# An established technique for teaching optimized vector paths?

**URL:** <https://community.openai.com/t/an-established-technique-for-teaching-optimized-vector-paths/202926>\
**Category:** Community\
**Tags:** embeddings\
**Created:** [May 11, 2023, 11:59am UTC](https://community.openai.com/t/an-established-technique-for-teaching-optimized-vector-paths/202926 "2023-05-11T11:59:28Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![Autometa](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/autometa/32/34145_2.png) [@Autometa](https://community.openai.com/u/Autometa)\
**Post date:** [May 11, 2023, 11:59am UTC](https://community.openai.com/t/an-established-technique-for-teaching-optimized-vector-paths/202926/1 "2023-05-11T11:59:28Z")

</div>

im working on a project and i was wondering if there were any ‘best practices’ or established techniques to do this that i could measure my own against, forgive me if this is a dumb premise im self taught and its for my own personal entertainment; and until i get lucky enough to afford some formal teaching these projects are probably about as close as ill get. thanks for taking the time!

im pretraining a model with sentence embeddings represented in the data per sentence i was hoping to be able to perform an optimization technique on the total “tensor” of embeddings over a varying window of text related to its semantic structure and then teach the model to calculate true embedding positions based on just the optimized pseudo-tensor representation.

overall the idea is to create a model capable of extracting enough information out of a vector map to use the recent [data](https://openai.com/research/language-models-can-explain-neurons-in-language-models) Openai published to work on creating a ‘trained’ weight mqp based on wanted behaviors. i know its a lofty thing to shoot for and i dont really expect to be able to do that, but my primary motivation has been to contibute towards dealing with alignment as much as i can and to me this seems like a very clear and obvious trail to follow if the goal is to solve explainability and alignment.

---

<div class="post-metadata">

**Author:** ![bill.french](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/bill.french/32/15015_2.png) [@bill.french](https://community.openai.com/u/bill.french)\
**Post date:** [May 11, 2023, 2:14pm UTC](https://community.openai.com/t/an-established-technique-for-teaching-optimized-vector-paths/202926/2 "2023-05-11T14:14:32Z")

</div>

Hi, and welcome to the community.

> [@Autometa](#):
>
> dealing with alignment as much as i can and to me this seems like a very clear and obvious trail to follow if the goal is to solve explainability and alignment.

This is a really good question. I don’t have an answer, but in a very limited experiment using an FAQ as an example, I assumed the embeddings that I needed to generate had to have as much related content as possible to provide a broader vector. I was wrong - the embeddings needed to be discrete and focused. The more each embedding was about a key idea, the better the similarity matching worked. It was my duh moment. 😉