Tools

Search

Explore

Videos Channels Figures

Atmrix

About

Tools

Search

Explore

Videos Channels Figures

Atmrix

About

LF

Lex Fridman Cast

02/03/25

@ Lex Fridman

To effectively remove facts from a model, it must be done at every stage of the training process, including pre-training and post-training adjustments.

Video

LF

DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459

@ Lex Fridman

DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459

02/03/25

Related Takeaways

DS

David Silver

04/11/25

@ Google DeepMind

Removing human feedback from training may lead to models that are not grounded, as they would not learn from the actual consequences of their actions in the world.

LF

Lex Fridman Cast

02/03/25

@ Lex Fridman

Censorship or alignment can be embedded at various stages of the AI training pipeline, making it difficult to audit specific facts within the models.

SB

Sergey Brin

05/24/25

@ Google for Developers

The shift in model training is moving from a focus on pretraining to a greater emphasis on post-training, which is becoming a more substantial part of the overall training process.

O

OpenAI Cast

04/11/25

@ OpenAI

Training large models is challenging because new issues arise that weren't anticipated at the project's inception, requiring ongoing adjustments and problem-solving.

LF

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.

YC

Y Combinator Cast

09/22/23

@ Y Combinator

Fine-tuning AI models on new types of data is crucial to maintain accuracy, especially as trends and terminology evolve rapidly.

RH

Raza Habib

03/01/23

@ Y Combinator

Fine-tuning a model involves gathering examples of desired outputs for specific tasks and performing additional training on a base model to specialize it for those tasks.

LF

Lex Fridman Cast

02/03/25

@ Lex Fridman

The challenge of filtering out specific facts from AI models is compounded by the complexity of language and the potential for encoded language to evade detection.

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.