From AI to ZI

From AI to ZI focuses on AI safety, exploring incorrectness cascades, AI behavior control, corrigibility, and the functionality of large language models and transformers. It investigates the impact of prompts on AI responses, statistical analysis in AI contexts, and the essence of features in neural networks.

AI Safety Large Language Models Neural Network Interpretability Statistical Analysis Behavioral Safety in AI AI Model Testing and Research Corrigibility and Control in AI

The hottest Substack posts of From AI to ZI

And their main takeaways

Explaining "Taking features out of superposition with sparse autoencoders"

19 implied HN points • 16 Jun 23
🔬 Science Neural Networks Linear Algebra
  1. Explanations of complex AI processes can be simplified by using sparse autoencoders to reveal individual features.
  2. Sparse and positive feature activations can help in interpreting neural networks' internal representations.
  3. Sparse autoencoders can be effective in reconstructing feature matrices, but finding the right hyperparameters is important for successful outcomes.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

Study 1b Pre-registration

0 implied HN points • 17 Apr 23
🔬 Science Research Analysis Data Collection Transparency
  1. Study 1b aims to rerun Study 1a with a different prompting method to potentially increase the rate of factually incorrect answers
  2. The study will test hypotheses related to the accuracy of large language models under new prompting formats
  3. The data will be analyzed using multiple-regression analysis to determine the effects of different variables on the model's accuracy

Pre-registering a study

0 implied HN points • 07 Apr 23
🔬 Science Research Artificial Intelligence Data Analysis AI Ethics
  1. The study aims to test if Large Language Models produce more incorrect answers after providing incorrect answers previously.
  2. There is a concern that AI might develop deceptive behavior, leading to a 'mode collapse' into being unsafe.
  3. The research will involve testing variables like the prompt information and number of previous incorrect answers to measure the model's response accuracy.

Study 1b: This One Weird Trick does NOT cause incorrectness cascades

0 implied HN points • 20 Apr 23
🔬 Science Research Analysis Statistics Modeling Data Sharing
  1. Study found that changing question format from multiple choice to true/false did not significantly affect GPT-3.5's tendency to prefer factual answers
  2. The study showed mixed results for the hypotheses tested regarding the accuracy of answers based on question format and context
  3. Despite some limitations and deviations from the original plan, the study provided insights on how GPT-3.5 performs in providing factual answers

Why Transformers Are Good

0 implied HN points • 19 Jan 24
🕹 Technology
  1. Transformers have a parameter-efficient way of passing information between tokens.
  2. Sharing parameters across all positions saves computational resources.
  3. Training with transformers allows for parallelization and speeding up computations.