I spent 15 hours on ARENA and decided it’s not for me

Abstract

ARENA is well-structured, practical and supported by a self-study community, but the author stopped roughly 15 hours in — after ¾ of chapter 0.0 — because progress was far slower than the published estimates given rusty linear algebra and calculus, because the maths was demotivating, and because she had already extracted the understanding of the course she needed to advise BlueDot graduates on whether to take it.

Framing

• Written as loose cached thoughts, primarily for her future self if she reconsiders doing ARENA.
• ARENA is one of the top recommended courses for getting into AI safety research.

What she liked

• Very structured — gives you the information needed to decide how to approach the curriculum, whether that means shoring up fundamentals or skipping sections. Following along, you never feel lost about where to go next.
• Very practical — constant exercises to build and test understanding. A small amount of theory for orientation, then learning by running the code and seeing what changes when you tweak things.
• Support for self-study — a self-study Slack community where members proactively answer each other’s questions.

Why she stopped

• It took far longer than expected. A LessWrong guide estimated more than a day for chapter 0.0; it took her nearly two days to get three-quarters of the way through.
  • She did data science in undergrad three years prior and was confident in her Python, but underestimated how much linear algebra and calculus she’d forgotten.
  • Remaining chapters are meant to take about the same time, so at that rate it would be another week before training her own neural network.
• Demotivation — “Math and I have a tough relationship.”
• She’d already got enough utility for her role. She leads the Technical AI Safety course at BlueDot, and now understands ARENA well enough — in time, coding experience and maths required — to advise course graduates on whether to take it. She judges the extra granularity from another 100+ hours not worth it, while allowing she could be wrong.

What she’d do differently if she picked it up again

• Get a mentor. Using Cline helped, but it wasn’t good at indicating which parts genuinely needed deep understanding and which could be skimmed. Someone who has completed the curriculum would advise better.
• Move on more quickly. She was fixated on understanding every concept deeply, which may have been the wrong approach — better to look at solutions sooner and work backwards than to reason everything out in her head first.