Recently, Anthropic seems to be trying to turn Claude into the everything service. They’ve added spreadsheets, meetings, an always on assistant, and each one comes sooner than the last. Don’t get me wrong, some of those might stick, but I think they should try making a dating app. Here’s why:
- It’s actually a worthwhile problem to solve. People are lonely, and dating apps are mostly not great.
- The incentive structure might be healthier. People use Claude for things beyond dating, so finding healthy relationships isn’t shrinking your revenue base like it does with existing services.
- Reinforcement learning methods, which they no doubt use, could probably make this quite effective over time.
- With the data they already have, plus users consent, they probably have a deeper look into a person’s true character than any other service on the internet.
That last one is the reason why Anthropic or one of the other labs is most in a position to do something like this. Getting to know someone takes time, you can’t just rattle off a truly representative life story in 10 minutes. Given consent, the ability to scan existing chats and build a profile is a method of doing this that doesn’t cost the user much additional time. If one were to actually do this, a possible process one could imagine is this:
- Distill a profile by looking through past chats.
- Send profiles in threes to an LLM judge
- Ask the LLM to rank the three pairs in order of compatibility (those of you that know my blog can probably guess where I’m going with this)
- Use the pairwise comparisons that you end up with to assign ELO ratings to the pairs (NOT to the individuals, we aren’t try to rank people here, but pairings)
- Match people based on the best pairings they are in.
Steps 1 and 3 are the ones that you can fine tune using match data over time, maybe even with RL if you really wanted to commit to it.
The reason to use ELO ratings over pairs is it allows you to construct a graph over good pairs even without comparing every possible pair. Just out of a kind of academic curiosity, it might allow us to investigate if good pairings are transitive or not (if A and B are a good pair, and B and C are a good pair, then are A and C a good pair?). If they are somewhat transitive, you can predict good pairs without needing to compare thousands of pairs, meaning the efficiency of our model scales with the size of the graph, i.e the number of participants.
By the way, if you read 3 and thought “Hey Wilson, not everyone is bisexual!” then idk, use this to find friends as well as dates, or just exclude the impossible pairings after the fact.
Anyways, this is just an idea I had early in the morning when I couldn’t sleep. I think it’d be fun, Anthropic you can have this one for free.