Member-only story
Some Are Terrified About OpenAI’s New AI. But Why?
Looped Transformers are now a — controversial — thing
With Anthropic’s release of Fable 5.1 yesterday, OpenAI is about to release Astra, a model that could shatter records and terrify many researchers.
But why? Well, because this release might be special: it's been leaked that the model is a ‘looped transformer’. This is a fascinating turn of events and, quite possibly, the biggest change in how models work over the last few years.
But the leak of this “new” architecture has mainstream media and many research scientists alike panicking, as some believe this could be a serious security risk, even a threat to humanity, to the point that OpenAI’s Chief Scientist had to step in publicly to clear up some confusion.
But why are people so scared?
This is undoubtedly the topic of the week (or, perhaps, the month), and there’s certainly a lot of confusion and misunderstanding, so let’s digest what is going on.
Scaling? Yes. How? Unclear.
You’ve probably heard (from me in the past, or from others) that the not-so-secret secret of AI progress is scaling laws.
But what are they?
