ARENA 7.0 Impact Report
London's intensive 4-week program trained 29 participants, achieving the highest satisfaction score in its history.
The London Initiative for Safe AI (LISA) has released the impact report for ARENA 7.0, a flagship AI safety training bootcamp held in January 2026. The intensive, 4-week in-person program hosted 29 participants who gave it a record-breaking average satisfaction score of 9.7/10, a significant jump from ARENA 6.0's 8.7/10. The cohort was exceptionally strong, with seven participants holding or pursuing PhDs and seven having over a year of professional software engineering experience. The program also saw improved diversity, with 10 out of 29 participants being female.
Participants reported substantial upskilling, with self-assessed confidence in key AI safety disciplines soaring. Average confidence in mechanistic interpretability (studying how models work internally) jumped from 3.7 to 7.0 out of 10, reinforcement learning from 4.0 to 7.0, and LLM evaluations from 4.5 to 7.9. The program's selectivity hit a new high, with roughly 370 applications for just 30 spots, making it the most competitive cohort yet. Participants estimated the program condensed what would have taken 9.4 weeks of solo study into 4 weeks, with the vast majority finding the duration 'just right.' The report concludes ARENA 7.0 was arguably the most successful iteration of their 2025 grant period, effectively accelerating careers in the critical field of AI safety.
- Record 9.7/10 participant satisfaction score, up from 8.7/10 in the previous cohort.
- Major skill gains reported: confidence in mechanistic interpretability improved from 3.7 to 7.0 on average.
- Most competitive application round with ~370 applicants for 30 spots, featuring a strong cohort with 7 PhDs/PhD candidates.
Why It Matters
Demonstrates a highly effective model for rapidly training technical talent to address the urgent and complex challenges of AI safety.