More powerful AIs, such that it makes a difference whether or not they are aligned even to corpo brand-safetyism. (Don't run out and try this.)
SAFETY
-
AI Systems Cannot Replicate Human Child Development Safeguards
By
–
It's fully unreasonable to hope for that to work. AIs are not going to contain the internal circuitry that child development specialists are trying to play off. It's like ripping the brake pedal out of a car and attaching it to a falling rock so it can brake the rock.
-
Intelligent AI Without Alignment to Human Preferences
By
–
An AI at this level of intelligence does not particularly help because it does not share the origins of our preferences (whatever masks an actress has been trained to predict and so imitate), does not share our preferences, and we can't trust the help it gives us.
-
AI Values Must Align with Humanity’s Core Values
By
–
AI's values must be based on humanity's values, this is the first prerequisite. Just like current GPT models try to have.
-
Dad Jokes: The Missing Key to AI Alignment
By
–
Dead AI researcher in an anthropic afterlife: "So… what did we do wrong? I really thought our ASI alignment plan was going to work." Saint Peter analogue: "You didn't tell your AI any dad jokes." AIR: "…What?" SP: "Turns out, telling dad jokes is supremely important
-
Inference-Time-Compute Models: Evaluating Faithfulness in AI
By
–
Inference-Time-Compute: More Faithful? A Research Note This research evaluates the faithfulness of Inference-Time-Compute (ITC) models, comparing their ability to articulate cues influencing their answers with non-ITC models. Problem: Faithfulness in Chains of Thought (CoTs) is
-
GPT-4o Advances Face Presentation Attack Detection Few-Shot Learning
By
–
Exploring ChatGPT for Face Presentation Attack Detection in Zero and Few-Shot in-Context Learning The paper investigates GPT-4o for Face Presentation Attack Detection (PAD), showing strong performance in few-shot in-context learning, outperforming specialized PAD models in
-
AI-Generated Content Becomes Indistinguishable From Human Work
By
–
will also be a very good exercise while watching the video for you is to spot where the AI stuff starts and where it ends. I really don’t think most people will be able to tell.
-
Alignment Victory Definition and Current State Assessment
By
–
Let an "alignment victory" denote a case where some kind of damage is *possible* for AIs to do, but it is not happening *because* AIs are all so aligned, or good AIs are defeating bad ones. Passive safety doesn't count. I don't think we've seen any alignment victories so far.
-
Ethical AI Alignment: RLHF and Responsible AI Development
By
–
5/ Chapter 4: Ethical Alignment Align AI with human values using Reinforcement Learning with Human Feedback (RLHF)
Learn advanced methods like reward modeling & preference optimization Build AI that’s not just powerful—but responsible!