Method: train a sparse autoencoder on the activation on the residual stream. The sparsely activated components ensure only few features are activated for similar activation patterns in residual stream. Each of the feature is in turn interpreted by an LLM for its semantics. One can use these feature to semantically interpret the working of the model and steer the model towards desired goals.

A LLM uncensoring technique by finding the embedding direction of refusals in the residual stream outputs. One can choose to negate the refusal direction in the output to block the representation of refusals.

More on LLM steering by adding activation vectors: https://www.lesswrong.com/posts/5spBue2z2tw4JuDCx/steering-gpt-2-xl-by-adding-an-activation-vector

Chatbot Arena: blind test on the performance of LLM models to assign them a ELO rating. It also serves as a playground for various models.

A quick explanation on how speculative decoding works.

Explanation of the working of different parts in a transformer by manually coding the weights.

A comprehensive video tutorial on the Transformer architecture by Andrej Karpathy.

The author poses a sundry of text generating tasks aiming to test various abilities and evaluated them on various models. The result is posted on this site. It's a lot more intuitive than just a scoreboard.

A really interesting demo on visualizing on perplexity of a LLM. I knew about perplexity metric from the theory, where it's the average log-likelihood for the distribution of current tokens given previous tokens. The demo makes the idea more intuitive by showing the perplexity for each token and how they're calculated.

A visual and interactive explanation on how grokking (understanding) emerges in neuron networks.

What fascinates me is that the vast difference in neuron connectivity during memorizing and generalizing looks not unlike a human's brain in development.

A sundry of optimization techniques to transformer models to reduce the computation complexity associated with longer context.

A technical article on how to run large scale models efficiently on CPU.

An actually informative guide on prompt engineering. It talks about how to communicate format, prompt injection, how to prompt hack, etc.

What are the simple questions to ask that exploit the limitation of LLM? Interesting examples.

This article discusses conventional and newer fine tuning techniques for LLM. Keyword for search: prompt tuning, prefix tuning, in-context learning, frozen layers, adaptor, LoRA.

LLM prompting tricks and patterns.

A catalogue of AI tools and news.

A post on fine-tuning Alpaca with a lot of workable actions.

Visualize tokenization for OpenAI models.

“Quantity creates emergence. Simple elements, complex interactions, new patterns.” -- by Bing Chat

I am always fascinated by how interesting phenomenon seem to emerge suddenly in complex systems. The recent breakthrough in Large Language Models has brought me surprises after surprises.

This article aggregates a list of articles on emergent capabilities found in LLM.

A collection of prompts to jailbreak ChatGPT.