I participated in BlueDot’s Technical AI Safety Project Course

• What this course is: Technical AI Safety Project
  • Publish a project in roughly 30 hours.
  • Receive direct mentorship from an AI Safety expert.
• Why I decided to apply for this course:
  • I’m currently focused on AI Safety field-building. You can find my project details here.
  • As Neel Nanda argues, mentorship is key.
  • I am planning to organize my own mentorship programs. I figured the best way to understand the ingredients of a high-quality program is to experience a great one firsthand as a mentee.
• Project Overview
  • AI models might employ deceptive strategies as a form of scheming or misaligned behavior.
  • Relying solely on output monitoring (the black-box approach) is insufficient.
  • This work demonstrates that a simple linear probe can distinguish between honest and deceptive internal states.
  • We extended this methodology to evaluate Japanese models and datasets.
  • Key Finding: We observed that performance degraded significantly when applied to Japanese models.
  • Full details are here
• How the course went (Logistics & Impressions)
  • Participants scope, pitch, and drive their own independent projects.
  • Weekly Check-ins: We had 1-hour sessions with our mentor and fellow participants.
    • We shared progress and gave each other feedback.
  • High attrition rate: Only 2 out of 6 people in my group actually completed the course.
    • I guess people just get too busy.
    • My intuition here is that we desperately need mechanisms to address this if I organize a similar program.
      • For example, TARA uses a financial commitment bond—a highly promising hack for accountability. Participants pay a deposit of around $60 before the course, and it is only refunded upon successful completion.
  • 1-on-1 Mentorship (Optional):
    • Participants can proactively request 1-on-1s with mentors.
    • You aren't bottlenecked by your specific cohort mentor. BlueDot provides a diverse pool of expert mentors you can ping for feedback or quick sanity checks.
    • Conversations aren't strictly bounded by the project—seeking general career advice or discussing broader AI Safety topics is highly encouraged.
    • I personally scheduled one session.
    • This is an incredibly valuable opportunity. In my experience, organically finding mentors is one of the hardest parts of entering the field.

What I’ve learned from the course

• Managing Cloud GPUs:
  • I learned about compute costs, environment configuration, and which machines to choose.
  • Read my learnings on Cloud GPUs here(Japanese blog).
• The Research Workloads:
  • I learned what kind of work is needed and how much effort it takes.
  • I initially overestimated the resources (time and compute costs) required.
    • I thought an experienced researcher would need a month to do what I did. Now, I estimate an expert could replicate this in just one week.
    • I expected compute costs to be $100–$1000, but it only cost $30.
• Working on Projects Make Reading Papers Efficiently and Deeply:
  • I found myself reading papers not only much faster but also with a deeper understanding of the content.
  • I believe this is because I was in "chasing mode"—actively searching for specific answers, like why the authors chose layer 22 for activation extraction. Having a concrete goal made it easier to truly grasp the concepts instead of just passively reading.

What I’ve learned about mentorship

• Disclaimer:
  • I wrote this with Neel's post heavily in mind, so I encourage you to skim it first if you haven't already. I am not adding fundamentally new frameworks; rather, I am providing empirical validation from my own experience.
  • Limitations of a short sprint:
    • Because this project was basically a one-month sprint, it is inherently hard to demonstrate the kinds of value that require a longer feedback loop.
      • Mentor values that take time to compound, such as providing "supervised data" to help you build your own research taste.
      • Building slow skills, like how to prioritize experiments, when to pivot, and how to generate good research ideas.
• My mentor pointed me toward a ton of literature I wouldn't have found independently.
  • It’s tempting to argue that modern LLMs can act as research assistants just as well as (or better than) human mentors.
  • However, my intuition here is that human mentors remain crucial for tackling unknown unknowns.
  • For example, when brainstorming extension directions, my search space was completely constrained to 'probing'.
  • He zoomed out and introduced adjacent concepts like "steering"—a broader perspective I wouldn't have generated on my own, but which proved highly valuable for contextualizing and discussing probing.

What I've learned about running a mini-project