Lachlan Ewart

Lachlan Ewart

3rd Year Mathematics undergraduate, University of Warwick. Also Director at Mercial and SPAR fellow.


Hello! I am a final-year maths undergraduate student at the University Of Warwick. I am interested in AI Safety, and spent a month working in London over August, gaining context. I am part of EA Warwick, and I write on Substack and occasionally LessWrong. I care about maximising my impact, and I like interesting takes on how I can do this. Right now, I am a SPAR research fellow under Alec Harris (Pivotal, Redwood), and am director and founder at Mercial, previously the Warwick AI Safety Initiative. I have taken the GWWC pledge.

Writing

I wrote blogs for PauseAI, on the dangers of AI. I also had lunch with my MP to talk about AI risk, and managed to get her to sign PauseAI's open letter to the PM. I also write on substack, on a few topics, mainly AI Safety, or morality.

Mercial

I founded Mercial in September 2026, as I saw value in increasing the overall capacity of remote research programs, and increasing the AI safety fieldbuilding presence at Warwick. Now (Nov '26) we have just under 50 fellows working in the current cohort of the research fellowship, and we host hackathons and talks on Warwick University's campus. The team has grown from just me to 9 volunteers, and I am excited to see the organisation grow further! You can read more about Mercial at mercial.uk.

Projects

How the White House Accord Affects Open Weight Models (a project I mentored)

Mercial October '26 Cohort · Writeup

An overview and pushback on the White House Accord on Super Intelligence, in regards to how well it would regulate open weight models.

Testing Monitoring Models with Encoded Transcripts

2026 Apart AI Control Hackathon · Writeup

Stress-tested chain-of-thought monitorability by feeding Claude Haiku 4.5 transcripts of malicious activity under different encodings, and found monitoring degraded on half-encoded transcripts more than fully encoded ones.

LLM Continuous Probability Inference

Low probability estimation · Code

We try to use an energy-based sampling method to estimate the probability of rare outputs of an LLM while it is in the training phase, as distribution shift of the inputs in deployment can dramatically alter the probability of outputs. Therefore we need to be able to tractably estimate very rare outputs in the training phase.

Education

University of Warwick

BSc Mathematics · 3rd year, expected 2027 · current grade: 1st

Royal Grammar School Newcastle

A-Levels in Mathematics, Further Mathematics, Physics, Computer Science (4 A*) · 2012–2023