Latent Space: The AI Engineer Podcast

Aman Singh, Head of AI at AI-native wealth firm Arca, explores the machinery behind agentic financial services. He discusses model-agnostic agent harnesses, persistent memory for multi-day planning, human-approved portfolio changes, deterministic safety controls, audit trails, sandboxed financial artifacts, and proactive client insights. A fascinating look at making AI reliable in high-stakes wealth management.

Joseph Kim, OpenAI’s financial services executive, discusses building ChatGPT for finance. He explores enterprise data, firm-specific workflows, plugins, and startup opportunities. The conversation also covers human oversight, AI safety, professional judgment, financial evaluations, and how models may advance from junior analyst tasks toward senior decision-making.

Paul Copplestone, Supabase co-founder and CEO, and Ant Wilson, its co-founder and CTO, explore the infrastructure behind AI-built software. They discuss long-running agents, faster and scalable Postgres, Turso’s database architecture, zero-downtime growth, and Supabase’s multi-cloud ambitions. They also unpack vibe coding, declarative schemas, enterprise database management, open source, and AI’s impact on software development.

Angela Jiang, Anthropic product leader for Claude, and Katelyn Lesse, engineering leader for its platform and agents, unpack the machinery behind Opus 5.5. They explore API scaling, protecting reasoning traces, model launch discipline, safety controls, managed agents, evaluation, memory, and the “brain versus hands” architecture. They also discuss Rust rewrites, long-running workflows, personalization, and avoiding AI-generated slop.

Ian Quigley, a scientist and entrepreneur at Leash Bio, explores AI agents for molecular discovery. He discusses an eval with a full answer key for scientific decision-making, how models navigate DNA-encoded library data, and why stronger systems abandon unproductive molecular regions. The conversation also covers medicinal chemistry as a game, frontier versus open models, molecular representations, Bayesian optimization, and simulated lab environments.

Erik Bernhardsson, Modal co-founder and CEO, previously built Spotify’s recommendation systems and led technology at Better.com. He explores why AI workloads defy rigid GPU contracts, Modal’s strategy for pooling global capacity, and the financial risks of procurement. They also discuss ephemeral sandboxes for agents, alternative accelerators, developer experience, and AI’s next verifiable applications.

Anekha Sokhal, founder and CEO of Moshi, brings experience in mathematics, machine learning, and energy derivatives trading at J.P. Morgan. She discusses turning massive market-data streams into real-time intelligence for commodity traders. Topics include Moshi’s recommendation engine, Bloomberg comparisons, trusted-source filtering, private data, workflow-friendly AI, customer discovery, and scaling demand.

Alex Lieberman, Tenex co-founder, and CJ Hess, a software engineer focused on AI-native workflows, explore how large enterprises can become AI-native. They discuss outcome-based engineering pay, MetaHarness for reliable long-running coding agents, token economics, the gap between AI demos and production, rebuilding SaaS, vertical AI’s margin challenges, and how automation may reshape jobs and New York’s tech scene.

Evan Conrad, CEO of SF Compute, builds markets for tradable GPU capacity. He explores why physically settling compute contracts requires operating the cloud, how workload interruptions and GPU differences shape pricing, and the rise of spot, forward, and lease-breaking markets. They also discuss compute futures, sovereign infrastructure, financialization, and programmable data centers.

Ian Quigley, co-founder and CEO of Leash Bio, explores the messy promise of AI-driven drug discovery. They unpack DNA-encoded libraries, four-billion-measurement benchmarks, models that accidentally learn chemists instead of chemistry, and molecular “activity cliffs.” The conversation also ventures into language-model scientists, oral small-molecule drugs for cytokines, and lab-in-the-loop simulations for scientific reasoning.