Trust Issues Newsletter Archive

Epoch 14 - November 9, 2023

Direct preference optimization and a study on LLM factuality. Plus the White House’s executive order on trustworthy AI.

Epoch 13 - October 12, 2023

Chain of density prompting and sliding window attention. Plus Mistral’s new SOTA model and summarization evaluation in TruLens.

Epoch 12 - September 14, 2023

Influence functions for explaining LLMs, automated valuations for reducing hallucinations in ChatGPT, OpenAI fine-tuning, and federal AI licensing.

Epoch 11 - August 17, 2023

Researchers jailbreak ChatGPT, the right to be forgotten, automatic generation of LLM attacks, and machine unlearning.

Epoch 10 - July 13, 2023

We do a deep dive into retrieval augmented generation along with a method for improving it called data importance evaluation. Also, the mode of LLM usage is changing with the rise of AI teaching assistants and the decline of ChatGPT usage.more efficient ways of training models and constructing augmented language model systems.

Epoch 9 - June 8, 2023

We explore more efficient ways of training models and constructing augmented language model systems. We also ponder upward and downward pressure on AI.

Epoch 8 - May 18, 2023

Scaling laws for LLMs, advancements in LLM adaptation, and Google/OpenAI have no moat unless regulations build them one.

Epoch 7 - Apr. 13, 2023

Emergent risks of generative AI, promptable image segmentation and Twitter’s open sourced recommender system - a red herring?

Epoch 6 - Mar. 9, 2023

How do vision transformers (ViTs) work and a new method for explaining them with ViT Shapley. Multimodal Language Models, and the reproducibility crisis reaches Principal Component Analysis (PCA). The FTC steps in on unjustified AI claims and OpenAI releases their gpt-3.5-turbo API.

Epoch 5 - Feb. 16, 2023

Tree-based models still outperform deep learning for tabular data, explainable AI is also a social challenge, and a failed attempt at classifying LLM-generated text.

Epoch 4 - Jan. 12, 2023

We touch on the alignment problem in deep learning, SparseGPT reached 50-60% sparsity with negligible accuracy loss, Microsoft announces OpenAI integration in Bing, and DoNotPay announces its AI legal assistant.

Epoch 3 - Dec. 8, 2022

The generative AI issue. We take on the societal implications of generative AI systems like ChatGPT and DALLE-2, an AI that can play Diplomacy, and touch on concept-based explanations.

Epoch 2 - Nov. 10, 2022

What makes an explainability method for NLP good, and how do you evaluate it? Plus, StackOverflow goes offline and the hype starts to pick up for generative AI.

Epoch 1 - Oct. 12, 2022

Robust models have smoother decision boundaries; enforcing that deep learning models are more interpretable during training can lead to benefits in performance; NYC mandates audits for AI in hiring, and the White House releases the AI Bill of Rights.