Java for AI

Many Java features, existing and future, can meet the demands of AI. Existing features include the Foreign Function and Memory API and the Vector API. Future features include those proposed by Project Valhalla and Project Babylon. This video discusses these features and how they might be used by Java libraries and applications to build competitive AI solutions.

Presented by Paul Sandoz - Java Library Architect (Java Platform Group - Oracle) at JavaOne 2025 (CA, March 2025).

Tags: #Java #JavaOne #AI

SES1650

[thenaughtyanil]

Slightly late but much needed. Java needs to build lots of ml and ai based packages.

[JohnnyDada-qw]

I wish Java was dominant in every IT field.

[RanjanKumar-zzn]

We all know Java is a parent and other languages are child, so I want Java to always be at the top instead of python or JavaScript... So guys you are the guys who can make it possible 💪 so keep it up...
Somewhere I heard Java is going to be Deprecated so hope this will be wrong

[remek]

Would you ever consider making a video about how to deal with pressure during Sprints as a Java developer? Working in Sprints stresses me out, especially Sprint Commitments. I often end up doing unpaid overtime just to deliver everything by the end of the Sprint or to meet deadlines.

[Zakaros]

still, most tutorials for that field are on Python....

[ironickeyboard]

This why the AI rises up and murders us all.

[duxiao]

Compared to other programming languages and their thriving ecosystems, Java is increasingly showing its heaviness/verbosity

[mauriciozapata]

I can't wait for fireship and/or ThePrimeagen to react to this 😂😅

[lhxperimental]

I am using Python for AI because there is no choice at the moment. But python's dependencies are very flaky and difficult to manage once you have a dozen or so dependencies. There are breaking changes in libraries quiet often. One lib wants an older version of a dependency while other wants the latest version of same dependency. It feels like the whole system is geared towards doing one off PoC on a single machine where once your environment is setup just right, you don't touch anything for the fear of breaking. You have things like virtual environments where you essentially create copies of every library including the python runtime! You run out of SSD space very quick. Your code automatically atrophies after a few months and will refuse to run. The concept of reproducibility of builds is apparently not considered important in Python world. Java has a very mature dependency distribution model with signing and immutability baked in. If AI becomes accessible via it will be a blessing for development and operations.

[lattehour]

just do it because Python is really making the case for a mental cognitive disorder, the language is bad and deviates to much from automata and formal grammar but that is not really the biggest problem but the 1000 ways in which something can and IS done, this is fully mental you have to switch neural graphs in your brain from code to code just to be able to read it, worse when i want to write an algorithm i can't decide how to write it shall it be, ,pythonic`` shall i use array based data structures this start to mix in my head and i deviate from purity, hopefully in long term Java will abolish Python

[EliezerBrasilian]

Javafx must be more imperative to. Like kotlin multiplataform

[Xdddd]

We are too little too late. We kept thinking java is faster and better more matured it will always be at top and ignored python. But now python has caught up in modern world. Apart from banking systems or new frameworks nobody cares anymore about java. Automation- Python, Data Scientist and Analytics - Python, Data Engineering - Python, Ai & Ml - Python

[drownwarlord]

do u even understand the current development of AI is leaning toward performance sensitive architecture, serving an LLM using a lowlevel runtime that highlyoptimize for production sometimes still encountered performance issue bcs of how costly the actual model is. And now someone suggesting to use JAVA yea bro no thanks, aint no way I'm running my model with freakin JVM runtime 💀💀💀💀💀💀💀💀

[johnnycoast]

15 years behind python and very complex syntax

[Sven_Dongle]

Performance update: 12 core threadripper pro, 32 gigs, windows 11, A2000 GPU 1b_Q8_0 GraalVM JDK25 I am getting anywhere from 50-100 tokens/sec! I was able to set the heap -Xmx26g and increase context to 128k. this thing just needs space and it runs like a scalded dog, but not under native. JIT is beating native all to heck and gone.

[Sven_Dongle]

The 1b_Q8_0 seems to use 5 threads as a native exe and 25 threads running in the GraalVm under JDK25. In either case the performance seems strangely glacial. The parse is 3.45 seconds and the load is 1.01 seconds so its in line with the demo results, but after that the inference is crawling on intel 8 core i3 under win10.

[Sven_Dongle]

They vastly overstate the performance. I am running it now on an older laptop, granted, but after about an hour of processing the answer to "who are you" is so far "I am an artificial..."