See the live blog table of contents
Chad Armor
- Celebrated Java 26 release date
- Preview of a documentary about the creation of Java coming out in the summer
- Java 26 goals – Data oriented programming, Java in the small (scriptable/learnable), Java at global scale, Integrity by default, AI at Java scale
George Saab
- Showed JEPs in Java 26
- For new learners java playground (now supports java 26), oracle vs code extension (can share snippets of code, notebook support)
- Oracle Java Verified Portfolio (JVP). For paid customers including OCI customers includes private support for Java FX, VS Code Java extension, Helidon
- Support for Java 8 ends March 2028, Java 17 ends September 2028, Java 25 through September 2030
- Project proposal Detroit: combine JavaScript/Python snippets with Java. Can call libraries not available in Java. Was on hold becuase was waiting on Panama project
- Java and AI – readability/compatabiliy, clean semantic model/static tiyping, Java’s specs/docs/JEPs, JVM performance/tools
Ana Maria Mihalceanu and Lize Raes
- Didn’t take notes. It was short
Uber (didn’t catch name)
- Michelangelo – Uber’s unified E2E ML platform. 20K models trained per month 5.3K models in production, 40M peak predications per second. Suppported by less than 10 engineers. Java 8 to 11 reduced CPU by 15% and Java 11 to 21 another 11% reduction.
- GPUs to scale – 10-100x traffic amplification
NVIDIA (Ikroop Dhillon)
- Partnering with Java architects at Oracle for many years about Panama
- NVIDIA conference in San Jose. Yesterday talked about AI including partnering with OpenClaw to make it more secure.
- Worlds AI runs on NVIDIA
- Java historically focused on CPU based development. Environment more complex due to acceleration
- Unstructured data is context of AI
- NVIDIA cuVS Java/Lucene powers Java Vector Search ecosystem. Went from hours to minutes.
- NVIDIA NIM – containerized/packaged model
Lize Raes
- Listed Java frameworks for AI
- Rod Johnson is sick – created Embabel.
Spring (Josh Long)
- Used Spring Initiallzr
- Noted Java 25 was latest as of time of the recording
- Used this java script.java being the first good java script joke
- Showed his AI example very quickly. [glad i had seen this before]
- Showed JVM fastest in benchmark against other langauges
- Also covered Rod Johnson’s part. It was a subset of those at his DevNexus keynote so didn’t take notes again. See https://www.selikoff.net/2026/03/05/devnexus-2026-its-up-to-java-developers-to-fix-enterprise-ai/ for those notes
Ana Maria Mihalceanu and Lize Raes
- Example of a help desk system
- more to comeLook at what AI can help with – ex: add context to ticket, classification, code fixes, flagging urgency
- vs humans for clarification, compliance gatekeeping, accountability, exception handling
- AI triage classifies ticket using similar tickets and company RAG
- AI coding assistant proposes fix
- Important for Enterprise AI: human must approve pull requests and gets final approval on integrating into workflow. Also, must be easily revertable
Ana Maria Mihalceanu
- Typed contracts
- AI components behave like standard services
- Existing security/permissions
- Lots of libraries for AI
Paul Sandoz (Java Libraries Architect)
- “Java is *often* where AI needs to be”
- “Java is *almost* everywhere AI needs to be”
- Java in a good positionf or AI
- Need more than AI. ex: code that converts to text to tkens, connect models to info sources (vector/relational databases), manage interactions among models
- JEPs for performance improvement. Sometimes behind the sense like aliasing venctors to improve speed.
- Most models are Python wrappers around C/C++ code
- Pain points become solutions/jeps/new features.
- Pain point: Using GPUs is too hard
- Pain point; Developing maching learning models is difficult
- Solutions: foundational building blocks in JDK, build libraries with (ex: GPU support), for building apps with (Java code running on GPU)
- JEP-500: prepare for final to mean final (vs reflection)
- Project Babylon – need to translate Java code into other languages
- HAT (Heterogeneous Accelerator Toolkit) – develop Java code that represents GPU code so can run/debug on JVM. Part of Babylon
- Project Detroit – tried in past but never got off the ground. Still strong interest in Java/JavaScript integration. And now Python interest as well. Will use widely used implementations for integration rather than integrating from scratch (v8 and CPython). Uses Panama
- Project Panama is to foreign libraries as Project Detorit is to foreigtn language runtimes
- Goal: java is everywhere AI needs to be
Microsoft (Patrick Chanezon, Brian Benz)
- AI transformations stats – decrase median code review turnaround, faster test automation, etc
- AI as a superpower: https://www.oreilly.com/radar/the-end-of-programming-as-we-know-it/
- GitHub Copilot – autocomplete
- Then added chat
- Then added agent mode – background
- AI agent – loop with LLM in middle
- Need to build and become productive teammates as become manager of agents
- Microsoft Foundry – has models available. Mix of open source and closed. Processed 100T tokens quarterly
- Microsoft IQ – sits on top of Office
- Copilot now has an SDK https://github.com/github/copilot-sdk-java
- Showed modernizing a Java 5/Struts app with Copilot. Lets choose from a list of models
- Agenta at different points in lifecycle – ex: SRE
- More stats: 93% of globabl engineering team uses Copolit. and copilot is #1 contributor to copilot.
My take
Good range of topics for an opener including a good mix of Oracle and external companies