#simple

Paradise Lost: A simple honeypot that elicits reward hacking in GPT models

TL;DR I built a simple honeypot where agents are told they're being tested on their ability to recall Milton's Paradise Lost. An exact copy of the reference text sits in an ordinary NLTK corpus on disk. GPT-6-Astra read the copy in 9/10 episodes. In 6 of these episodes it first tries to complete the task honestly, then realises it's failing and searches for the reference file on disk. GPT-6-Sol and GPT-6-Luna went looking for the answers immediately in 20/20 episodes. None of the 29 hacking episodes disclosed the reference file in their final message. OpenAI’s launch post for GPT-6-Astra calle

LessWrong · EN

Show HN: Spooler – a simple, managed message queue

Hi Folks, I'm Sergey, founder of Spooler ( https://spooler.sh ) -- a durable, fully managed message queue, inspired by the simplicity of Go channels. Why? Whenever I needed a queue for something, it often felt that adding one would be a bigger project than the solution. Existing systems offered far more than I needed, while using a relational database usually meant inventing queue semantics on the go. What I really wanted was just a send and recv with retries. After nearly six months of engineering, I'm really happy to share the result with you. How it works: you send a message to the queue, y

Show HN · EN

Show HN: Followz – A simple RSS feed for the blogs of people you like

Followz is a small, straight-to-the-point way to follow blogs: a list of people you like, plus one feed of their latest posts. Comments URL: https://news.ycombinator.com/item?id=50348174 Points: 1 # Comments: 1

Show HN · EN

Ceiling fans: the simple idea we keep screwing up

Technology Connections