| What do I Want from My AI? The Problem(s) of Alignment |
| What will the impact of AI be in education? … System: You are a helpful assistant User: What will the impact of AI be in education? Assistant: … |
| What will the impact of AI be in education? What will the impact of
AI on work? In other words, the LLM response doesn’t need to be an answer – it is just more words / tokens System: You are a helpful assistant User: What will the impact of AI be in education? Assistant: AI is set to significantly impact education in several ways, transforming teaching, learning, and administrative processes. Here are some key areas of impact: |
| What is this? → The “System” (or “Developer”) prompt is only visible in the API. It sets the tone / character / personality** of the chatbot Let’s try it out… System: You are a helpful assistant** User: What will the impact of AI be in education? Assistant: AI is set to significantly impact education in several ways, transforming teaching, learning, and administrative processes. Here are some key areas of impact: |
| ENGINEER YOUR “YOU ARE A |
DESIGN YOUR DREAM |
Tools like ollama and LM Studio make testing with local
(downloadable) models easy |
| And to sum up: Last week we discussed unsupervised (base model – $$$$**) followed by supervised (fine-tuned – **$$) Now we can look at a kind of “learning” which is just for the duration of the chat session – and involves “teaching” the system what it should do (one-shot / few-shot learning – cheap! DIY!) This is also an example of alignment – aligning a model with our preferences via this initial instruction Why does it work? Prior SFT has configured the model to respect this first system prompt (doesn’t always work) |
| Stochastic = random Parrot = imitators Therefore, LLMs are imitators with random variation Also see Ted Chiang’s influential “blurry JPEG” Are these metaphors (parrots, blurriness) accurate? Becoming more or less accurate over time? |