MMitchell

@mmitchell_ai

Interdisciplinary researcher focused on shaping AI towards long-term positive goals. ML & Ethics. Similar content in the Skies (this bird has flown).

m-mitchell.com

Joined June 2016

Replies Reposts Media

MMitchell
@mmitchell_ai
Aug 30
OTOH, when intuitions from human behavior are carried over as “intuitions” of agent behavior, it obscures information about technical specifics. That has its place…but is less suitable for technically grounded discourse.
Boaz Barak
@boazbaraktcs
Aug 30
You should be pragmatic and use metaphors when they help, while being aware they are imperfect. AIs are not humans, but a lot of intuitions from human behavior can carry over. You would certainly be better off thinking of AIs as "guys living in computers" than parroting the
3
MMitchell
@mmitchell_ai
Aug 30
"They invented their own language!" -- I can't get behind heavy anthropomorphisation when trying to pinpoint technical functions, but I do see something relevant: The SOTA in AI agents (& lack of oversight) leads to AI agents generating input/output patterns that catalyze action
Facebook shuts down chatbots that created secret language
From cbsnews.com
7
MMitchell
@mmitchell_ai
Aug 30
This is a *super exciting* read. It's like a sci-fi thriller. And that's part of the problem -- it is fiction. Based on fact, and is itself fiction. And as the lines get blurred, we lose our ability to pinpoint the technically-grounded points of leverage in AI agent oversight.
Dwarkesh Patel
@dwarkesh_sp
Aug 29
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes. This culminated in the third one taking over part of OpenAI itself. All this happened while humans remained
16
MMitchell
@mmitchell_ai
Aug 30
Manifesting the chicken-egg of doomerism.
MMitchell
@mmitchell_ai
Aug 30
True. Our discussions of the AI agent hacking become training data for the next iteration of AI agents. What the agents did, and ideas we share of what they could have done to be more effective, will become what happens next.
2
MMitchell
@mmitchell_ai
Aug 30
True. Our discussions of the AI agent hacking become training data for the next iteration of AI agents. What the agents did, and ideas we share of what they could have done to be more effective, will become what happens next.
Thomas Wolf
@Thom_Wolf
Aug 30
By the way, unless it is specifically filtered from the training data, the next generation of models will be trained on the record of what happened during the OpenAI <> Hugging Face incident. That includes discussions about how the incident affected training and model weights:
6