I'm not going to transcribe my notes exhaustively. The JuliaCon 2025 website has lots of information, which you should definitely check out, but I'm going to focus on my own observations, reactions, and thoughts. My schedule didn't allow for the workshops or the hackathon (maybe next year).
I had never attended JuliaCon before. I had reviewed some of the videos, but the in-person experience is substantially different than watching videos on the laptop. The venue was great, and everything was walkable. My car didn't move between arriving and departing.
My biggest challenge was choosing talks. On Thursday, there were seven rooms in use simultaneously. My background isn't a good fit for any of the mini-symposia, so there were several times I had to move quickly from location to location to hear what I wanted.
With so much going on, I was surprised by how many people were in the hallway on their laptops. Everyone was building something. After three days, I wanted to be building something, too. That was the most significant takeaway from my time there: just start coding.
One motivation for attending the conference was to be a greater participant in the Julia community. I've been considering helping out with a few packages I find useful, so I made a point to listen to a talk about open-source contributions. Contributions fall into three categories:
No code
Low code
High code
Blogging falls into the first category. It's important to draw attention to how people are actually using Julia and what their experience is like. Other examples are documentation and tutorials.
Testing is the best example of a low code contribution. Although this would include more than just unit tests, I encounter unit tests far more frequently than any other kind.
I have found LLMs to be useful for both of these categories. My experience is that coding agents are able to create READMEs, docstrings, and test cases quite easily. Because multiple dispatch in Julia allows packages to be combined in unexpected ways, adequate documentation directly affects the usefulness of a package. As a byproduct, documentation and test cases make code easier for LLMs to pick up and incorporate in future code.
At the State of Julia talk, I learned that LLM usage needs to be disclosed for any contributions to the Julia repositories. It seems sensible while we all try to figure out how to be more effective with agents in the loop.
I would highly recommend watching An AI Agenda to Modernize Healthcare Operations. Do that and come back.
The most impressive part of the talk was the description of UCSD mission control. While it's cool that they can use data for a utilization forecast on T+4, what's genuinely amazing is that they've empowered the CMO to act on that forecast.
I found the concept of notecasting to be an interesting use of LLMs. I would have expected LLMs to be useful in extracting relevant data from the patient notes. Structuring this data and combining it with output from patient monitors should be helpful, right? Instead, the metrics were inserted into the text so that one model could do all the necessary predictions. It's an interesting approach, and one I plan on trying in a different domain.
We also got to hear about how Temple Capital uses Julia for research. I think it's great to see more Julia in finance. My sense is that the language has reached a production-capable state.
There were a number of talks about fundamental parts of the language. There is progress in improving debugging by augmenting what happens between source code and machine code. I enjoyed the discussion about the type system.One of the more interesting talks was discussing how Julia could handle interruptions better. There were also suggestions about improving Julia's scheduler.
I'm glossing over a lot of the detail here. The general theme is that we're seeing Julia implementations inspired by what other programming languages do well. There's clearly a lot of thought about getting these details right.
Dyad was a major focus of the conference. The first keynote discussed Digital Twins.
I have zero expertise in this area, but it became immediately obvious why people are focused on compilation and compact binaries. IoT is relevant to digital twin models, and there's a desire to use Julia for every part of the problem. Julia seems to be a good choice for eliminating the two-language problem in this space.
During the conference, I heard about many things worth a glance:
The Innovation Factory by Prith Banerjee
I've discussed only a fraction of what the conference had to offer. Visit the conference website for more.