Prioritise writing and sharing

How to write well

How to share your work

Get it working

Figure out what needs to be true

Deciding what to test first

Using your results

Evaluation criteria

Submit your project

You can make AI systems safer

Getting started

Project ideas

Scope it down

Every time you get a result, you face a decision:
- **It confirms your assumption** — narrow further, escalate to a more thorough test, or move to the next assumption on your list
- **It surprises you** — widen again. Why was your prediction wrong? Is the assumption flawed, or was the test? This is often where the most interesting findings come from.
- **It's inconclusive** — before abandoning it, you can try escalating to a slightly more expensive test that might give a cleaner signal. If that's still inconclusive, move on.

If your direction is a dead end and you have enough time for another loop, pivot. 

If you don't, start writing up. A clear account of "I tested this direction, here's what I found, here's why it didn't work, and here's what I'd suggest trying instead" is a genuinely useful contribution. Don't treat a dead end as failure — treat it as a finding that saves someone else the same work.

Check with your mentor if you're unsure. They can help you judge whether you have time for another loop or whether your remaining time is better spent writing up what you've learned.


For each test you decide to run, you can determine whether to:
- Run more rounds for statistical significance 
- Conclude that this isn’t as interesting as you thought 
- Reframe the question based on your learnings 



Technical AI Safety Project

Using your results

Resources

Exercises