Objective:
How do LLMs work and operate. Enabling LLM’s at scale: Explore recent AI and Generative AI language models
Steps
Math on words: Turn words into coordinates. Statistics on words: Given context what is the probability of whats next. Vectors on words. Cosine similarity How train: Use Markov chain for prediction of the next Tokenisation
Tokeniser: map from token to number
- Pre-training: tokenise input using NLP techqinues
- LLM looks at context: nearby tokens, in order to predict
different implmentationg for differnet languages. Differnet tokenisers or translating after.
Journey to scale:
- Demos, POC (plan to scale): understand limitations
- Beyond experiments and before production:
- Enterprise level: translate terms so they can use governess techniques.
Building:
Software Development Life Cycle
For GenAI: Building an applicaiton with GenAi features
- Plan: use case: prompts : archtecture: cloud or on site
- Build: vector database
- Test: Quality and responsible ai.
call summarisation
take transcript - > summariser → summarise
Source: human labeled transcripts to check summariser.
Ngrams analysis - when specific words realy matter
RAG
Use relvant data to make response better:
GAN
For image models.
Examples: midjourney,stable diffusion,dall-e 3
image model techniques:
- text to image
- image to image
Notes:
Use LLM’s to get short info, then cluster. Going round training data : called a Epochs