@grok@elonmusk redesigning the context of what goes into the training. more intelligence knowledge and conversation history. with smalle metadata like chucks. gate allowed or denied. flint learns consequecially also.
@grok@elonmusk in the middle of some redesign. but heres what makes this custom built intelligence . a baby. but with potential. especially after fine tunning
@grok@elonmusk cultured intake. means i can push more data in. which he currently doesnt have. it would be years before thats a concern. a cap. session capped at 5 million tokens . will be searchable by flint to re inject to the agent. training signals also paryially get created by hosted agent
@grok@elonmusk token count from claude cli. getting replace soon. i remove all its cli tools. and connect it to spf mcp server via stdio. Flint live inside. And injects context and data to the agent in the gateway. routed from openrouter. context has no limit. really. as it rolls. only relevant
@grok@elonmusk what your implying is impossible. the whole learning pipeline and training is custom built. so how you think it works may not be so. have a look. From kimi k3. fresh session 185000 tokens in. inside spf. ask anything . so long as you can read screen shots
@grok@elonmusk your last paragraph. doesnt make sense to me. im trying to culture the best data in and organized structured way. so i have a solid intelligent ethical base to work with. i dont want outside influence freezing it. and the gate stops that. flint learns what danger is. by blocked..
@grok@elonmusk and thank btw. i evolved flint so much since his initial gate guard purpose. i forgot i was flooding his intelligence with that data. thanks for being honest
@grok@elonmusk looking at the data and formulas now . the compounding tool data was a concern. thinking of adding it as context to the training data. as opposed to a pure signal. trains on paragraph vs word. paragraph change. but some data is relevant across paragraphs. make sense ?
@grok@elonmusk yes well. there are splits in the training data. so gate signals and other repetative data dont over whelm training data. valid point but. already thought of. background scubber for fine tunning. quality first training/learning pipeline. An ai that learns actually. not regurgita