(Applications have now closed. Express your interest in future iterations here.)

OAISI (the Oxford AI Safety Initiative) is excited to announce ARBOx, a two-week bootcamp to rapidly upskill in technical AI safety and machine learning. We'll do a condensed variant of the ARENA syllabus, aiming to build familiarity with mechanistic interpretability and transformers, along with a short introduction to RL and RLHF.^[1]

ARBOx is best suited for those who have a fairly strong technical background, but who lack experience in mechanistic interpretability or transformer models.^[2] We expect basic familiarity with linear algebra, Python programming and AI safety.

• Deadline to apply: end of day anywhere on earth, December 13th, 2024
• Dates: January 6th-17th, 2025
• What we provide:
  • Accommodation in central Oxford
  • Provided meals
  • Partial support for travel costs
• Content: an abbreviated version of the ARENA syllabus
  • gpt-2-small from scratch
  • A paper replication or two (e.g. Redwood's IOI paper)
  • A brief introduction to RLHF

See you there!

Questions? Reach out to us at arbox@oaisi.org.

1. ^{^}

   Our syllabus is inspired by CaMLAB, a similar programme run by our counterparts in Cambridge.

2. ^{^}

   We believe that understanding some basic ML/mechinterp is useful for a large range of roles in AI safety, beyond directly working in technical AI safety. Strong applicants who intend to work in more governance-related areas are particularly encouraged to apply.