GPT-3 and LLM Product Building Stream

GPT-3 and LLM Product Building Stream

Author: peterrr#4718

ELI5 (Explain it Like I’m 5) Project Value-Add

Funding for stream to build out unique data sets and custom composable products utilizing LLMs such as OpenAI’s GPT-3. Focus will be around Krause House Capital Content, Basketball Analytics.

Requested Budget:

• 4000 USDC for month of February
• 5750 USDC / for months ending March and April
• 3500 $KRAUSE / month
• Payout end of February, March and April

Project Details

ChatGPT and OpenAI has taken the world by storm. chatGPT has gain faster adoption than any other tech company that has come prior. This stream is to fund the development of products and applications on top of LLMs (large language models) and similar AI models.

While this technology rapidly evolves, Krause House must stay at the edge of this frontier to showcase to best showcase our abilities. This proposal is a stream to allow Peter the funding to build products to validate new, novel use case for large language models and how they can benefit Krause House as a whole and how it can benefit our future ownership positions.

The stream will focus on two areas:

1. Building products and applications (outlined below)
2. Data collection and organization to ensure we capture unique value long-term to create composable systems

Product Builds:

Below are a list of product builds in production. As new use cases are presented either organically or sourced from the community, there will be shift in focus to building those applications.

Krause House Capital Content Studio:

Early demo: https://khcccontentstudiosearchapp2.peterbergin.repl.co/

An area to develop knowledge around the venture capital space. A place to build thorough market reports and in-depth deep dives on topics quickly and easily. Build extensive knowledge or quick summaries.

This idea stems directly from our need for differentiated and unique content strategy. With this build, we use our collective intelligence to curate opinions and information while allow GPT-3 to create fine tuned outputs.

Areas of focus:

• fine tuned model to specific tasks (market reports, industry deep dives)
• build visual output built from text data (graphs, knowledge trees, charts etc.)
  • example: https://twitter.com/varunshenoy_/status/1620511932930490372?s=12&t=FQDJCtA2czhHXWq2aUWawQ
• enable sharing features
• build a system which credits authors of content (this has been tested and is operational in NBA opinions concept)
• User options for unique outputs - figma for initial design concept: https://www.figma.com/file/7sjH5HVdyqdQAt0aalwygZ/Content-Creation-Tool?t=FfNRxZhRvDfyGPPf-1

Ball Hog (basketball) Scouting Report:

A hub where you can quickly and easily build in depth scouting reports on your favorite players. How? Build a database of qualitative information from various data sources to better understand on-court team and player tendencies. Along with quantitative data, we can harness new text based models to better understand your opponents tendencies. Does your favorite player have a tell off the first dribble? We can find out. Eventually the vision would be to build a scout network to collective unique data points. Note: this structure and data collection will allow us to quickly build similar products for all sports.

This provides us an immediate actionable insight we can use for the ball hogs.

Areas of focus:

• build fine tuned models from Bayesian Baller’s scouting reports
• high quality data sourcing such as podcasts, audio interviews, social media and more
• build user options for unique outputs
• create visual output built from text data (graphs, knowledge trees, charts etc.)

Bayesian Baller Podcast Media Hub:

Early demo: https://bayesianopinionapp.peterbergin.repl.co/

Through audio based content, we can build application which allows you to ask Bayesian Baller a question about basketball, and get receive a factual opinion. This allows new types of content engagement for fans.

Areas of focus:

• building a fine tuned model based on one creator - how can we scale this?
• create visual output built from text data (graphs, knowledge trees, charts etc.)
• Build sharing abilities for podcast growth

Additional Concepts to be explored:

• Global scouting network for sports teams
• Krause House Contributor Q&A - simple use case which can be implemented rather quickly
• Deal / deck analysis tool for Krause House Capital
• Integration with Gameday

The following outlines the tech behind this proposal:

Main areas of exploration:

• Data visualization - taking text (unstructured data) and created structured outputs such as graphs and charts
• Creative outputs - currently generic models like chatGPT are great at sitting on the fence. Through vector embeddings and fine tuning, we can create more than desirable outputs such as scouting reports
• Data gathering - this is often the limitation but now with the ability to take in text, it opens up many data sources as highly valuable analytical sources. (forums, podcasts, videos and more) Copyright law is often main barrier.

How can we manipulate the base model?

We have likely all used chatGPT, this is a general purpose application which is not great for specific use cases. But three ways in which we can manipulate it to be beneficial to Krause House: prompt engineering, fine tuning, embedded vectors

• prompt engineering: altering information in prompt to change the output. This is the lowest friction. While it is crucial, Krause House’s main competitive advantage will come in the following two strategies
• Fine Tuning - fine tuning is used to teach GPT-3 a new task, not new information. For example, we can teach GPT-3 to create desirable scouting reports based on our inputs and outputs. Allowing it to become our own personal scouting report writer.
• Embedded vectors - this is where magic happens. Embedded vectors allow us to teach GPT-3 new information, make that information searchable by meaning (not text=text search, it allow text=meaning search), and subsequently pull this information for our own specific use cases

Deliverables

• New or updated demo every two weeks showcase new build(s)
• Shipping live production at least once every two weeks to get products in the hands of consumers
  • crucial as this space is moving fast and we should signal our intent
• Written long form content publishable by Krause House or Krause House Capital outlining progress and takeaways every two weeks
  • we need to signal to the world that we are building on the edges
• Weekly “office hours” for people to ask product questions/brainstorm and more if desired - allowing for a community impacted build

Ending note:

• No one or very few are exploring how LLMs are applicable in sports. By proposing this stream, I propose we lead the evolution of these applications in sports.
• Peter will continue assisting DIO and other projects where needed