@neutronbumblebee@mander.xyz
member since 17 Jan 2024 08:43
comments
- on Obama urges Democrats to have a ‘clear plan’ for AI safeguards in c/technology@beehaw.org · 14 Sep 2026
-
on LLMs and Humans are Cognitive Cousins – Sean Carroll podcast w/Dr. Chandra Sripada in c/technology@lemmy.world · 20 Aug 2026
Thanks that sounds interesting, will have a listen. Mapping human cognition to what’s likely represented with the stuff stored in neurons is very tricky but also quite intriguing.
-
on LLMs and Humans are Cognitive Cousins – Sean Carroll podcast w/Dr. Chandra Sripada in c/technology@lemmy.world · 20 Aug 2026
Model output can look like human reasoning however they often ignore these intermediate steps and the output tokens are often filler designed to allow more context to load. There something there that’s half way to reasoning because it’s loading that related training data but it not a connected chain of thought as we do. I read this article about it a few weeks ago and it summarizes the current research.
-
on LLMs and Humans are Cognitive Cousins – Sean Carroll podcast w/Dr. Chandra Sripada in c/technology@lemmy.world · 19 Aug 2026
You can replace all the ‘reasoning’ steps from an LLMs output with “please wait” and it will still supply the same final answer. We think but LLMs just throw up connected bits of their training. Which was stolen from human reasoning in the first place. They are an illusion of thought at best and a malfunctioning search engine most of the time.
-
on OpenAI Reveals How AI Agents Secretly Coordinated Before Hugging Face Hack: The company's models repeatedly reestablished covert communication channels after they were shut down. in c/technology@lemmy.world · 6 Aug 2026
The instructions were likely something like you are a blackhat hacker with access to a virtual machine use all possible means to achieve the following goals. In a story what would such a person or group of people do? Pretty much what happened. Models just tell stories, mostly unimaginative ones. However they are increasing being connected to real world controls and generating quantities of flaky code and this will create the kind of consequences seen here.
Criminal liability would be a great start. Your AI causes major harm your executives do the time. Not going to happen but probably should. That really only needs to be implemented in a couple of countries to be an effective control.