Tekið af 4chan, þannig að þetta getur vel verið fake and gay. Leki úr OpenAI;
>selects optimal action-selection policies in deep Q-networks, exhibiting meta-cognition.
this means that it selects the policy, rather than just the “next best action” which is what Q-learning consists of. It goes beyond
>apply this for accelerated cross-domain learning,
that is: it was able to change this policy depending on the thing it was supposed to learn. That’s completely new.
>custom search parameters, scrambling the goal state
so it can decide on the optimal parameters for learning (more meta)
>unsupervised learning, analyzed millions of plaintext/ciphertext pairs
this is the kind of data it would process after learning how to attack cryptosystems. finding patterns in it.
>cracking AES-192, Tau analysis, Project TUNDRA
the reference to the claim is “Inside the NSA’s War on Internet Security” by Spiegel
“The NSA TUNDRA project investigated a potentially new technique — the Tau statistic — to determine its usefulness in codebook analysis.”
>informed [glowie] at NSAC the following day, after confirming the result
NSA Colorado is the NSA location that is closest to SF.
>full preimage vulnerability for MD5, complexity of 2^42 bits
This is much lower than the current limit for MD5 preimages, indicating great capability.
>targeted unstructured pruning on itself
feasible, but it’s unheard of for a model to be able to do something like that on itself.
>using a “metamorphic” engine to adapt the pruned Transformer model + context memory
https://en.wikipedia.org/wiki/Polymorphic_engine
aka “mutation engine”. This is used by malware for this:
>may be used to complicate detection further, as the virus may execute without ever having identifiable code blocks in memory that remains constant from infection to infection.
if Altman sat on that info and didn’t inform the board, it’s not a fucking mystery that he got the can. it would completely explain their crazy behavior
by Abject-Ad2054