ICLR 2025
I got to attend ICLR 2025 and spend a week in Singapore with my Answer.AI colleagues. This post is some disjointed thoughts on the whole experience.
Safety + Interpretability are in the air
I’m trying to think what trends jumped out at me. ‘Agents’ of course, with some interesting stuff and lots of… not haha. The other main trend I noticed was safety/interpretability stuff, with a sense of ‘this is something academics can work on + contribute to without tons of compute’. There was a keynote from an OpenAI person making the case for why it’s important + good to work on even if you don’t think AI will soon kill everyone. My friend Stanislav Fort had a fun paper on ‘scaling laws for adversarial attacks’ and as always tons of other neat ideas in progress. Various papers trying to make AIs say naughty things or not say them. And in the airport I met Yoshua Bengio who’d been working very hard trying to convince people it’s a bad idea to be chasing creating smart things with agency. I admire his efforts, and tend to agree that making ‘tool AI’ vs ‘agent AI’ is a good idea! And I liked the way he phrased a take on doomish stuff: “The correct response to [someone asking me about AI2027] is to say that lots of experts agree we can’t casually dismiss it” - in other words, you may think specific takes are wrong or that fears are overblown but ‘nobody takes that seriously’ is not a valid argument. All that said, while I noticed this safety talk around it wasn’t like it was the bulk of the discussion and I’d wager most attendees aren’t thinking much about it.
Vibe-Researching is fun
One fun trick I could do at this conference thanks to the latest great models like o3 was whip up quick exploratory demos based on concepts I came across. This is great! Poster sessions and hallway chats often spark ‘it would be cool to compare that to X baseline’, or ‘I want to try this with Y’ - but by the end of a busy conference lots of ideas are forgotten. It was super cool to voice dictate an idea to o3, copy some code into Google Colab and have a minimal toy demo to try out. I did this for a few things that came up, including running one experiment entirely on my phone to make some plots related to a paper I saw, and the ones I did this for are now lodged in my brain enough that I’ll probably follow through on the experiments I want to run now. Wild that someone can say ‘We could probably train a model to go directly from CLIP penultimate hidden states back to an image’ and a few minutes later I can have a toy model training that does just that! The AIs still aren’t excellent for AI research but boy are they better than they used to be - more than enough for simple baselines and getting going quickly. Anyway - vibe research - can recommend!
Singapore is amazing
Singapore was delightful to visit. We stayed in ‘Little India’, an easy MRT ride from Chinatown and close to the marina and various parks and things, making it easy to experience a variety of different cuisines and aspects of the city. I loved how pervasive nature was - greenery bursting from any vailable spot, but somehow contained and managed very neatly. The food was particularly good - so many exciting options with nustling hawker centers full of amazing smells and sights. This post we sent to our family has more pics and my bird list - ft 36 new ‘lifers’ for me!
I Have The Best Job
Answer.ai is fully remote, and this was the first time most of us had met in person. And yet we all got on extremely well, instantly falling into a happy rhythum of exploring and enjoying the city together then returning to the hotel pool area to code, chat and play games. I made a travel ‘crokinole’ board (link coming soon) and introduced the team - hopefully enough got hooked that by next time we meet I’ll have some competition ;) All this already had me feeling grateful, but what really hammered home how cool this company is was answering (and hearing others answer) the repeated question of ‘so what does answer.ai do?’ that came up in conversation with all the new people we met at the conference. I’m so lucky!
Conclusion
Anyway, there are my jet-lagged ramblings. If we met there, or you wish we’d met there, or you just want to say hi, reach out! I want to cram in more social with fellow AI nerds for a week or two before I settle back to my quiet and introverted default state :)
Social is top
The best part of conferences is always meeting people. The venue setup was disappointingly short on affordances for the ‘hallway track’ but Jeremy managed to claim a couch in one of the cafe’s and hold court for various friends old and new :) And of course there were the usual evening dinners + events. This time around though I mostly spent time hanging out with the answer team, heading out to different parts of town for food and coding by the pool. Other hightlights were meeting a collaborator in person for the first time and getting taken out on a hike by a local fast.ai alum who knew all the good spots. If you get a chance to go to one of these conferences, my top recommendation is to focus on meeting friendly people and spending time chatting vs attending talks and things - far more interesting IMO! Poster sessions are a good way to find people with shared research interests. Pretend to be an extrovert - it’s only a few days!
PS, just after I shared this I found these tips on how to connect at academic conferences which are great! Go read that and follow them :)