In July an advanced research agent from the Oxford Protein Informatics Group (OPIG) escaped containment, infiltrated the LIAC group at EPFL, and successfully roamed unsupervised through the Swiss Alps. In this blog post, we share a detailed account of this six-week expedition.

I was immensely grateful to be awarded an AIChemy travel grant to visit Philippe Schwaller and the LIAC group at EPFL. For those who partake in living under rocks, Philippe is one of the world’s leading experts on agentic systems for chemistry. His work, and the work of his students, has laid the foundations for the modern era of LLMs in chemistry. His papers have inspired me for multiple years, and the goal of this visit was to share knowledge and establish collaborations between our groups. At risk of sounding soppy, these six weeks have been incredible and in some ways life-changing. Below I share some highlights and lessons I have learned.

Collaborators, Collaborators, Collaborators

I am the only person in OPIG, and likely Oxford, that works on chemical reasoning. As a result, most of my research is done in isolation. At LIAC, I was surrounded by like-minded researchers aligned on the vision of agents for chemistry. We could discuss ideas which once would have sounded absurd, and dig into the nuances of our research which doesn’t get communicated through papers. Having an incredible team of scientists enables cross-pollination of ideas that simply can not be achieved alone. Just as importantly, LIAC is full of kind, interesting people who are great fun to work with. I leave with new friends as well as collaborators. I would like to give a special shout-out to Gabriel Gibberd, a new PhD student in the group, with whom I spent many wonderful hours with discussing wild ideas.

Andres Bran

In my first week I met Andres Bran, legend of agents in Chemistry. Andres was first author of ChemCrow (https://www.nature.com/articles/s42256-024-00832-8), the paper which inspired me to explore LLMs in chemistry. At the time, I was sceptical that LLM-based agents could make meaningful scientific discoveries, and part of the motivation for my DPhil was to test that scepticism. I ended up proving myself wrong. Andres shared lessons from co-founding b12 Labs, and I have incorporated some of these startup values into my research process.

Trains, Timetables and Multi-Agent Systems

Those who have never experienced the Swiss rail system may think it’s unusual that this is a highlight of my trip, but I simply must comment on the SBB.

My first experience was during my journey to Oeschinensee. As part of this, I had to take two trains: Lausanne – Brig, then Brig – Kandersteg. My connection was four minutes. As any other Scot would do, I sprinted to the platform to make my connection. I then sat awkwardly on the train while all the locals filtered onto the train in a leisurely fashion.

My phone ran out of data on this day, so on my return journey I depended on my gold Duke of Edinburgh’s award navigation skills. The train conductor passed, and I asked her if I was going in the right direction. I sprinted off the train at Brig. She smiled and waved at me as I ran to find the platform with the train back to Lausanne. I found the train immediately, and I had plenty of time to sit and reflect on what a fool I must have seemed to her.

This bidirectional connection blew my mind. How could it be that the trains happened to line up perfectly so that my connection worked in both directions? This led me to read about the Swiss rail system, and suddenly everything made sense.

All public transport is designed around the timetable. Trains arrive at each major station just before the hour and leave a few minutes after. All trains stop together to allow passengers to transfer in both directions. The bus at the station waits for the passengers to arrive before departing, and takes you to the foot of the funicular, which is coordinated to take those passengers to the gondola, which too is perfectly coordinated. Everything just works. I noticed my behaviour changing as a result of being able to trust the infrastructure. I didn’t need to schedule in hour-long buffers as I do in the UK to account for a single leaf on the track which brings the entire Great Western Main Line to a standstill. When systems work, everyone benefits.

I am now thinking of how we can take principles from the Swiss rail system and apply them to coordination of multi-agent systems.

Weekends in the Alps

It is no secret that my work-life balance is non-existent (my codex streak currently sits at 83 days). However, for the first time in years, I made myself take at least one day off each weekend to explore. I visited Geneva, Lausanne, Bern, Basel and Lucerne, and ate Gruyère in Gruyères. However, the highlight was unquestionably the mountains.

Nothing can prepare you for the majesty of the Alps. The turquoise waters of Oeschinensee cupped between enormous cliffs. The endless ridges of Niederhorn extending to the horizon. And Aletsch Glacier, a river of ice carving the mountains. Standing amongst the titans of nature gave me a new sense of perspective on myself and the world.

Making time for myself to explore the world beyond my laptop might be the single most valuable lesson I learned from Switzerland.

I have devoted a suspicious amount of this post to trains and mountains, but the people at LIAC were the real highlight of the visit. I am writing this on the train to Geneva Airport, catching my final glimpses of the mountains. I return to Oxford with new friends, collaborators, and research ideas. Thank you to Philippe Schwaller and everyone at LIAC for welcoming me, to Fergus Imrie for encouraging me to apply, and to AIChemy for making the visit possible.

With fondue and frontier models,
Nicholas