🧵

Paul Christiano

Original namePaul Christiano

American Computer Scientist, Pioneer of AI Alignment Research

The Names Remembered · Small Roles, Famous Names

Early Research on Reinforcement Learning From Human PreferencesThe Alignment Research Center (ARC)AI Alignment Theory

Who they are

Paul Christiano is an American computer scientist, formerly at OpenAI, who in 2017 with his team co-authored a paper proposing the method of letting human evaluators rank their preference among a model’s several candidate outputs and training the model accordingly, an early research thereafter further developed into one of the important theoretical bases of the RLHF training paradigm widely applied in products such as ChatGPT. Christiano thereafter founded the Alignment Research Center (ARC), continuing to focus on the question of how to design AI systems whose behavioral goals stay consistent with humanity’s true intent, and he is also one of the earlier researchers to systematically think about the practical methodology of "how to evaluate the potentially dangerous capabilities of frontier models."

Primary sourcesChristiano et al., "Deep Reinforcement Learning from Human Preferences" (2017)

Key stories

A Paper That Became the Theoretical Starting Point of ChatGPT’s "Obedience"

The "training a model from human preference" method Christiano proposed in 2017 had at first a relatively simple experimental scene, involving only some basic simulated control tasks, but this core idea of "letting humans rank the model’s outputs by preference and then adjusting the model’s behavior accordingly" was thereafter further developed and extended by researchers such as John Schulman to the alignment training of large language models, becoming one of the key theoretical sources behind ChatGPT’s better ability to "understand human speech" over earlier models.

Relationships

Disciple

🎛️John Schulman

Ally

🐘Connor Leahy

Echoes today

Below are how modern works borrow or reinterpret this name or story — not the original material. The two differ, so keep them apart.

The theoretical source of the RLHF methodThis paper, at first involving only simple simulated control tasks, was thereafter proven to be one of the most important theoretical sources behind the ability of conversational large language models such as ChatGPT to "understand human speech."

Appears in

Algorithmic Bias and the Awakening of AI EthicsThe Early Research of RLHF

← Back to Figures