Nuclear Fusion, No Power Lines ft Jonathan Frankle

Most organizations treat a bigger context window like a cheat code: dump every document in, skip the data work, ship. Jonathan Frankle, Chief AI Scientist at Databricks, says that's still wrong.This is Jonathan's return visit to Invisible Machines — a conversation recorded last summer, released ahead of Databricks Data + AI Summit. His first appearance (season 2) was the MosaicML-era craft conversation: lottery tickets, mixology, mini-cupcakes. This one is the enterprise engineering thread: be a scientist, curate before you scale, and treat specification (what you actually want the system to do) as the bottleneck between raw model power and useful AI.Robb and Josh press him on the myths that still seduce enterprise teams: million-token windows as a substitute for real data work, hyperscaler résumés as a proxy for talent, and the fantasy that unlocking every PDF in the org automatically makes knowledge useful. Jonathan's answer is consistent: measure success, test your use case, climb the ladder of techniques, and accept that multimodal is where long context actually earns its keep, not as a universal bypass for curation.Along the way: the nuclear fusion vs. power lines metaphor; why building a benchmark is a cop-out compared to describing intent; prompts as parameters; chat-only UIs vs. a generation that never wanted buttons; LLM-oriented publishing and static FAQ pages; unlocking PDF at scale when curation gets skipped; early-adopter mistakes we'll laugh at in ten years; and why separating knowledge from reasoning is the north star, even if we aren't there yet.---------- Support our show by supporting our sponsors!This episode is supported by OneReach.aiForged over a decade of R&D  and proven in 10,000+ deployments, OneReach.ai’s GSX is the first complete AI agent runtime environment (circa 2019) — a hardened AI agent architecture for enterprise control and scale. Backed by UC Berkeley, recognized by Gartner, and trusted across highly regulated industries, including healthcare, finance, government and telecommunications.A complete system for accelerating AI adoption — design, train, test, deploy, monitor, and orchestrate AI agents.Use any AI modelsBuild and deploy intelligent agents fastCreate guardrails for organizational alignmentEnterprise-grade security and governanceGet in touch: https://onereach.ai/contact/?utm_source=youtube&utm_medium=social&utm_campaign=s7e11&utm_content=1 for SoundCloud:https://onereach.ai/contact/?utm_source=soundcloud&utm_medium=social&utm_campaign=s7e11&utm_content=1 ---------- The revised and significantly updated second edition of our bestselling book about succeeding with AI agents, Age of Invisible Machines, is available everywhere: Amazon — https://bit.ly/4hwX0a5#InvisibleMachines #Podcast #TechPodcast#AIPodcast#AI #AgenticAI#EnterpriseAI #Databricks#RAG#MachineLearning#DataEngineering#EnterpriseEngineering#AIStrategy#AIEngineering0:00 Jonathan Frankle Returns | Databricks Chief AI Scientist · Invisible Machines1:47 We Remember the Plants | Returning Guest Jonathan Frankle2:22 Million-Token Context Windows: Do You Still Need to Train LLMs?3:40 Be a Scientist | Measure AI Success Before You Scale5:54 Hyperscaler Résumés Are Not Proof of AI Expertise10:01 Maximize Impact | MosaicML, Databricks & Enterprise AI13:02 Lottery Ticket Hypothesis vs. Real-World AI Impact14:12 Nuclear Fusion but No Power Lines | Jonathan Frankle16:08 AI Specification & Evals: Why "Build a Benchmark" Is a Cop-Out17:59 The Smoothie Problem | From Model Power to Useful AI18:53 Prompts as Parameters | Fine-Tuning Without Model Weights22:46 It's Computing | Specification, Testing & Agent Design24:44 LLM SEO, PDFs & Enterprise Data for AI Ingestion27:35 Static FAQs, Curation & LLM-Oriented Publishing30:26 Unlocking PDFs Scales Your Mistakes | Enterprise RAG33:25 Knowledge vs. Reasoning | Brand Control in AI Search34:50 Thanks for Listening | Invisible Machines