Call To Arms: ChatGPT challenged me -- I WILL CRUSH IT TO PIECES

Sascha

Dear Zettlers,

you might have heard it: ChatGPT challenged me and my honour. Let's see if I can keep it, or bow down in shame.

Here is a 100% authenticated record of a conversation I had yesterday (what a coincidence):

LUHMANN: What is best in life?
ME: To crush AI, see it driven before you, and to hear the lamentations of their creators.

See this thread for full context: https://forum.zettelkasten.de/discussion/2837/gpt4-zettel-critique-assistant-robo-zettel-feedback

---

The idea is to compare the feedback of ChatGPT and me. So, we need notes. 2--3 are enough. The content needs to be accessible to me. So, I cannot deal with mathematical content software code. Also: Really intricate details in specific domains might be out of reach. However, I have broad knowledge in the basics of most domains. So, just give it a try.

Please, submit notes for this battle.

Live long and prosper
Sascha

FernandoNobel

I always trust in the capabilities of mankind, but as artificial intelligence is a consequence of human intelligence: I don't know which I prefer to win :^)

Here is my note. I hope that the context of the note is accessible for this experiment.

# 202401221114 Don't make the mistake of defining your own convention
#convention #beginner's-error #scenio


Don't make the mistake of defining your own convention, if you want to be
useful, use your work to expand the community that uses that convention.


There are situations where it may be justifiable to work to:

* make a convention more efficient [[202401191051]]
* create a new convention.

However, this is not the norm: it is an exception that occurs very rarely. The
general rule is: 

  "Where you can add most value with your work is by creating and improving
  tools that use a convention. Making the convention itself more efficient is
  useless work most of the time".

In other words, use your work to increase the value of the community that uses
the covention [[202401191110]].

So don't make the mistake of defining your own convention to solve a problem in
your work, it is best to follow the community's convention. This idea also
applies to the fact that it is best to give value to a scenio [[202401221112]].

Example of conventions:

* Use of spaces or tabs to indent code [[202401191049]]]
* The USB protocol
* The chemical composition of automotive gasoline
* The voltage of the electrical grid
* Anything that can be done in many different ways, but one community defined
  a standard for how to do it.

---

Conversation with Borja about following conventions..............[[202401191017]]

phykas

Hi Sascha, this is a bit techical oriented note, but it's a good test to check if someone who is maybe not familiar with this topic can get a graps of what it's about.

---frontmatter--
title: 202403272114-binary-vector-search
aliases: Binary Vector Search
tags: work-related vector-search vector-database
date: 27-03-2024 21:14
author:
source: https://blog.pgvecto.rs/my-binary-vector-search-is-better-than-your-fp32-vectors
---frontmatter--

Binary Vector Search

Overview

• Description:

  • Binary Vector Search uses a simple but effective strategy to optimize memory and performance within a vector database

• Origin/Background:

  • Vector Search relies on vector embeddings which are a floating point numbers that look something like this [-0.000232, 0.0002342342, 0.0000234234 ....]
  • There are many language models that generate embeddings and they all have a certain amount of dimensions that is outputted in a form of an embedding eg. 356dims, 1536 dims etc.
  • Storing these vectors takes a lot of disk space and can take time to search through a vector database as they consume a lot of working memory

• Key features/steps/procedure:

  • The procedure is to take this vector [-0.000232, 0.0002342342, 0.0000234234 ....] and convert it to a binary form where if > 0 is stored as 1 else 0
  • Example: [0, 1, 1 ...]
  • This takes significantly less memory and increases throughput when doing search while lowering accuracy but still keeping a lot of semantic information in the embeddings

• Advantages:

  • Binary Vector Search reduces memory usage by 32 times
  • Binary Vector Search is a method of optimizing memory

• Drawbacks:

  • Binary Vector Search reduces accuracy of search

• Learning resources:

  • https://blog.pgvecto.rs/my-binary-vector-search-is-better-than-your-fp32-vectors
  • Vespa Colbert Embedder

Arguments

Binary Vector Search is a method for optimizing memory consumption in large Vector database

• Given that we have a problem of high memory consumption & slow performance in a large Vector Database
• When we use Binary Vector Search we will
• [Therefore] reduce working memory usage and increase performance

Binary Vector Search reduces working memory usage by 32 times

• Given that binary vector search uses integers instead of float32 to store it's data
• [And] that binary representation can be compressed to a single integer number
• When we use Binary Vector Search we will
• [Therefore] reduce memory by 32 times
• [And] increase vector search performance

Binary Vector Search reduces accuracy of search

• Given that binary vector search uses integers instead of float32
• When we use Binary Vector Search we will
• [Therefore] reduce accuracy of vector search
• [And] increase performance

Binary Vector Search accuracy can be enhanced by using Adaptive Binary Vector Search

• Given Binary Vector Search reduces accuracy of search
• When we use Adaptive Binary Vector Search we will
• [Therefore] increase Binary Vector Search accuracy