yt-semantic-search VS client-vector-search

The dev LLM stack

- OpenAI + Pinecone + GPT-Index or Langchain

- Perhaps also dust.tt for playing around with prompts, kinda like a more advanced gpt playground --

The production LLM stack

- The dev stack

- OpenAI + Pinecone + GPT-Index or Langchain

- arXiv for finding new research to build on

- Prompt platforms such as Humanloop

- ML frameworks such as PyTorch, Keras, Tensorflow

- MLOps tools such as MLflow, Kubeflow, Metaflow, Airflow, Seldon Core, TFServing

Example OpenAI Projects

-

-

-

-

-

What OpenAI/MSFT should do

- Fund "AI white mirror" -- a tv show that has beautiful visions a future where intelligence costs ~0

Things you'll probably discover

- Embeddings work ok, but not great, from a user perspective. As a developer they're great to work with. As a user, the results aren't ranked quite right. Embeddings use cases will be better with GPT-4 or GPT-4.5.

- All of the obvious gpt apps will be built. We'll get hundreds of basic gpt wrapper apps (and some of them will be big businesses!), hundreds of basic embeddings search apps. If someone can think of the idea and make it without needing specific relationships, credibility, or experience, then it'll probably exist by Summer 2023.

- The developer energy in this space is intense. Adults are going to hackathons to build ai apps. This is awesome.

- Devs using gpt will soon be a large enough market that startups will exist and succeed just by selling to developers that are using gpt-3 in production. We already saw it a little bit, but we'll get many more startups here.

- How could AI not be better than me at all computer based things within 10 years?

- AI is kinda like a kid. When they're young, they're not that smart. Then all of a sudden, they've gotten enough training data, and their brain (compute!) has grown, and they're doing useful stuff. This is related to why people will say that building models can feel frustrating because it doesn't work well for ages and then all of a sudden it works (CEO of Oasis said this, CTO of OpenAI said this, and Instagram co-founder said this).

Would love input and feedback on this. I have similar things that I'm going to submit, covering what builders and engineers should do, what vector database to use, why no one else made ChatGPT before OpenAI, things holding ai powered apps back, and some other stuff like that. If you want a preview and are happy to give feedback, then email is in my profile.

Are you using whisper for transcription?

For https://github.com/transitive-bullshit/yt-semantic-search, I'm using YouTube's built-in transcriptions which definitely aren't as high quality, but they work well enough to power the semantic search.

1. easy setup: you only need to add 5 lines of code to build a semantic search

2. no embedding api needed: you don't need an api and have to pay for it unless ure scaling up millions

3. faster search: modern hardware is better than cheap cloud computers (0.5vCPUs)

4. zero latency: no back-and-forth with server-side

5. easy integration with our api if you scale fast: 10x cheaper than OpenAI and you don't have to pay for Pinecone et. al.

play with it: https://clientvectorsearch.com

install the library: https://www.npmjs.com/package/client-vector-search

check our repo: https://github.com/yusufhilmi/client-vector-search

find us on twitter, hit us up with questions:

- https://twitter.com/yusufhilmi_

- https://twitter.com/karmedge