---
tags: lluf-prompt
---

# lluf-prompt: ai, gpt-4, and steamship

## the idea, the pitch

[link for submitting your work to this prompt](https://airtable.com/shrMZUvuNlIgPffpV)

ai is going to be a thing.

## the steps
### step 1: go to [this event] and take notes!(https://docs.google.com/forms/d/e/1FAIpQLSe0tRQoJeXZoTHkKrxKCGMPYyCNZPLxOczDIYp4jK67FXCipw/viewform).
the notes:

### step 2. make something using what you've gleaned from this event.
You are encouraged to organize (or get madeleine and jordan's help organizing) a group of folks who have various expertise if you want to work collectively on this!

### Divided into 2 Key Questions:

## What is GPT?

#### Many adequate ways to describe:
* Large Language Model
* Universal Approximator
* Generative AI
* Neural Network
* Artifical Intelligence
* Simulator
* Text Predictor
* Writing Assistant
* Content Generator
* Agent
* Chatbot
* Transformer

#### For our purposes, most accurate is language model
* GPT centers on probability distribution over some vocabulary
* Samples predicted word repeatedly to generate text, assigning each word (out of 50,000) a probability and selects most probable response
* Regenerates with each sample

#### What makes GPT-4 special is its ability to answer questions
* Combines knowledge of Internet and ability to solve a problem
* More training makes more accurate
* ChatGPT beat Tik Tok's record for most number of users in first month
* Concept of reinforcement alignment with human feedback
* Instruction tuning

## How are People Building AI Apps?
Future prediction: it will be built into softwares and we will have networks which communicte with it

#### Multiple Functions
* Companionship: can be a friend with a purpose (ex: tell me when/how to do xyz)
* Question-Answering (ex: answer this homework question for me)
* * user queries GPT, GPT grabs relevant documents, 'cuts' it up (window pulls out text fragment), transforms into **embedding vector** (chunks of text in vectors generating meaning), enters vector database, information passes to prompt, AI is trained to answer based on its personality
* * once in software, this process can just repeat
* * key idea: tactical rearrangement of prompts
* Utility (ex: read all of these Tweets and only show me ones I would find relevant)
* * automating tasks that can be context-free
* * user provides context
* * basic building blocks of labor could change
* Creativity: can generate text and images
* * general creative process: have an idea, overgenerate possibilities, edit and trim ideas to relevant ones, repeat
* * AI is fantastic at possibility stage because you are pre-agreeing that it can be wrong (AI searches in a database and pulls relevant pieces for you to decide on)
* Number of uses are still experimental, but will soon be more usable

#### BabyAGI, AutoGPT
* Multi-step planning bots
* Essentially GPT in a for-loop, telling itself what to do (self direction)
* Emergent behavior
* Currently still experimental but will unfold
* Means that Python could become less relevant and more important skill will be prompt-engineering

#### Terminology
* Agent: GPT and its bigger body that it lives in
* Tools: ways Agent can do things
* Twitter: excellent source of inspiration (many leadinf AI-builders share work)

#### Big Question: How will this Become Common?
* Answer: prompts!!
* For example: tell GPT-4 to talk to you like a kind teacher for 5 mins (technology responds as you tell it to)
* Soon, all will be within reach for everybody

## Q & A Session Summary
* "Hallucination": term for when AI not always outputting accurate information from lens of expert
* * but synthesis is part of game
* * if we develop technology that more closely mimics human ability to synthesize, then consequence is the same aggressive assumptiuon for knowledge
* * AI model has nothing to rest on to determine if something is true or false (parallel: someone learning English but never actually interacting with speakers)
* * a part of architecture, but more training can help, along with hard fixes like asking technology to provide its **"best guess"** or telling it you think it made a mistake somewhere
* * domain-specific questions more prone to phenomena

### step 2: gather folks to work on a prototype

## LLUF NOTES/TIPS to make it better/clearer: