Do we need "alignment"?
Probably not in the large sense. An iceboat driver does not pursue goals outside the game, and the worst it can do is win races it should not or make a server less enjoyable for the people on it. For now we think the useful work is preparation and evaluation.
A smaller version of alignment does apply once bots act as helpers. Does an NPC race at the difficulty you asked for? Does a coach give honest feedback? Does a bot follow the server's rules? We will measure those as part of our evaluation. There may be more to this than we expect, and we will update this page if so.
What we want to prepare
- Measure the driver. Exploring Benchmarks against human lap times, on tracks and conditions the model has not seen.
- Measure the tells. Exploring How distinguishable is a bot from a person? Input timing, consistency, how it recovers from mistakes.
- Help server owners. Planned Write-ups and simple tools so owners can decide what automated play they allow.
Protecting servers
If bots get good, a server owner may want to know who is driving. We expect no single signal to settle it, and we expect detection to stay an ongoing back-and-forth.
- Have a person review any flagged player before acting. Some humans are very fast, and a wrong call affects a real player.
- Prefer bots that declare themselves: a clear label, a registered account, or a server-side flag.
- Let owners set the rules per event: no bots, labelled bots only, or open.
- Take the most care with ranked races, leaderboards and prizes, since an unlabelled bot has the biggest effect there.
We do not have a reliable detector yet. If we ever build one, we will publish how we tested it and where it fails.
Where racing bots can help
The same driver that worries an organiser could also be useful to a server and its players. These are the uses we would like to explore:
- NPC racers. Fill the grid when only two players are online, with difficulty set to match the room.
- Coaches. Show a racing line, compare your lap with a good one, and explain where the time went.
- Ghosts and pacers. A rival to chase in time trials, at your pace or slightly better.
- Track testers. Let builders drop a bot on a new track to find bad corners, broken jumps and shortcuts.
Principles we are starting with
- Labelled. Our bots should be identifiable as bots wherever they race.
- Opt-in. Server owners decide whether our bots join their server.
- Honest. We publish our results, including the failures.
- Humans first. If a bot makes racing less fun for people, we change the bot.
These are early drafts. If you run a server or race on one, we would like to hear what you would want. See Join, or the broader plan on Research.