I agree that there is a huge technical and execution/growth-hacking/product risk in making the flywheel of [BetterDiscourse] self-sustaining.
But I want to emphasize that I think the “good reactions feedback data from production use → better-trained SSM → more value to the user even in a completely “local” type of use → more data” loop is, IMO, more important initially,than the network effect (the latter is supposed to attract users who are not active voters, and revenue from them). In particular, training the original version just on LessWrong data, although it’s high-quality, may not produce a good SSM because it will likely overfit on the particular topics, language, and worldviews that appear on LessWrong, rather than, let’s say, in political discourse in Portuguese.
I agree that there is a huge technical and execution/growth-hacking/product risk in making the flywheel of [BetterDiscourse] self-sustaining.
But I want to emphasize that I think the “good reactions feedback data from production use → better-trained SSM → more value to the user even in a completely “local” type of use → more data” loop is, IMO, more important initially,than the network effect (the latter is supposed to attract users who are not active voters, and revenue from them). In particular, training the original version just on LessWrong data, although it’s high-quality, may not produce a good SSM because it will likely overfit on the particular topics, language, and worldviews that appear on LessWrong, rather than, let’s say, in political discourse in Portuguese.