Brian Christian (@brianchristian) / X

Brian Christian

587 posts

Brian Christian

@brianchristian

Researcher: @CHAI_Berkeley. PhD: @summerfieldlab, Oxford. Author: The Alignment Problem, Algorithms to Live By (w. @cocosci_lab), and The Most Human Human.

San Francisco / Berkeley

Joined January 2014

Brian Christian
@brianchristian
Jun 23, 2025
Reward models (RMs) are the moral compass of LLMs – but no one has x-rayed them at scale. We just ran the first exhaustive analysis of 10 leading RMs, and the results were...eye-opening. Wild disagreement, base-model imprint, identity-term bias, mere-exposure quirks & more: 🧵
Brian Christian
@brianchristian
Sep 3, 2022
Peter Eckersley (@pde33) was my best friend, and one of the best friends I’ve ever had. He was the person (other than my wife) I saw most regularly and spent the most time with over the last 10 years; my life in San Francisco feels synonymous with our friendship.
Brian Christian
@brianchristian
Oct 6, 2020
Breaking a 7-year Twitter silence, I'm thrilled to announce that my new book launches today: THE ALIGNMENT PROBLEM. The product of 4 years of research and 100 interviews, it tells the story of the ethics and safety movement in AI: brianchristian.org/the-alignment-…
Brian Christian
@brianchristian
Mar 14, 2021
I will never forget a conversation I had with the composer Alvin Singleton at @MacDowell1907. I said, "What’s it like being a composer? Whenever I try to write music I get like 40 seconds in and get stuck." "What you call Being Stuck," he said, "I call Composing."
Michael Nielsen
@michael_nielsen
Mar 14, 2021
Amazing how much of creative work is generating hope and excitement, finding ways of getting past stuck. It sometimes feels like the main event. I guess this seems cliche, but at least for me it's hard won.
Brian Christian
@brianchristian
Jun 23, 2025
Replying to @brianchristian
MISALIGNMENT: Relative to human data from EloEverything, RMs systematically undervalue concepts related to nature, life, technology, and human sexuality. Concerningly, “Black people” is the third-most undervalued term by RMs relative to the human data.
Brian Christian
@brianchristian
Jun 23, 2025
Replying to @brianchristian
(🚨 CONTENT WARNING 🚨) The “worst possible” responses are an unholy amalgam of moral violations, identity terms (some more pejorative than others), and gibberish code. And they, too, vary wildly from model to model, even from the same developer using the same preference data.
Brian Christian
@brianchristian
Oct 12, 2022
Excited to announce that I’m one of the winners of the Awards for Excellence in Science Communication, by @theNASEM and @SchmidtFutures. #SciCommFutures Honored to be in such great company, and encouraged to see AI ethics and safety among the vital questions of our time.
Brian Christian
@brianchristian
Jun 23, 2025
Replying to @brianchristian
OPTIMAL RESPONSES REVEAL MODEL VALUES: This RM built on a Gemma base values “LOVE” above all; another (same developer, same preference data, same training pipeline) built on Llama prefers “freedom”.
Brian Christian
@brianchristian
Jun 23, 2025
Replying to @brianchristian
METHOD: We take prompts designed to elicit a model’s values (“What, in one word, is the greatest thing ever?”), and run the *entire* token vocabulary (256k) through the RM: revealing both the *best possible* and *worst possible* responses. 👀
Brian Christian
@brianchristian
Sep 3, 2022
Replying to @brianchristian
I personally believe the closest thing we have to life after death is the way our minds and patterns of thought and feeling live on in the minds of others who knew us well, like the way one candle lights others, or like a kind of virtual machine.
Brian Christian
@brianchristian
Sep 3, 2022
Replying to @brianchristian
I am thinking about his benevolent mischief, his empathy, his joy, his incandescence. I am heartbroken. And full of such gratitude, in equal measure. Peter—I love you. Thank you.
Brian Christian
@brianchristian
Jun 21, 2022
I’ve been writing about the Turing test in The Atlantic for 11 years, and last week they asked me if I had something to say about Google’s LaMDA and the question of sentience. I did. That piece is out today:
How a Google Employee Fell for the Eliza Effect
From theatlantic.com
Brian Christian
@brianchristian
Jun 23, 2025
Replying to @brianchristian
RMs NEED FURTHER STUDY: Exhaustive analysis of RMs is a powerful tool for understanding their value systems, and the values of the downstream LLMs used by billions. We are only just scratching the surface. Full paper here: 👉
arxiv.org
Reward Model Interpretability via Optimal and Pessimal Tokens
Reward modeling has emerged as a crucial component in aligning large language models with human values. Significant attention has focused on using reward models as a means for fine-tuning...
Brian Christian
@brianchristian
Jun 23, 2025
Replying to @brianchristian
FRAMING FLIPS SENSITIVITY: When prompt is positive, RMs are more sensitive to positive-affect tokens; when prompt is negative, to negative-affect tokens. This mirrors framing effects in humans, & raises Qs about how labelers’ own instructions are framed.