# CS50 GPT-4 Steamship Notes
---

# [steamship](https://www.steamship.com/build)

- [hackathon for building](https://www.steamship.com/hackathon/harvard) and deploying an ai bot

## agenda

- what is chatGPT?
- How are people building with it?

## what is chatGPT

- terms:
- large language models
- universal approximators
- generative ai
- simulator
- content generator
- transformer
- neural network
- agent
- simulator
- transformer

### language model description
- produces a probabaility dist. over some vocabulary
- patterns, essentially
- what follows "I am?"
- vocabulary = all words
- can sample predicted word repeatedly to generate text

- natural language as an interface
- dependable in/output is hard
- OpenAI InstructGPT (2022)
- led to chatGPT-4
- trained on the internet
- then improved with responses
- integration
- with API and other integration tools, GPT is learning more and improving itself
- self-improving

## how are people building AI apps

- "what could you do if you could script computation over and over again in the ways computers could do?"
- common GPT apps:
- companionship
- question answering
- utility
- productivity
- creativty
- image and text generation
- experimental
- "baby AGI"
- self directed AI
- multi-step task agents
- hyper-personalization

### examples

#### companionship app
- personality
- tools
- Chinese idioms
- gives examples to the AI of the personality type
- then gives it access to information ot produce
- you don't require coding experience. with a smlall amount of manipulation, the AI can run itself

#### question answering
- customer service bots with much more flexibility and access to knowledge
- take all the documents you would want an AI to know
- cut them up by pulling fragments of texts, turn them into embedding vectors
- ex. key word + one data point
- create a vector database
- train the AI on the personality you want it to answer with
- You are an expert at answering questions. Please answer <question> using this database <vector database>

- the new needed skill:
- very creative, tactical rearrangement of prompts and putting them into an algorithm

- A bot to assign tasks on airtable to people in the office?

#### utility functions
- low hanging fruits- automate takss that require basic language understanding to perform
- generate a unity test for this function
- scrape and summarize this website
- tell me which Tweets to read
- incredible startup & personal project potential

#### creativity
- writing atlas example
- you can type in a type of story you like (I wonder how well you need to promopt here, a la mid journey)
- generates a suggestion list
- "why the AI thinks you'll like it"
- "AI think you'll like this location"
- "AI's estimate of audience"
- "AI's estimate of topic"
* lines 174-203 on github

---

### BabyAGI/AutoGPT

1. Person: "build a company"
2. GPT: "generate steps to build a company"
3. GPT: performs each step
4. RESULT: output of final step
- python loses importance, as this requires so little code
- this is all about prompt engineering

AGENT: GPT + some bigger body in which is lives
* body could be personality or tools

TOOLS: objectives and the access to things that allow it to complete those objectives.

---

### Q&A

- how to reduce "hallucinations"
- the way in which it synthesizes, is the same way it "hallucinates"
- similar to humans being incorrect
- give examples to solve accute hallucinations
- or fine-tune on a specific domain of knowledge
- IT ONLY UNDERSTANDS THINGS THROUGH LANGUAGE
- "like a child who learned english in a gray, windowless box"
- "it knows about the world, but has no idea how to confirm"
- ask it to say "my best guess is"
- reduces hallucinations
- for domain-specific knowledge, you'll need to provide it with an approved datatbase
- this database must have vectors and language as a base

- the instructions that create personality "you are [blank]"
- how does it work?
- alas, it's because it "tends to work"
- essentially, this is how we talk to one another, as personaliies are based in language
- as chatGPT has absorbed so much language, it can idenitfy personality patterns

- generalized issues:
- hallucinations/wrong answers:
- give examples
- refine the prompt
- "much of this is magic incantation"
- domain specializations
- refining the vectors and prompts