@TownsUsa as long as manfred continues to systemically destroy baseball and it’s ability to attract fans, horse shit like this will be necessary to keep the game alive
@DataChaz 95% reduction that chops off the bits the agent needs isn’t a reduction. You need to balance retrieval and token cost, ala https://t.co/wyUBhN2K3E
@_Evan_Boyle I am begging people to alter their harnesses with RTK to only have rtk wrapping find and grep commands. All others trade too much retrieval for minimal token reductions
@alokbishoyi97 i do this with a lot of technical papers & books. ask your harness to run a map-reduce style workflow over the chapters of all the books at once for optimal token usage/retrieval balance
@bleuonbase i’m taking a frighteningly similar approach with a semantic graph baked in as the core knowledge base as well - i’ve found forcing the loop/graph to ground itself in relation to the resources (tokens, compute, in this case time) helps reduce low-entropy ideas
@Yuchenj_UW@karpathy i’ve found that without heavy prompting focused on exploring new spaces and not on low-entropy ablations, you get generic looking sweeps with low project value
@arnie_hacker really cool project. i personally have formed a “company” around a growing project in an attempt to get free startup credits, maybe try that avenue.
@Espngreeny then you simply don’t know ball anymore. franklin didn’t get anything more out of his players than the norm. that’s why he threw every big game except the ones his players made incredible plays by themselves. recruited pennsylvanians from pennsylvania, wow, such an accomplishment