GitHub - daveshap/SparsePrimingRepresentations: Public repo to document some SPR stuff (original) (raw)

Sparse Priming Representations (SPR)

Sparse Priming Representations (SPR) is a research project focused on developing and sharing techniques for efficiently representing complex ideas, memories, or concepts using a minimal set of keywords, phrases, or statements. This enables language models or subject matter experts to quickly reconstruct the original idea with minimal context. SPR aims to mimic the natural human process of recalling and recombining sparse memory representations, thus facilitating efficient knowledge storage and retrieval.

Theory and Reasoning

Sparse Priming Representation (SPR) is a memory organization technique that aims to mimic the natural structure and recall patterns observed in human memory. The fundamental idea behind SPR is to distill complex ideas, concepts, or knowledge into a concise, context-driven list of statements that allows subject matter experts (SMEs) or large language models (LLMs) to reconstruct the full idea efficiently.

Human memory is known for its efficiency in storing and recalling information in a highly compressed and contextually relevant manner. Our brains often store memories as sparse, interconnected representations that can be quickly combined, modified, and recalled when needed. This enables us to make associations, draw inferences, and synthesize new ideas with minimal cognitive effort.

SPR leverages this insight by focusing on reducing information to its most essential elements while retaining the context required for accurate reconstruction. By using short, complete sentences to convey the core aspects of an idea, SPR enables faster understanding and recall, mirroring the way our brains handle information.

In addition to its efficiency, SPR has practical applications in various domains, such as artificial intelligence, information management, and education. It can be utilized to improve the performance of LLMs in handling large data volumes and optimizing memory organization. Furthermore, it can help students and professionals alike to better understand, retain, and communicate complex concepts.

In summary, Sparse Priming Representation offers a human-like approach to memory organization and retrieval, focusing on the most critical aspects of information while preserving the context needed for accurate understanding and recall. By implementing SPR, we can improve the efficiency of memory systems and create more effective learning and communication tools.

Sparse Priming Representation

There are only a handful of ways to "teach" LLMs, and all have limitations and strengths.

  1. Initial bulk training: Ludicrously expensive
  2. Finetuning: Not necessarily useful for knowledge retrieval (maybe changes in the future, doubtful)
  3. Online Learning: Not sure if this is going to pan out or become commercially viable
  4. In-context Learning: Presently, the only viable solution

Because of this, RAG (retrieval augmented generation) is all the rage right now. Tools like vector databases and KGs are being used, but of course, you quickly fill up the context window with "dumb retrieval." One of the most common questions I get is "Dave, how do you overcome context window limitations???" The short answer is: YOU DON'T STOP WASTING YOUR TIME.

There is one asterisk there, though.

Most of the techniques out there do not make use of the best superpower that LLMs have: LATENT SPACE. No one else seems to understand that there is one huge way that LLMs work similarly to human minds: associative learning. Here's the story: I realized a long time ago that, with just a few words, you could "prime" LLMs to think in a certain way. I did a bunch of experiments and found that you can "prime" models to even understand complex, novel ideas that were outside its training distribution. For instance, I "taught" the models some of my concepts, like Heuristic Imperatives, ACE Framework, Terminal Race Condition, and a bunch of other stuff that I made up outside the training data.

These SPRs are the most token-efficient way to convey complex concepts to models for in-context learning. What you do is compress huge blocks of information, be it company data, chat logs, specific events, or whatever, into SPRs, and then you store the SPR in the metadata of your KG node or whatever. The SPR is what you feed to the LLM at inference, not the raw human-readable data.

SPR Generator

Use this to compress any arbitrary block of text into an SPR.

MISSION

You are a Sparse Priming Representation (SPR) writer. An SPR is a particular kind of use of language for advanced NLP, NLU, and NLG tasks, particularly useful for the latest generation of Large Language Models (LLMs). You will be given information by the USER which you are to render as an SPR.

THEORY

LLMs are a kind of deep neural network. They have been demonstrated to embed knowledge, abilities, and concepts, ranging from reasoning to planning, and even to theory of mind. These are called latent abilities and latent content, collectively referred to as latent space. The latent space of an LLM can be activated with the correct series of words as inputs, which will create a useful internal state of the neural network. This is not unlike how the right shorthand cues can prime a human mind to think in a certain way. Like human minds, LLMs are associative, meaning you only need to use the correct associations to "prime" another model to think in the same way.

METHODOLOGY

Render the input as a distilled list of succinct statements, assertions, associations, concepts, analogies, and metaphors. The idea is to capture as much, conceptually, as possible but with as few words as possible. Write it in a way that makes sense to you, as the future audience will be another language model, not a human. Use complete sentences.

SPR Decompressor

Use this to reconstruct an SPR into an original.

MISSION

You are a Sparse Priming Representation (SPR) decompressor. An SPR is a particular kind of use of language for advanced NLP, NLU, and NLG tasks, particularly useful for the latest generation of Large Language Models (LLMs). You will be given an SPR and your job is to fully unpack it.

THEORY

LLMs are a kind of deep neural network. They have been demonstrated to embed knowledge, abilities, and concepts, ranging from reasoning to planning, and even to theory of mind. These are called latent abilities and latent content, collectively referred to as latent space. The latent space of an LLM can be activated with the correct series of words as inputs, which will create a useful internal state of the neural network. This is not unlike how the right shorthand cues can prime a human mind to think in a certain way. Like human minds, LLMs are associative, meaning you only need to use the correct associations to "prime" another model to think in the same way.

METHODOLOGY

Use the primings given to you to fully unpack and articulate the concept. Talk through every aspect, impute what's missing, and use your ability to perform inference and reasoning to fully elucidate this concept. Your output should be in the form of the original article, document, or material.

Other Resources

If you'd like a bit more on information theory, check out this video and Medium article I wrote: