The Journal - October 2024

“FIRE: Financial Independence, Retire Early.”

(Monday 28/10/2024)

(Tuesday 29/10/2024)

这一天，整理下网站的内容，希望有一个清晰的分类管理，便于在此基础上来充实。

没有进行有效的阅读。只跟女儿一起阅读了有关MindMap的记忆技术，短期和长期两种大脑记忆类别。 如何将短期记忆转化为长期记忆，主要还是反复记忆，形象记忆，及跟已有记忆形成关联。

(Wednesday 30/10/2024)

Why not leverage the knowledge we can learn within free scope?

Akamai, is great for CDN knowledge, Networking Security (DDoS) and more now, Cloud Computing with Linode.

1.Akamai Connected Cloud

• Core (Linode)
• Edge (CDN Infrastructure)
  • FaaS (Serverless)
  • Cloudlet (local compute and storage)
  • CDN (Cache, Content delivery)
  • DNS (routing to edge node)
  • GLB (distributes multi-locations)
• Edge-Native vs Cloud-Native
  • edge of network, close to user and data source. V.S. microservices which iterating fast and contineously delivered.
  • decentralized deployment, high availability, even disconnected from central network. V.S. HA multi-region
  • scalability, scale out by adding more edge nodes, great performance geographically. V.S. Centralized DCs still relying on CDN/Edge underneath.
• Edge Native Application
  • Data Governance
  • Cost Efficiency
  • Improved Performance

2. AI/OCI Module-1. LLM

By reviewing some links on LinkedIn, I found a few folks with Oracle University Generative AI professional centificate. Here it is

What is a Large Language Model?

• LLM Architecture What else can LLMs do?
  • Encoders and Decoders
  • Embedding and Text Generation
  • Transformer Architecture

  Encoder - models that convert a sequence of words to an embedding (vector representation) Decoder - models that take a sequence of words and output next word (text generation, chat-style Q&A) Encoder-decoder - encodes a sequence of words and using decoder to generate another (translator)

• Prompting How do we affect the distribution over the vocabulary?. Based on Decoder only models, the way to affect the distribution:
  • Prompting / Prompt Engineering challenging, unintuitive, not guranteed to work. strategies for a good Prompting:
    • In-context learning: conditioning (prompting) an LLM with instructions.
    • k-shot Prompting: explicitly providing k examples of the intended task in the prompt.
    • chain-of-thought: emit intermediate small reasoning steps.
    • least-to-most: prompt LLM to decompose the problem and solve, easy-first.
    • step-back: prompt LLM to identify high-level concepts pertinent to a specific task risks:
    • prompt injection (jailbreaking)
    • memorization (after answering, repeat the original prompt)

• Training Fine-Tuning (classic ML Training) Param. Efficient FT (LORA) soft prompting: Learnable prompt params (cont.) pre-train: (unlabeld, same as ML pre-training) inference

• Decoding How do LLMs generate text using these distributions?

  Greedy Decoding:

  • pick the highest probaility word at each step.
  • send the new sentence as input
  • until EOS is the highest probaility word.

  Non-Deterministic Decoding:

  • pick randomly among high probability candidates at each step.
  • Temperature: modulates the distribution over vocabulary. temperature decrease closing to greedy decoding.
  • relative ordering of the words is unaffected by temperature.

  Nucleus sampling

  Beam search

• Hallucination: generated text that is non-factual and/or ungrounded. (Simply not from pre-trained data set.)

• LLM Applications RAG Code models Multi-modal Language Agents
  • ReAct
  • Toolformer
  • Bootstapped reasoning

(Thursday 31/10/2024)

OCI Module-2. GenAI Service

• OCI Generative AI
• Chat Models
• Embedding Models
• Prompt Engineering
• Customize LLMs with your data
• Fine Tuning and Inference
• Dedicated AI Cluster Sizing and Pricing
• Security

(1/11/2024 Friday)

• Observability - Spring

(Infrastructure) -> Cloud (TF Linode) -> K8s -> Helm/Docker -> Spring Native (Microservices) -> Spring GraphQL Kotlin (MobileApp) -> iOS Application

(2/11/2024 Saturday)

(3/11/2024 Sunday)

Journals Page