My thoughts on NeurIPS 2025
Coming back from NeurIPS 2025, I thought I’d write some of my thoughts about the conference and the current state of AI research.
Given that this was my first time attending the conference (huge mistake for someone who’s working in AI research), I was astonished by the amount of people there, I overheard someone saying that there are around 35,000 people this year, a 60% increase from last year alone. That’s insane (didn’t fact check that)
Thank you
First and most importantly, I want to thank the organizers for managing to pull off such a huge event, the logistics behind it must have been a nightmare, and they did an amazing job. The security, catering, cleaning, and admin staff all managed to give us an excellent experience where we can focus only on learning about AI research while everything else is taken care of. Thank you for making this possible.
What I enjoyed
While I couldn’t attend many talks (due to scheduling conflicts), I really enjoyed the poster sessions; I tried to attend as much as I could and go through most of the posters. The ability to talk directly to the author(s) and ask them questions about their work is something that I really enjoyed. I tried to ask them how I can use their work with what I currently do and did get some interesting ideas that I will try in the next couple of weeks.
Another thing I really enjoyed were the workshops. Even though I only attended a couple of them, I found them very useful; maybe it’s because they’re more focused on a certain topic and they go deeper into it, but I found myself learning a lot more in those workshops than in the main conference talks.
What I didn’t enjoy
NeurIPS is not a cheap conference, in fact I think it’s one of the most expensive ones out there; though totally worth it
San Diego is not a cheap city to stay in, it’s also one of the most expensive cities in the world.
What really annoyed me is the abuse that most hotels did to the attendees. The first hotel I stayed in (shame on you!) was charging 900$ per night for a 12m2 1950s room that has the sink inside the bedroom next to the bed!
This extremely high combined cost makes it really hard for students and researchers from low-income countries or labs with limited budget to come and show their work, and this shows in the conference! Having to pay thousands of dollars just to attend the conference isn’t really better than journals asking for subscriptions to give you access to papers, we should work to make science open for all!
What’s the one idea that really stuck with me
One topic I wanted to learn about for some time is Mechanistic Interpretability, so when I saw a workshop about it, I immediately signed up for it and it really didn’t disappoint.
Dr. Been Kim’s keynote talk about it (Towards a Pareto Frontier of Interpretability: 15 Years of Research in 15 Mins) was eye-opening. I learned a lot about the current state of interpretability research, what works and what doesn’t and the challenges that we still face in this field.
During her presentation, she talked about how we can use mech interpretability to extract concepts from models in a certain domain (chess in that case) and use these extracted concepts to teach experts in that domain about new strategies that the model learned but the experts didn’t know about.
What really caught my interest is that this might be a way to make a much bigger use of AI without actually waiting for AGI or whatever we’re waiting for. AI (LLMs in this case) are exceptionally good at extracting patterns in data that us as humans miss, so if we use this technique ( and if it works well ) to extract these patterns and teach them to humans that are experts in a certain, we can make a lot of progress in various fields without actually needing to have a super intelligent AI that can do it completely for us.
I managed to talk to Dr. Been Kim after her talk and she was kind enough to answer my questions; I wanted to understand what kind of “concepts” these techniques were extracting and whether this can be applied to other data domains (DNA/RNA and Protein LLMs) and she seemed quite positive about it, so that’s something I will experiments with in the next couple of weeks
Final Remarks
Even with some downsides ( how expensive it is ), NeurIPS 2025 was an amazing experience, I learned a lot, met a lot of interesting people and got to see some amazing research. If you are working in that domain or interested in joining it, definitely attend it if you can and make the best out of it. Networking is key in this field and NeurIPS is the perfect place to do it. Most of the people there share similar interests so I found it easy to have a good long conversation with them about their work.
See you next year NeurIPS!