Your Impact at LILA
We're building a talent-dense, high-agency AI safety team at Lila that will engage all core teams within the organization (science, model training, lab integration, etc.), to prepare for risks from scientific superintelligence. The initial focus of this team will be to build and implement a bespoke safety strategy for Lila, tailored to its specific goals and deployment strategies. This will involve technical safety strategy development, broader ecosystem engagement, as well as developing technical collateral including risk- and capability-focused evaluations and safeguards.
What You'll Be Building
- Evaluations to test for scientific risks (both known but especially novel) from cutting edge scientific models integrated with automated physical labs
- Initial proof-of-concept safeguards, such as ML models to detect and block unsafe behavior from scientific AI models, as well as from physical lab outputs.
- Understanding of a range of model capabilities, across primarily scientific but also non-scientific domains (e.g. persuasion, deception) to inform Lila's broader safety strategy.
- Broader, high-quality research efforts - as and when needed - for scientific capability evaluation and restriction.
What You’ll Need to Succeed
- Bachelor's degree in a technical field (e.g., computer science, engineering, machine learning, mathematics, physics, statistics), or related experience.
- Strong programming skills in Python, and experience with ML frameworks (including, for instance, Inspect) for large-scale evaluation and scaffolded testing.
- Experience in building evaluations, or conducting red-teaming exercises, for CBRN / cyber risks (or for frontier model capabilities more generally, including both unsafe and benign capabilities)
- Experience in designing and/or implementing (directly or through consultation) AI safety frameworks for frontier AI companies.
- Ability to communicate complex technical concepts and concerns to non-expert audiences effectively.
Bonus Points For
- Masters or PhD in a field relevant to safety evaluations of AI models in scientific domains, or a technical field.
- Publications in AI safety / evaluations / model behaviour in top ML / AI conferences (NeurIPS, ICML, ICLR, ACL) or model release system cards.
- Experience researching risks from novel science (e.g. biosecurity, computational biology, etc.) or working with narrow scientific tools (e.g. large scale foundation models for science).