Introduction¶

Generative AI¶

AI systems that can produce high quality unstructured content: text, images, audio

Impact on Jobs¶

More effect on

• higher-paid jobs
• knowledge workers

Lifecycle of GenAI Project¶

1. Scoping
2. Build prototype
3. Internal evaluation
4. Improve system
5. Deploy
6. Monitor

LLMs¶

Large Language Models

Supervised learning to repeatedly predict the next word

Applications¶

• Finding new information
• Writing
• Assistant
• Translation
• Reading
• Proof reading
• Summarization
• Chatting

Advice for chatbots: Start with internal-facing that works with staff

What an LLM can do¶

Rule of thumb

Whatever a fresh undergraduate can do with the given prompt and

• No internet/other resources
• No training specific to your prompt
• No memory of previous tasks

Prompting Tips¶

1. Be detailed: Give LLM sufficient context & information required to task at hand
2. Be specific
3. Guide the model to think through its answer: Suggest steps for performing task
4. Experiment and Iterate

Objective¶

Helpful, Honest, Harmless

How it works¶

• Instruction tuning
• RLHF: Reinforcement Learning from Human Feedback
• Train another Supervised Learning model for answer quality rewards
• Train LLM to generate responses with high response scores

Can be used to reduce bias

Tool-Use¶

Action¶

Reasoning¶

Agents¶

• Use LLM to close and carry out complex sequence of actions
• Not ready at the time of typing this

Image Generation¶

Diffusion Model

Noise + Prompt -> Generated Image

Limitations¶

• Knowledge cut-off
• Hallucinations: LLM can produce confident responses which are completely false
• Prompt size is limited
• Does not work with structured data
• Does not do arithmetic well
• Bias & Toxicity

Caveats¶

• Be careful with confidential information
• Double-check: LLMs do not necessarily give reliable information
• For user service, better to have confirmation dialog before performing transaction

Cost of LLM¶

Relatively cheap to use

4 tokens ~ 3 words

RAG¶

1. Given question, search relevant documents for answer
2. Incorporate retrieved text into updated prompt
3. Generate answer with new prompt with additional context

Fine-Tuning¶

1. To carry out a task that isn’t easy to define in a prompt
2. To help LLM gain specific knowledge
3. To get a smaller model to perform a task

Pre-Training¶

• Very costly
• Requires large amount of data

For building a specific application, pre-training is the last resort

LLM Model Size¶

General guidelines

Comments