Episodes
-
⚡️ The 2026 Software Factory Standard — Zach Lloyd, Warp.dev
A software factory should be a concrete, declaratively defined system with repositories, skills, MCPs, workflows, and orchestration encoded rather than a loose collection of agent automations.
-
Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
Legora’s product evolved from Swedish legal search into an agent that drafts, edits Word files, analyzes documents, organizes legal work, and coordinates lawyers, clients, and other agents.
-
Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Arca uses AI to replace the operational work surrounding financial advisors, not the advisors themselves; clients are intended to experience better human-led service rather than direct interaction with an AI system.
-
⚡ Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services
ChatGPT for Financial Services combines ChatGPT capabilities with firm-specific templates and licensed financial sources; Morgan Stanley and Evercore were design partners, and customers have especially valued the bundled data.
-
Why Agents Love Supabase
Supabase Compute is designed to remove Edge Functions’ roughly two-second limit by providing long-running, container-like serverless workloads colocated with the database, including Rust, Python, and agent execution.
-
How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands
Anthropic’s reported efficiency improved sharply: revenue per employee rose from $1.4 million to $6.6 million in a year, while the company moved from losing money on each revenue dollar to generating roughly $2 for every dollar of cost.
-
⚡ An Eval for AI Scientists That Knows the Answer Key - Ian Quigley, Leash Bio
Leash Bio’s measured DNA-encoded-library data can serve as a virtual experimental environment in which agents receive immediate feedback while optimizing molecular binding.
-
Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO
Modal began as a general-purpose way to execute code in the cloud, but Stable Diffusion revealed that AI inference and custom model workflows were its strongest early use case.
-
The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi
Moshi is designed to give commodity, energy, and macro traders real-time insights by combining market time series, fundamentals, news, alternative data, and users’ private information.
-
AI for the Fortune 2000: How Tenex Builds an AI-Native SDLC — Alex Lieberman & CJ Hess
Tenex pays engineers primarily for agreed output, measured in story points, rather than hours; the company accepts that this creates substantially more variance in earnings, with roughly a 4X gap between its highest- and lowest-paid engineers.
-
The Programmable Data Center — Evan Conrad, SF Compute
SF Compute evolved from trying to sublease one cluster into operating the full cloud infrastructure needed to deliver GPU capacity reliably.
-
🔬 Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
DNA-encoded libraries let Leash test millions of pooled small molecules by attaching DNA barcodes, pulling out protein binders, and reading the surviving barcodes through sequencing.
-
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Dyna-2 is designed to predict future world states as well as robot actions, using video, joint-state information, and language instructions to produce more dexterous and generalizable behavior.
-
🔬 Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub
Scaling compute or data helps only after researchers find a problem-specific scaling relationship and ensure that the data contains the information needed for the desired capability.
-
🔬 Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
Periodic’s central thesis is that intelligence alone cannot advance science: conjectures must be tested through repeated interaction with physical reality.
-
⚡️ Which GPU Clouds Are Actually Good? | ClusterMAX 3.0
ClusterMAX has expanded from roughly 24 tested providers and about 120 tracked providers in its first edition to 323 tracked providers, with substantially more automated testing for performance, reliability, security, and burn-in.
-
Academia is for Ambition — Alex Zhang, MIT
GPU Mode was created as a learning-oriented community for GPU-kernel programming, and its Popcorn leaderboard grew from the idea that kernel optimization has a small enough problem space to be scaled with data and automation.
-
Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week
Ari argues that computer use has advanced sharply: current agents are much better at debugging and retrying, can combine screenshots with accessibility and DOM data, and can write JavaScript to perform multiple computer actions at once.
-
Claude Code’s Next Era — Thariq Shihipar, Anthropic
The main challenge in Claude Code has shifted from proving that agents can code to teaching users how to specify goals, uncover unknowns, supervise execution, and use compute efficiently.
-
OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha
OpenRouter’s core thesis is that AI consumption is continuous and dynamic, so developers need a neutral PubSub-like marketplace that lets them discover, subscribe to, and switch among models and providers.
Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account