precisely: o3 and gpt5t are great models, super smart and helpful for many things; but they love to talk in this ridiculously overcomplex, insanely terse, handwavy way. when it gets things right, it's awesome. when it confidently gets things wrong, it's infuriating.
Intel TDX unfortunately suffers from the exact same vulnerability as Scalable SGX. The underlying root cause is the lack of randomized encryption; using a static-adversary encryption scheme (XTS) rather than a dynamic-adversary one. The result is that plaintext-ciphertext mappings are unchanged at a fixed memory address.
While the choice of scheme might initially seem puzzling, it is due to a randomized encryption scheme requiring counters for each memory block, which has a prohibitive on-chip memory cost when scaling to hundreds of GBs of memory.
M2 Max processor.
I saw 60+ tok/s on short conversations, but it degraded to 30 tok/s as the conversation got longer.
Do you know what actually accounts for this slowdown? I don’t believe it was thermal throttling.
This has been my thinking as well.
Quite anti-human that a technology that improves automation and productivity works against common interest rather than for it.
> Never use random IVs with GCM; this breaks the authentication [2] [3]. Given the pitfalls of AES-GCM with respect to random nonces, you might prefer switching to XSalsa20+Poly1305. The advantage of XSalsa is it has an extended nonce length, so you can use random nonces without fear.
Those papers are a bit over my head. Could you please explain what's wrong with using random IVs here? What should we do instead (assuming we can only use GCM, and not switch to chacha)
Wow, that situation with Android phones looks DIRE.
The charts for Android look intriguing; are they claiming they can break into BFU (before first unlock) phones that have secure startup (i.e. pin at boot) enabled?