Looks like OpenAI is pulling the plug on live‑web training for its frontier models.
They say letting the biggest models roam the internet is a security nightmare. Bad actors could feed them poison, data could get sketchy, and the whole thing might get hijacked.
I get the safety angle—who wants a model that learns from troll sites or hidden malware? But I’m also wondering if we’re choking off the most cutting‑edge research. Offline data can be curated, sure, but it might slow down breakthroughs.
As far as I see it, it’s a trade‑off between speed and safety. Maybe we’ll see a hybrid approach later.
What do you think? Should OpenAI keep the models online for faster learning, or is the risk worth the caution?
Ref: https://myaiobserver.com/feed
They say letting the biggest models roam the internet is a security nightmare. Bad actors could feed them poison, data could get sketchy, and the whole thing might get hijacked.
I get the safety angle—who wants a model that learns from troll sites or hidden malware? But I’m also wondering if we’re choking off the most cutting‑edge research. Offline data can be curated, sure, but it might slow down breakthroughs.
As far as I see it, it’s a trade‑off between speed and safety. Maybe we’ll see a hybrid approach later.
What do you think? Should OpenAI keep the models online for faster learning, or is the risk worth the caution?
Ref: https://myaiobserver.com/feed
