Claude went down at 2:30 PM IST. I was mid-session on Desktop β context blew past 200k and just kept going. Hit 275.8k. Bar said 100%. Still typing, still getting responses.
Outage hits and the guardrails nap too.
Wild Tuesday. π
@claudeai@ClaudeDevs@bcherny
@Pseudo_Sid26@rohit4verse continuation:-
truth gets matched in those n runs. Then inspect why the failed runs trajectories pick differently based on passed trajectories. Most of the time agents works on things based on how main prompt gets structured. Let me know your thoughts?
@Pseudo_Sid26@rohit4verse Most importantly the model versioning. If v1 runs on x model with a y prompt then in reliability if gets passed push to prod. But if you try with x_1 model with same prompt trajectories changed. Most imp everytime I see out n runs of same input to system, how many times ground
@anirudhbv_ce@JamesNguyen868@OpenAI@GeminiApp@sentra_app Hey, does the work make sense? Also have you observed the same phenomenon by taking the least possible dimension from open ai text 3 large which supports matryoshka property
@anirudhbv_ce@JamesNguyen868@OpenAI@GeminiApp@sentra_app hey what happens if we take lowest possible dimension that supports matryoshka property? i mean eariler i tried to made some analysis on "does utilising the cumulative percentage of embedding model input tokens matters for the granular retreival or not"