Hoagy Cunningham (Anthropic & Former MATS Scholar) “I joined the previous version of LISA in around April 2023, when it was the Moorgate workspace for the MATS 3.1 cohort. At the time I was a first-time independent researcher on an LTFF grant, coming towards the end of my first piece of research and had been improving my research and programming skills but not feeling particularly satisfied with the current quality, and feeling rather lonely and lacking in people to talk about my work with.

When I started coming into the office I generally felt more productive and healthy- I'd been working in the British library which was good for day-to-day productivity but hardly spoke to anyone and I felt this made it very easy for the weeks to drift by while going in circles, and I also couldn't work late or on weekends when I wanted to. Having the space gave me a big energy boost from having a community and a sudden influx in ideas and a huge jump in how much I understood what junior researchers like MATS scholars were doing.

More important than the direct productivity boost were the conversations in the office, at the time around the Toy Models of Superposition paper which I started to think about on my own and with Spencer, and then emailed Beren with my thoughts who put me onto Lee Sharkey who'd already made significant progress in the direction which I had in mind. I then very quickly took up the sparse autoencoders agenda which Lee didn't have the time to follow up on, which naturally segued into him becoming my MATS mentor, and we had a very productive summer, culminating in the first paper on sparse autoencoders (https://arxiv.org/pdf/2309.08600.pdf), which was well received, and the intermediate LessWrong posts were also heavily cited by Anthropic in their recent work on superposition (https://transformer-circuits.pub/2023/monosemantic-features/index.html).

This direction is now a core research focus for Anthropic's mechanistic interpretability team, and I've now received an offer to begin working full-time for them, starting early in the new year. It feels very unlikely to me that I'd be in this position without having had access to the LISA space. I think I would perhaps have had a 50% chance of getting into MATS otherwise, and probably a much lower quality of output, and would have been unlikely to receive a job offer in this round, instead going through another round of perhaps 6 months of independent research, in which having LISA would again be invaluable.”

Callum McDougall (Anthropic/ARENA) “Having an office to work from, and especially one full of other alignment researchers, has massively boosted my productivity. It's one thing to have available office space for work, but it's another for that office to also be a source of fascinating spontaneous conversations about the things we're working on, the output we're producing, or the alignment field overall. I've also been able to co-work with people while using this space, including Kathryn O'Rourke while putting together a timeline for ARENA, and Hoagy Cunningham while we were both preparing for interviews with Anthropic. We both ended up receiving offers, and I'm confident that I wouldn't have received mine without the benefits that the LISA office space & community has provided over the last few months.”

Simon Lermen (MATS) “I think I had one of my most productive episodes while working in LISA, I published a paper and worked on another significant research project while working there. I also think that it was great for networking and presentations.”

Sohaib Imran (Lancaster University) “Productivity is something I have struggled with for a while. I am most productive when working around colleagues, since that maintains a helpful illusion of accountability. Since I started working remotely from Lancaster, if it were not for LISA, I would have found having such a work environment nigh impossible. I estimate that working at LISA has increased my productivity by nearly 50%.

Spending more hours working is not the only way LISA has benefited my research. LISA appears to have carefully crafted an environment perfectly suited for networking, collaboration, and innovation. One of the most important ingredients of this has been Tuesdays and Thursdays lunchtime talks, which provide food for thought and an opportunity to talk to senior researchers in the field in a friendly setting. I have also had the opportunity to engage in thought-provoking conversations in many of the after work events LISA has organised. Lastly, collecting so many of London’s AI safety researchers in one place is an excellent recipe for promoting collaboration in of itself.

As a result of this exposure and increased productivity, I have explored many new research ideas, and am now planning to write a paper on evaluating situational awareness in Language models, hopefully collaborating with others interested in the topic.”

**Rusheb Shah (Apollo) “**I participated at ARENA, which was hosted at the LISA office in moorgate. I got a huge amount of value from this.

• I spoke to an Apollo team-member about project ideas, which I later implemented and emailed them about, which was a factor in me getting a job offer from Apollo (they mentioned it in my feedback)
• I met collaborators for my LLM red-teaming research, and was able to start off the research project in-person which helped things to go more smoothly. This research was another factor that let to me getting a job offer. I may have met the other participants if it hadn't been in person, but would have had a lower chance of working with them. There was 1 collaborator that I doubt I would have met because they were a MATS scholar and not an ARENA participant. This person's prompting expertise was vital for the success of the project.
• ARENA being in-person meant that I got a lot more value out of the technical parts of the program, since I could pair-program with other participants and the barrier to asking for help was lower. Therefore I learnt more. I also made connections in the AI safety community whom I still have discussions with about AI Alignment topics.
• I just started at Apollo 2 weeks ago. Coming to the office every day has made me much more productive. I've already benefited from lunchtime conversations with LISA members, which have helped me improve my internal mental models of AI Safety.”

Anonymous (MATS scholar) “LISA has been a very productive working environment and significantly better than the counterfactual of being at home, and better than the counterfactual of being at LEAH. I have not felt more productive at any other office (though this is due to circumstances as well as LISA). Being able to work with other AI safety researchers in-person has been valuable. The potential of the space is limited by the high fees LISA currently has to charge, which deter many and prevent greater network effects.

I wrote a paper accepted to a workshop at NeurIPS as part of my MATS extension, and since then I have made much progress towards a major safety project that will likely be a conference paper at a major AI conference. All in collaboration with several other LISA members.”

Anonymous (Technical AI Safety Researcher at organisation)  “LISA has connected me directly to relevant people in government; provided generous diet coke supplies (v important); great place to host meetings; sparked joy and new ideas from lunchtime presentations; lovely event/social space! I'm more productive – it feels like home.

My output has been “A few LW posts, a NeurIPS workshop paper, built an interpretability library based on our new methods, and raised significant investment to progress interpretability research”