if this is true, we are entering a new era, latent reasoning is a fundamental architecture change and opening entirely new doors. it only makes sense that astra outputs half the tokens and performs high without CoT, when most of it happens in latent space. papers from over a year ago already show latent reasoning in its basic form outperforms classic architectures. it makes sense that Altman says future models will get much better if this is what's happening bts
new: OpenAI & others quietly using loop transformers that don't show their 'thinking' when scaled up
a leap forward on performance, but sparking concerns inside & outside OpenAI re: security as this takes off
This is huge: According to their no CoT analysis GPT-6 Sol was not a looped model and thus a different architecture than Astra entirely despite the 6. GPT-6.1 Sol though shares the same characteristics as Astra and is thus very likely looped
https://t.co/fNmf4hVocG
@corj1k this post is misleading. data center cooling does not inherently need water, and even those that do can recycle most of it. the thing that needs vastly (!) more water is the energy generation to run said site since that energy currently comes mostly from the grid