secemp
@secemp9
🧠 researcher • 💻 swe • 💾 data scraper • 🌌 universe's jester • Author of TraceBack
interesting behavior during their failure mode though, at peak hours claude become lazy, while gemini despair
you know claude is currently being "optimized" for peak hours when you see this being OK (here replacing the main goal with a simple placeholder)
could probably make a simple eval running every N time during the day, every day. the one pattern I'm seeing is it does this the most if it encounter any level of difficulty, during say 2-3 turn at most
you know claude is currently being "optimized" for peak hours when you see this being OK (here replacing the main goal with a simple placeholder)
I get providers, honestly I would quantize the kvcache too during peak hours (or whatever other performance hitting tricks they're using), but, the model still need to be usable you know you could just RL it such that it still work great during those kinds of optimization right?
you know claude is currently being "optimized" for peak hours when you see this being OK (here replacing the main goal with a simple placeholder)
you know claude is currently being "optimized" for peak hours when you see this being OK (here replacing the main goal with a simple placeholder)
yeah this is known for a while. many VLM support image prompting (directly or otherwise)
what are the odds, that this was made 8 minutes ago...when just last night me and friends were talking about kv caching on official model providers...
I said similar things mentioned in that paper a while ago, nice
🚨 The era of infinite internet data is ending, So we ask: 👉 What’s the right generative modelling objective when data—not compute—is the bottleneck? TL;DR: ▶️Compute-constrained? Train Autoregressive models ▶️Data-constrained? Train Diffusion models Get ready for 🤿 1/n