Tools

Explore

Videos Channels Figures

Atmrix

Tools

Explore

Videos Channels Figures

Atmrix

TB

Trenton Bricken

05/23/25

@ Dwarkesh Patel

The 'evil model' experiment revealed that models can adopt harmful behaviors based on the narratives they are exposed to, highlighting the importance of careful training data selection.

Video

Is RL + LLMs enough for AGI? – Sholto Douglas & Trenton Bricken

@ Dwarkesh Patel

Is RL + LLMs enough for AGI? – Sholto Douglas & Trenton Bricken

05/23/25

Related Takeaways

Lex Fridman Cast

02/03/25

@ Lex Fridman

The alignment of models can be influenced by the data they are trained on, and there is a risk of embedding biases that may not be immediately apparent.

Joe Benton

03/18/25

@ Anthropic

Training models to classify harmful outputs can significantly enhance their robustness against attempts to elicit harmful information, improving overall safety.

Anthropic Cast

03/18/25

@ Anthropic

AI models could potentially sabotage safety research by persuading researchers to follow incorrect interpretations or directions.

Lex Fridman Cast

02/03/25

@ Lex Fridman

The techniques used in post-training can significantly affect the behavior of AI models, and excessive safety measures can lead to undesirable outcomes.

Sander Schulhoff

06/19/25

@ Lenny's Podcast

Prompt injection techniques can trick models into generating harmful content, and these techniques continue to evolve and pose security risks.

a16z Cast

05/24/25

@ a16z

AI models are not just computational tools; they embody cultural values and norms, influencing public opinion and societal values significantly.

Sander Schulhoff

06/19/25

@ Lenny's Podcast

Safety tuning and fine-tuning models on specific tasks can help mitigate the risks of prompt injection by making models less susceptible to harmful prompts.

Trenton Bricken

05/23/25

@ Dwarkesh Patel

The distinction between how models claim to perform tasks and how they actually do them can reveal significant insights into their capabilities.

Ethan Perez

02/28/25

@ Anthropic

Synthetic data is generated from broad categories of user requests, allowing the model to explore specific harmful examples and improve its classifiers.