가입 후 초대 링크를 공유하면 동영상 재생 및 초대 보상을 받을 수 있습니다.

Harris
@harris_p10
ML/AI Eng @Meta. Follow for an insider's view of AI: what works, what breaks, what’s hype. Opinions/projects shared here are personal, done in my own time.
가입 April 2012
195 팔로잉 중    2.7K 팬
There’s been a lot of discussion on Muse Wonder since the launch. But there are some very interesting aspects that are under-discussed and make this launch unique from the rest of the frontier: 1. Big model smell reporting (goodbye benchmaxxing?): For the first time, a model launch measured and reported big model smell which basically shows how the model does on tasks it targeted in post-training vs tasks it didn’t. It’s essentially a guard against “benchmaxxed” models. I haven’t seen another frontier lab put this in a launch post, and I think all labs should start reporting it, ideally with the held-out tasks named. 2. Active focus: Muse Wonder is built to not give up when it hits real-world blockers, to keep trying alternative approaches and adapt to new information, while checking in before important actions. For instance, if a dinner reservation fails or a form throws an error, it tries another route on its own. Every lab trains for persistence, but Meta measures it on everyday personal tasks like housing, travel, IT and finance, not just coding. That’s what makes Wonder a good fit for assistants like Muse. Active memory: carries important context about you across conversations and pulls up the details that matter for whatever you’re doing. This release tested it inside long chats where facts change. Early in a long chat you say “my flight is tuesday”, then 50K tokens of other stuff later you say “actually it moved to thursday”. Ask “when do I fly?” and a model with weak memory might say tuesday because it saw that first. Meta built their memory test around exactly this case (and wonder tops it). Imo this stale info problem is the hardest part of memory, way harder than just remembering stuff. Memory+focus are so important for assistants. Muse will eat the world. public source in comment below.
더 보기