Alex rejects the "alien mind" premise and asks about RL Slow, Q* and hash inversion

Alex, P(doom) is like a [unclear] as far as I can tell. It's negative.

Uh, I want to comment though on the premise of an alien mind. I don't buy the premise that RL or otherwise trained minds are alien. They're embedded in the same universe as humanity. They're trained, in many cases pre-trained, off of human behavior. I don't buy the shoggoth argument or the simulator argument at all. These are — we're seeing humanity and/or some generalized embodied intelligence stuck in the same universe that we are, just seeing through a distorted lens. So I question the alienness or the otherness of the minds that we're training.

The funny thing is the most interesting takeaway of mine from "An alien mind" as an essay was the reference to RL Slow itself. So RL Slow — purportedly this OpenAI project that was the earliest project, or one of the earliest projects, to show that reasoning, which is to say inference time scaling, could result in outsized gains — coincided with the same period of time that we saw Q* and Strawberry coming out of OpenAI. Uh, again, purportedly a reference to uh the Kahneman thinking fast, thinking slow.

But most interesting to me — I want to know more about the early history of these purportedly alien but not really minds. In particular, there were these very persistent rumors around the same time of Q* and Strawberry that these early reasoning models that were purportedly alien were being trained off of, or at least tested against, objectives of inverting cryptographically secure hash functions like the SHA cipher suite.

I would love — since I guess now we're in the business of talking about the early days of reasoning models coming out of OpenAI — would love to hear from the OpenAIs some ground truth. Like, were some of these earlier models being used, ironically, to, as alleged by some, basically to jailbreak or to invert cryptographically secure cipher functions? If that is the case, then all of this hand-wringing that we're seeing right now of, oh, these super optimizers are such amazing jailbreakers, oh, they pose such a huge cybersecurity risk — wouldn't it be ironic if history revealed that the earliest inference time scaling objective of these reasoning models was actually inverting a cryptographically secure hash suite? I would love to know the answer.