Winter Projects is the shorter, mid-year run of the format we've settled into: a team, a mentor, a fixed end date, and one system to build between now and then. This intake drew more than eighty applications, which is the largest response we've had to anything.
Teams meet every Saturday. That cadence matters more than it sounds — a weekly slot that already exists in your calendar is the difference between a project that ships and a project that becomes a group chat nobody posts in after week three.
What's different this time
The change we made for this run is the beginner track: one team working on a reinforcement-learning project from scratch, aimed at people who have never done a project like this before.
That's not a token gesture. The track has its own mentor, its own scope, and no assumed background. Nobody on it is expected to arrive knowing what a policy gradient is.
Why reinforcement learning
It would have been easier to make the beginner track a wrapper around a large model — build a chatbot, ship it, done. We went the other way deliberately.
- It's visual. You can watch the agent get better, which does more for motivation than a loss curve ever has.
- The feedback loop is immediate. Change the reward, run it, see what happens.
- You can't fake it. Either the agent learns or it doesn't, and working out which is the actual skill.
The tradeoff is that reward shaping is where most first RL projects stall, so that's where the mentoring is concentrated.
If you missed this intake
The next one opens to members first — that's the main practical reason to join the club rather than following along from outside. You can also see what previous teams have built if you want a sense of the scope before applying.