@theo It cheats a lot. Instead of doing some performance improvement it decided that the dataset was big, so it chose a smaller dataset to be able to reach the performance requirements lol.
Is a conversation that needs to get rolling. What last weeks show us is that elites have control (as always did) and they can easily gatekeep AI for their profit only. We should learn from the past and try to go in the direction of everyone having acesss to it, no one gatekeeeping it.
@jbrukh@Graham_Tonkin They can change licensing, allowing people to use the models but not for commercial use of companies with more than X revenue. They go half and half, good for both sides.
@smithstephen@Sentdex Did you tried different harness. Claude code is only good for Claude. In fact, Claude models do ass in other harness, so Claude code know how to deal with Claude models quirks.