开启时会模糊预览图，关闭后正常显示
releases like this give startups 6 months longer runways and something to hype ship a month later. incredible value for 10B active params.
making gradients flow • eXperiments lab
not sure about you but 2026 is about efficiency and engineering for me. creative endeavours would have to take a backseat until and unless they have 100x RoI potential.
making gradients flow • eXperiments lab
> Hallucinations are actually "Over-Compliance." Interventions prove these neurons force the model to prioritize satisfying your prompt (even with false premises) over telling the truth. so is there like a concept of noise here as well, where training a model on cases where it doesn’t know the answer and chooses to refuse clearly would help mitigate hallucinations?
making gradients flow • eXperiments lab
RT @cloud11665: A good thought experiment is how do you count something like an IMO gold medal by the METR standard - it is a task that tak…
making gradients flow • eXperiments lab
i really wish to know what changed in gemini 3 flash training on top of anything other than OPD/regular RL. is it just good env engineering? benchmaxxing? or something that would make next gemini pro even deadlier?
making gradients flow • founder - eXperiments lab
RT @xeophon: okay, small thread! in times of chatgpt & claude code, designing tasks which are accesible and teach a lesson, while being to…
making gradients flow • founder - eXperiments lab