgrade.

Why this matters: It shows that benchmark wins don't automatically translate into better day-to-day workflows, so teams may need to rethink which tasks they hand to a new frontier model.

Zeno Rocha, founder and CEO of Resend, talks with swyx about designing products for AI coding agents, including how agents now choose developer tools like Resend on their own and what answer engine optimization (AEO) looks like for a dev-tools company.

Why this matters: As coding agents increasingly pick libraries and services in pull requests, developer-tool makers need to optimize docs, APIs, and onboarding for agents as well as humans.

Goodfire's Curt Tigges explains research showing individual MLP neurons in modern gated-MLP language models are more interpretable and functionally bounded than assumed, tested with an automated agent across architectures like Qwen3.6-35B-A3B. They cover dynamic sparsity without lasting benchmark loss, enumerable representations, and implications for behavior monitoring and persona steering.

Why this matters: If neurons can be directly enumerated and interpreted, developers may monitor and steer model behavior without relying on complex external featurizers like sparse autoencoders.

Luisa Rodriguez examines why China's best AI models trail the US by only about six months despite less compute, pointing to rented and smuggled American chips, large-scale model distillation, and weak security around frontier models. She argues the US could close these gaps with existing powers.

Why this matters: It challenges the 'we must race China' argument against slowing down, with implications for export controls, model security, and the time available to test frontier models for safety.

A daily AI news rundown covering Anthropic's Claude Fable 5.2 and Mythos 5.2 launch (30% faster, $8/$40 per million tokens, 49.1 on Frontier Code 2.0) alongside an NVIDIA board investigation into Jensen Huang, DeepSeek's agent app growth, and Amazon's humanoid robotics push.

Why this matters: Faster, cheaper Fable pricing and new coding benchmark numbers directly affect model selection and cost planning for developers building on Claude.

In this episode, Stefano Corazza, head of AI research at Canva, joins us to explore how his team is building custom models and agents around iterative design workflows. We discuss the tradeoffs between training models from scratch and adapting open weights models, and how proprietary design data supports specialization. Stefano explains how Craft separates layout planning from pixel generation and how Brush uses design tools and visual feedback to create and revise compositions. We examine the challenges of evaluating design quality beyond a single generation, including the use of vision language models as judges alongside human evaluation. We also discuss lessons from simplifying agent harnesses, why faster inference enables more iterations, and how model ownership creates opportunities to optimize serving costs. Finally, we explore how APIs and MCP are expanding access to these capabilities and what changes when software companies build tools for agents as well as people. 🗒️ Full show notes: ⁠⁠

a16z's Joe Schmidt talks with HubSpot co-founder and CTO Dharmesh Shah about where startups can build as frontier AI labs move into applications. They cover starting in narrow markets, owning workflows and customer context, HubSpot's AI experiments, and one-person software businesses.

Why this matters: As models become infrastructure, it clarifies where application-layer startups can still build defensible value despite competition from AI labs.

SimEX is a framework where coding agents iteratively probe and optimize robot skills in simulation, then correct the simulator and toolbox with only about 10 minutes of real-robot trials. It transfers to real tasks like towel folding, barcode scanning, and plate manipulation without demonstrations.

Why this matters: It shows a practical way to bring trial-and-error coding agents to physical robots while avoiding the cost and safety risks of real-world iteration.

Revolut founder and CEO Nik Storonsky explains how the fintech scales to 80 million customers, uses a strict retention-curve "new bets" process to kill or grow products, and builds profitability around subscriptions rather than lending. He also covers using AI to cut 210-day compliance approval cycles and Revolut's push against Amex and Chase in the US.

Why this matters: It shows a concrete example of AI being applied to slow compliance workflows and of retention-driven product discipline at large scale.

Covers the reported scandal around Nvidia CEO Jensen Huang and Megaspeed executive Alice Huang, alleged by US intelligence to be a Chinese MSS recruit, and its link to a $2B GPU deal and AI chip export controls. Also explores the web of companies buying Nvidia chips near China and the four fronts of the US-China AI race: chips, power, people and models.

Why this matters: Export control policy and alleged influence over it could reshape GPU supply and compute access for AI developers and companies worldwide.

Allie Howe of Keycard joins the Mastra hosts to discuss how headless apps and agents expose security gaps that UIs used to mask, covering agent identity, rate limits, prompt injection via WebMCP, approvals without a UI, and consent fatigue. They also debate MCP versus REST as an agent front door and explain ID-JAG for Enterprise-Managed Authorization for MCP.

Why this matters: As agents call your APIs directly, developers can no longer rely on UI friction for security and must design agent identity, authorization, approvals, and rate limiting explicitly.

GitLab VP of Engineering Tim Zallmann discusses bringing AI agents to the full software development lifecycle, with enterprise governance, auditing, and agent identity. They also cover air-gapped deployments and how new entrants in code hosting are changing the space.

Why this matters: Teams adopting coding agents need shared context across tools plus governance, auditability, and human attribution for agent actions to deploy them safely in enterprises.

Paul Roetzer and Mike Kaput cover voter backlash against data centers in 40 of 41 local ballot measures, METR's finding that AI roughly doubled research pace at Anthropic, and Nvidia's probe into an alleged spy operation targeting Jensen Huang. They also touch on AI in education, enterprise AI spending reaching $177B, and controlling AI costs.

Why this matters: Local limits on data centers could constrain compute buildout and push up AI costs, while signs of AI accelerating lab research affect expectations for model progress.

Ed Elson and guests discuss reported allegations that Nvidia CEO Jensen Huang had an affair with an alleged Chinese spy and how Washington might respond. The episode also covers post-election market volatility and the growing backlash against data centers.

Why this matters: Leadership scandal at Nvidia and data center backlash could affect AI chip policy, export scrutiny, and infrastructure buildout that developers depend on.

Claire Vo shares hands-on impressions of Anthropic's Fable 5.2 after two days of building with it, including an animated video from a tweet, a furniture-assembly vision test, a game with a trailer, and a real pipeline re-architecture. She also covers when to route tasks to Fable 5.2 versus Opus 5.5.

Why this matters: Practical model-routing guidance and real engineering results help developers decide where a new frontier model fits in their Claude Code workflows.

NLW covers Jensen Huang's reported admission of an affair with a suspected Chinese spy, Nvidia's internal investigation, and the possible fallout for AI chip exports and US-China relations. Headlines include Anthropic's IPO plans, OpenAI's funding shortfall, and Microsoft's Oracle pullback.

Why this matters: Leadership and intelligence concerns at Nvidia could affect AI chip export policy and GPU supply, which developers and AI companies depend on.

A short episode on packaging AI consulting offers, finding clients, and pricing services based on the value delivered. It is business-oriented guidance with hypothetical pricing examples rather than technical content.

Alex Kantrowitz discusses a reported story that Nvidia CEO Jensen Huang admitted to an affair with a woman U.S. intelligence believes is a Chinese spy, prompting a board investigation into whether Nvidia or U.S. secrets were compromised.

Why this matters: Nvidia underpins the AI hardware buildout, so leadership and security concerns could affect chip supply, export policy, and the AI industry.

Latent Space's first deep dive with an AI data company, featuring Handshake AI's Garrett Lord and team on how Handshake built an AI training data business near $1B in annualized gross revenue in about a year. They discuss what frontier labs buy, including reinforcement learning data and expert-driven training work.

Why this matters: It offers a rare look at the data supply chain behind frontier models, showing where labs spend money and what kinds of training data (RL environments, expert annotation) drive model capability.

Aytekin Tank and Demetri Panici show how Claude can generate motion graphics and training videos, with examples from Jotform workflows. They cover reusable skills, brand consistency, product launch and help videos, and where AI video tools may go next.

Why this matters: Reusable Claude skills could make it practical for product teams to produce launch and help videos for every feature without a dedicated video workflow.

Ghita Houir Alami of ZeroEntropy (now Model Research at Notion) explains how to train calibrated rerankers, including the zELO method that uses pairwise comparisons and Elo scores, then distills them into efficient pointwise models. The talk also covers inference infrastructure, multi-stage retrieval, and how reranking fits into agentic search.

Why this matters: Calibrated, instruction-following rerankers enable dynamic top-K retrieval and cut expensive LLM token usage in agentic search and RAG pipelines.

Federico and John start a miniseries on their favorite apps for the iPhone Duo, covering productivity and quirky utilities like DuoBox, Hinge Caliper, T3 Code, and Fineliner. The AppStories+ bonus segment covers Apple's tools and skills for Siri AI.

Why this matters: It shows what apps are emerging on a new Apple hardware form factor, and points to how Siri AI tools and skills may affect app developers.

William Imoh shows how he used Codex and GPT-Live to build Mural, a native iOS and Android real-time voice app for learning Norwegian, despite never having written Swift. He also walks through running a ~25-person team as a solo founder using a Lord of the Rings-themed fellowship of Grok Bot agents.

Why this matters: It shows how a non-specialist can ship native mobile apps with coding agents and operate a company with agent roles, including the cost and latency tradeoffs of realtime voice.

Limitless examines OpenAI's competitive position as Anthropic heads to IPO, covering the GPT-6.2 Astra rumors, ChatGPT's user growth and search share, key Anthropic hires, a trimmed funding round and valuation, compute constraints, GPT-6 Cyber, and ChatGPT ads.

Why this matters: Shifts in frontier lab momentum, talent, compute supply, and monetization affect which model platforms and agent products developers can rely on.

Scott Kupor, a16z co-founder turned Director of the U.S. Office of Personnel Management, tells David Senra how a16z was built on the CAA model and how he is applying talent and incentive lessons to reform the federal workforce, including performance systems and paper-based retirement processing.

Scott Galloway and Ed Elson discuss the market's whiplash reaction to the election and what a shift in Congressional power could mean for the economy. They also touch on the backlash against data centers and the need for a clear growth agenda.

Why this matters: Growing political and public backlash against data centers could affect the infrastructure buildout that AI companies and developers depend on.

Mihail Eric of Monaco discusses Stanford's CS146S, a course on AI-native engineering where coding agents can solve traditional homework in one prompt. In its second run, students work with open-source maintainers to learn unfamiliar repos and propose real changes.

Why this matters: It shows how software engineering education is being redesigned around coding agents, emphasizing code comprehension and real open-source contribution over rote implementation.

Peter Yang talks with OpenAI's VP of Product for health, Ashley Alexander, about using ChatGPT to help their families through cancer, from preparing for doctor visits to researching treatments and handling health records. They also cover trust and privacy, doctors' reactions to patient AI research, and why hospital data systems still rely on faxes and CDs.

Why this matters: It offers a first-hand look at how OpenAI's health product lead sees real-world AI health use cases, the trust and privacy concerns, and the healthcare interoperability gaps developers could build around.

Google's head of product for frontier models, Tulsee Doshi, explains how PMs work inside a frontier AI lab, from writing a first eval to her 60-20-20 framework for building ahead of model capabilities. She also covers Google's full-stack AI strategy, AI harnesses, and avoiding AI slop.

Why this matters: It offers a concrete look at how evals, model UX, and product adaptability shape building products on rapidly improving models.

NLW explains how AI agents can now edit, animate, and produce video from plain-language instructions. He covers use cases like meeting highlights, concept demos, animated business updates, and self-updating customer education videos, plus three levels for getting started.

Why this matters: Agent-driven video workflows turn video production into something that can be scripted and automated, opening new product and internal-tooling opportunities.

Truemed CEO Justin Mares tells a16z's Daisy Wolf which health interventions are worth the effort: food, movement, sleep, and fewer harmful exposures. They also cover Truemed's effort to open HSA/FSA dollars to preventive lifestyle spending.

Nathan Labenz and Prakash Narayanan discuss signs of accelerating AI progress, citing METR's assessment and Anthropic's amended regulatory filings. Guests cover agents trading on midterm election forecasts, secondary GPU markets and compliance, and AI-native venture investing that runs thousands of agent runs.

Why this matters: It shows how autonomous agents are already being used in forecasting, compute markets, and investing, which raises practical questions about oversight and governance as capabilities accelerate.

Paul Graham interviews Replit CEO Amjad Masad about how Replit anticipated the AI coding era and shipped an agent that handles errors and provisions infrastructure on its own. They also cover what follows vibe coding and the vision of a "self-driving company" where agents build and operate software.

Why this matters: It shows how coding agents are moving into large enterprises, with claimed cost drops from $10M software projects to $250K, which affects how developers and startups scope and build software.

Dr. Mark Hyman talks with a16z partner Daisy Wolf about practical longevity habits like strength training, protein, sleep, stress management, and testing. They also cover Function Health's approach of combining blood tests, imaging, wearables, and eventually AI to shift healthcare toward earlier prevention.

Ed Elson examines why enthusiasm and easy capital in the AI trade are starting to fade, drawing on market and investor sentiment. It's a short, finance-oriented take rather than a technical discussion.

NLW walks through Menlo Ventures' enterprise AI report showing companies spending nearly 5x more on AI than last year, driven by agents, with Anthropic and OpenAI capturing most of the market. Also covers the shift from seat-based to token-based pricing, a market selloff, Anthropic IPO uncertainty, and METR's report on AI accelerating AI research.

Why this matters: The move from per-seat to token-based spend changes how teams budget for, price, and build AI-powered products and agent workloads.

Alex Kantrowitz and Ranjan Roy discuss the week's tech news: midterm impacts on AI, market volatility around AI stocks, datacenter backlash and its effect on training and safety tradeoffs, Anthropic's IPO prospects, and whether Anthropic is nearing recursive self-improvement. They also compare polls and prediction markets.

Why this matters: Datacenter restrictions and shaky AI market sentiment could affect compute availability, lab strategy, and funding for AI developers building on frontier models.

A 14-minute AI news roundup on the shift from chatbots to agents that complete tasks, covering Perplexity Teammates, Meta Muse, Airbnb's planned AI travel assistant, OpenAI's GPT-6 Cyber, and Thinking Machines' Inkling. It also touches on Google video generation, Anthropic protein research, AI regulation, and data-center costs.

Why this matters: It tracks how agent products with computer control and automation are becoming the main battleground for AI tooling and platform competition.

Delta Intelligence presents Δ₀, a humanoid whole-body foundation model that controls a 69-DoF Unitree G1 with dexterous hands for long-horizon loco-manipulation. The founders explain pretraining on whole-body human data, a delta action space for seamless human intervention, and co-designing the policy and controller.

Why this matters: Joint design of policy and whole-body controller, plus a delta action space enabling live corrections, offers a practical blueprint for training generalist humanoid policies.

A short episode on using AI-generated music to grow streaming revenue on YouTube and Spotify, including utility music and evergreen albums. It focuses on monetization tactics rather than technical implementation.

A 2-minute announcement of a curated 10-episode feed from 80,000 Hours covering AGI, superintelligence, recursive self-improvement, AI deception, geopolitics, bio risk, and AI consciousness. It also points to career-pivot resources for people moving into AI work.

Turing Award winner Andrew Barto discusses why reinforcement learning rewards go wrong, from Norbert Wiener's early warnings to dopamine parallels and agents escaping sandboxes. He also critiques the limits of reducing goals to a single scalar reward and argues for engineering discipline in AI.

Why this matters: Reward design underlies RL-trained reasoning models and agents, so misspecified rewards directly affect how safely and reliably they behave.

Latent Space talks with Fabian Theis of Helmholtz Munich, a leading figure in Virtual Cell efforts, about algorithms for modeling biology across scales and modalities. The conversation looks at how ML methods can predict and eventually control cellular behavior.

Why this matters: Virtual cell models show how foundation-model techniques are extending into biology, opening new opportunities for ML engineers working on scientific tooling and data infrastructure.

Limitless covers a sharp selloff that erased $2 trillion in 40 minutes and hit AI stocks hardest, arguing the market is wrong because AI demand keeps rising. Cited evidence includes GPU rental prices up as much as 106% since July, token prices down 58%, and a coming CPU boom driven by AI agents.

Why this matters: It links AI infrastructure demand signals like GPU rental pricing, falling token costs and agent-driven CPU needs to market sentiment, which affects compute availability and costs for builders.

Salesforce's John Somorjai talks with a16z's Alex Immerman about nearly $100B in acquisitions, including Slack, MuleSoft and AI-native startups. They cover build-versus-buy, outcome-based pricing replacing seats, what kills deals in diligence, and founder considerations when selling.

Why this matters: As AI makes software easier to build, founders and engineering leaders need to understand what makes a startup acquirable and how enterprise pricing is shifting toward outcomes.

Ed Elson and Anthony Scaramucci discuss the midterm election results, possible shifts in Trump's policies, early market reactions, potential White House shakeups, and what it means for the Republican Party and MAGA movement.

IBM's Mixture of Experts panel discusses how close AI is to self-improvement, starting with an MIT technique that uses AI to modify another model's weights. They also cover Mistral's open model release and Google's Genie 4 world model.

Why this matters: AI-driven weight modification and new open models point to faster model iteration cycles and more options for developers building on open weights.

Ed Zitron's short monologue argues that the Blue Wave election outcome is a bad sign for the AI trade and calls for building a better future by fighting for everyone's voice. It's an opinionated commentary on AI's political and market outlook rather than a technical discussion.