Simon Willison on Nostr: nprofile1q…8klka nprofile1q…ufa5p it's mainly about getting them to respond in a ...

nprofile1qy2hwumn8ghj7un9d3shjtnddaehgu3wwp6kyqpqes0y2pau8nwpzrzzjes6mpqd8xa94x3ywe7dnmxswamxkmetcneqx8klka (nprofile…klka) nprofile1qy2hwumn8ghj7un9d3shjtnddaehgu3wwp6kyqpqt8u7s2z2e8j5gdup94s7kjyjxqle03hekumc3kmd0ppermdwsqrs5ufa5p (nprofile…fa5p) it's mainly about getting them to respond in a way that's as useful as possible - to do that you need them to have seen many examples of similar conversations that went in a foodnfir action, and generating "synthetic" versions of those conversations is a great way to do that

Also common is synthetic code: generate lots of valid code examples (that actually compile for example) and use those to help train a coding model