This podcast dives into the evolution of post-training recipes, from InstructGPT to the 2026 multi-teacher on-policy distillation (MOPD) era. Nathan Lambert and Finbarr Timbers reflect on challenges in open-source models like OLMo-3 and analyze how frontier labs leverage specialized teachers and distillation to push performance boundaries.

At Data + AI Summit 2026, Databricks expands Agent Bricks into a comprehensive agent platform, addressing the 99% hidden technical debt (token capacity, deployment, security, evaluation, monitoring, context, sharing) in building agents. The platform offers model choice, context retrieval, and governance control, supporting over 100k agents and processing 1+ quadrillion tokens per year.

Unity AI Gateway is Databricks' enterprise AI governance solution. At Data + AI Summit 2026, they announced major innovations across four areas: optimizing AI usage with cost controls and smart routing, governing AI assets and interactions in one place, monitoring and investigating AI activity, and extending governance through an open ecosystem. The article details unified spend visibility, hard spend caps, smart routing, model and MCP service governance, contextual runtime policies, end-to-end tracing, and security integrations.

At Data + AI Summit 2026, Databricks announced major updates to Unity Catalog, including Unity AI Gateway for governing AI agents, Glossary and Domains for business context, and cross-cloud/cross-region governance. These capabilities address the need for control, context, and choice in the agentic era.

Databricks launched CustomerLake, an Agentic Customer Data Platform natively embedded in the Databricks Lakehouse. It unifies customer data, identity resolution, AI models, and governance to enable always-on, personalized marketing through autonomous agents, replacing fragmented martech workflows.

Databricks introduces Lakebase Search, a hybrid vector and full-text retrieval system built into Lakebase Postgres, now in beta on AWS and Azure. It leverages tiered storage and purpose-built indexes to reduce costs by up to two orders of magnitude, enabling billion-scale vector search with 32x compression via RaBitQ. Designed for AI agent workloads, it provides a single backend for memory and context, eliminating the need to stitch together separate databases.

At Data+AI Summit 2026, Azure Databricks announced a wave of new capabilities across four pillars: Agentic Data (first true LTAP architecture, serverless Postgres Lakebase, millisecond-response Lakehouse//RT), Agentic Dev & Work (Genie integration with Microsoft Teams and M365 Copilot, Excel Add-in, SharePoint Connector), Agentic Marketing (CustomerLake agentic CDP), and a governance framework powered by Genie Ontology and Unity AI Gateway. These innovations aim to transition enterprises from experimental AI pilots to production-grade automated workflows, enabling seamless operation of data, teams, and autonomous agents natively on Azure.

Databricks launches Genie One (data-smart AI coworker), Genie Agents (autonomous agents), and Genie Ontology (context layer) to address enterprise challenges with scattered context, slow reasoning, and high costs. Genie One integrates Slack, Teams, mobile apps, and more to move from insight to action; Genie Agents let users create autonomous agents via simple prompts; Genie Ontology automatically builds a knowledge graph for better accuracy and performance.

Databricks introduces Lakeflow, a unified platform integrating ingestion, transformation, and orchestration under Unity Catalog, enabling AI agents to build and operate data pipelines. New features include Genie Code integration, Lakeflow Designer for visual development, Genie ZeroOps for automated operations, 100+ connectors, Zerobus high-performance ingestion, and real-time mode for Spark Declarative Pipelines.

Databricks introduces Genie ZeroOps, an autonomous background agent built into the platform that monitors, investigates, and proposes fixes for data and AI assets like pipelines, jobs, tables, and ML models. It leverages full observability, data lineage, and sandbox environments to securely verify fixes, aiming to free data teams from maintenance burdens.