Every student club says it's beginner-friendly. Almost none of them are, and the reason is structural rather than cultural: a beginner in a room of people who already know the material is still a beginner, no matter how welcoming everyone is.
You can be as friendly as you like. If the only project on offer assumes a semester of linear algebra and a working PyTorch install, the friendliness is decoration.
The specific change
So for Winter Projects we made the beginner track a real track: its own team, its own mentor, its own scope, and no assumed background.
The important word is own. The failure mode we were avoiding is the one where beginners get attached to an existing team as junior members, which sounds inclusive and in practice means they watch someone more experienced do the interesting parts while they're handed the setup scripts.
A separate track means the beginners are the ones making the decisions, badly at first, which is the only way anyone has ever learned to make them well.
Why RL, specifically
We chose reinforcement learning for the track, which surprised some people. The obvious beginner project in 2026 is to wrap a large model and ship a chatbot — it's faster, it demos well, and it feels current.
But it teaches you the API, not the field. You end up fluent in someone else's abstraction with no sense of what's underneath it.
Reinforcement learning has three properties that suit a first project:
- It's visual. You watch the agent get better. A loss curve going down is abstract; an agent that stops falling over is not.
- The loop is fast. Change the reward, run it again, see what happened. Fast feedback is most of what learning is.
- It's honest. Either the agent learns or it doesn't. There's nowhere to hide, and working out why it didn't is the actual skill.
The cost is that reward shaping is genuinely hard, and it's where most first RL projects quietly die. That's a known risk and it's where we've put the mentoring.
The thing we're actually testing
The real question isn't whether the track produces a good RL agent. It's whether someone who arrives in June knowing nothing is, by the end of it, someone who can start a project on their own.
That's a slower thing to measure than a demo night, and we won't know for a while. But it's the one that matters — a club that's beginner-friendly for one semester and then leaves people where it found them hasn't done anything.
If that describes where you're at, membership is open and the next intake goes to members first.