4 ms·

This seems highly relevant: https://arxiv.org/abs/2406.01506 https://arxiv.org/abs/2406.01506 > In this paper, we study the two foundational questions in this

by davedx 2y ago

This seems highly relevant: https://arxiv.org/abs/2406.01506 https://arxiv.org/abs/2406.01506 > In this paper, we study the two foundational questions in this area. First, how are categorical concepts, such as {'mammal', 'bird', 'reptile', 'fish'}, represented? Second, how are hierarchical relations between concepts encoded? For example, how is the fact that 'dog' is a kind of 'mammal' encoded? We show how to extend the linear representation hypothesis to answer these questions. We find a remarkably simple structure: simple categorical concepts are represented as simplices, hierarchically related concepts are orthogonal in a sense we make precise, and (in consequence) complex concepts are represented as polytopes constructed from direct sums of simplices, reflecting the hierarchical structure. Basically, LLM's already partially encode information as semantic graphs internally. With this it is less surprising that augmenting them with external knowledge graphs has a lower ROI.

CharlieDigital 2y ago
> Basically, LLM's already partially encode information as semantic graphs internally. There's an (underutilized?) technique here to take advantage of that internal graph: have the LLM tell you the related concepts first and then perform the RAG using not just the original concept, but the expanded set of related concepts. So: concept → [related concepts] → [[.. rag-rc1],[.. rag-rc2],[.. rag-rcn]] → summarize With GPTs prior to 4o, it would have been too slow to do this as a two-step process. With 4o and some of the higher throughput Llama3 based options (Together.ai, Fireworks.ai, Groq.com), a two-step fan-out RAG approach takes advantage of this internal graph and could probably yield similar gains in RAG without additional infrastructure (another datastore) nor data pre-processing to take advantage of a graph approach.
- davedx 2y ago
  Yup. Fascinating stuff really. Kind of like mnemonics for LLM's, if you squint a bit.
- neeleshs 2y ago
  Even with old GPT, if the summary is decent, it works reasonably well with even no RAG. We are a data management platform and allow users to build data pipelines around a data model. This is basically a DAG. We autogenerate documentation for these pipelines, using gpt4, and feed a summarized version of the data pipeline - expressed as graphviz dot file format in the prompt. gpt4 understands this format well, and seeminlgy understands the graph itself reasonably well! It performs poorly expressing the higher level intent of the pipeline, but tactical details are accurately documented. We are trying to push prompting itself more, before turning to RAG & finetuning
mkehrt 2y ago
As an GenAI skeptic, I think this is a very cool finding. My experience with AI tools is that they are complete bullshit artists. But to a large extent that's just a result of the way they are trained. If this description of how the data is structured is correct, it indicates that these programs do encode a real model about the world. Perhaps alternative ways of training these same models, or fixing the data afterwards, will result in more truthful models.

by davedx 2y ago

CharlieDigital 2y ago

> Basically, LLM's already partially encode information as semantic graphs internally. There's an (underutilized?) technique here to take advantage of that internal graph: have the LLM tell you the related concepts first and then perform the RAG using not just the original concept, but the expanded set of related concepts. So: concept → [related concepts] → [[.. rag-rc1],[.. rag-rc2],[.. rag-rcn]] → summarize With GPTs prior to 4o, it would have been too slow to do this as a two-step process. With 4o and some of the higher throughput Llama3 based options (Together.ai, Fireworks.ai, Groq.com), a two-step fan-out RAG approach takes advantage of this internal graph and could probably yield similar gains in RAG without additional infrastructure (another datastore) nor data pre-processing to take advantage of a graph approach.

davedx 2y ago
Yup. Fascinating stuff really. Kind of like mnemonics for LLM's, if you squint a bit.
neeleshs 2y ago
Even with old GPT, if the summary is decent, it works reasonably well with even no RAG. We are a data management platform and allow users to build data pipelines around a data model. This is basically a DAG. We autogenerate documentation for these pipelines, using gpt4, and feed a summarized version of the data pipeline - expressed as graphviz dot file format in the prompt. gpt4 understands this format well, and seeminlgy understands the graph itself reasonably well! It performs poorly expressing the higher level intent of the pipeline, but tactical details are accurately documented. We are trying to push prompting itself more, before turning to RAG & finetuning

mkehrt 2y ago

As an GenAI skeptic, I think this is a very cool finding. My experience with AI tools is that they are complete bullshit artists. But to a large extent that's just a result of the way they are trained. If this description of how the data is structured is correct, it indicates that these programs do encode a real model about the world. Perhaps alternative ways of training these same models, or fixing the data afterwards, will result in more truthful models.