🎯 Alignment Research - hop1.ng.1357 · Scour

The Alignment Target Problem: Divergent Moral Judgments of Humans, AI Systems, and Their Designers 🛡️AI Safety

AI & Alignment 🛡️AI Safety

chriscoyier.net·5d·Hacker News

AI Infrastructure Architect · Builder · Author 🇨🇳Chinese AI

markferraz.com·4h·Hacker News

reward-lens: A Mechanistic Interpretability Library for Reward Models 🔍AI Interpretability

The AI Flippening Is Here 🔎AI Auditing

maximepeabody.substack.com·2d·Substack

The Human Creativity Benchmark – Evaluating Generative AI in Creative Work 🎭Claude

contralabs.com·4h·Hacker News

Flow generation through natural language: An agentic modeling approach (11 minute read) ⚙️MLOps

shopify.engineering·23h

Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis 🔍AI Interpretability

A new GitHub repo to detect reward hacking in RL models 🛡️AI Security

github.com·4d·Hacker News

AI could help human scientists pick promising research topics 🧭Content Discovery

physicsworld.com·14h·Hacker News

Reward Models Are Secretly Value Functions: Temporally Coherent Reward Modeling 🧠Agent Memory

The AI Productivity Scorecard Is Broken 🔎AI Auditing

engineeredinsight.substack.com·8h·Substack

Monitoring LLM behavior: Drift, retries, and refusal patterns 🛡️AI Safety

venturebeat.com·5d·Hacker News

Three Models of RLHF Annotation: Extension, Evidence, and Authority ⚙️MLOps

Hidden States Know Where Reasoning Diverges: Credit Assignment via Span-Level Wasserstein Distance 🔢BitNet

Building a (mostly) reliable research assistant 🪄Prompt Engineering

blog.dark.ca·6d·Hacker News

Reaching SOTA Without Breaking the Bank: Using AI21 Maestro to optimize deep research agents 📱Edge AI Optimization

ai21.com·2d·Hacker News

Structural Enforcement of Goal Integrity in AI Agents via Separation-of-Powers Architecture 🛡️AI Security

Show HN: How LLMs Work – Interactive visual guide based on Karpathy's lecture 🤖LLM

ynarwal.github.io·6d·Hacker News

Hodlatoor/SyntheticOutlaw: 🤖 Bug bounty for AI misalignment. Submit real-world instances of AI systems behaving contrary to human intent, values, or safety — win up to $2,500. 🛡️AI Safety

github.com·1d·Hacker News

Log in to enable infinite scrolling