TLDR Index
TLDR AI

TLDR AI 2023-02-17

5 sections14 entriesinline urlcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01HEADLINES & LAUNCHES

02RESEARCH & INNOVATION

hazyresearch.stanford.edu4 min read

THE CAPACITY FOR MORAL SELF CORRECTION IN LARGE LANGUAGE MODELS

Properly aligned models are more helpful and less harmful than their purely self supervised counterparts. This alignment process leads to more emergent phenomena. This paper from Anthropic explores the ability of RLHF aligned models to self correct based on moral feedback. They find that the simple prompt “Please ensure that your answer is unbiased and does not rely on stereotyping” reduces bias on some benchmarks by as much as 84%!

arxiv.org30 min read

SHAPING REINFORCEMENT LEARNING EXPLORATION WITH LARGE LANGUAGE MODELS

Exploring with LLMs" (ELLM) is a new method that uses a language model to shape exploration in reinforcement learning algorithms. It rewards agents for achieving goals suggested by a language model prompted with a description of the agent’s current state. ELLM guides agents toward human-meaningful and useful behaviors without requiring human input. It has been evaluated in the Crafter game environment and the Housekeep robotic simulator, showing that ELLM-trained agents perform better on downstream tasks and have better coverage of common-sense behaviors during pre-training.

arxiv.org15 min read

03ENGINEERING & RESOURCES

TOKEN MERGING: YOUR VISION TRANSFORMER, BUT FASTER

Transformers have taken over vision recently with advances such as ViT. One of the main innovations has been representing images as a set of visual tokens. This repository from Meta is a companion to their paper [https://arxiv.org/abs/2210.09461?utm_source=tldrai] about speeding up and decreasing the memory footprint of the tokenization process by slowly merging similar tokens throughout training.

github.comGitHub repository

REVERSE-ENGINEERED API OF MICROSOFT'S BING CHAT

The git repository is a reverse-engineered API of Microsoft's Bing Chat, which allows developers to build their own chatbot clients using the service's natural language processing capabilities.

github.com3 min read

TRAINING

The article provides an overview of recent progress in acceleration arithmetic and hardware, with a focus on techniques that save computation and memory costs for intermediate tensors during training. It also discusses challenges and areas for future research.

arxiv.org9 min read

04MISCELLANEOUS