🌱 (seedling) | Literature note |

Course Notes

Source on LinkedIn Learning - LinkedIn made this course available for free for 24h My completion

Was co-created using ChatGPT

Exercise: Pause and reflect

  • Take a view moments to think of one thing your business could use help with today
  • How could AI assist with that need

Course: Becoming and AI first leader

Major Factors in Al Advancement

  • Innovative deep learning architectures
  • Expansion of computing power
  • Trained on the entire internet

  • Create new products
  • Reimagine user experiences
  • Reinvent workflows

I you do have reservations .. write them down.

Define Algorithm objective - what is your goal and what do you want to accomplish?

Recent AI lower the entry bar for companies with less data (helping with the cold start problem).

Data alone will not be sufficient.

GPT (Generative Pretrained Transformer) General purpose Al models with a broad objective function, already trained on all public knowledge

Fine tuning and train the models on the unique dataset that we have.

  • to make our product unique

The AI model

  • a foundational model is considered a new paradigm for building an AI system
  • Rule based algorithms
  • Search & optimization algorithms
  • machine learning - supervises, unsupervises and reinforcement learning
  • deep learning
  • transformers and GPT (LLM)
    • based on transformers - Deep Learning architecture
  • diffusion model

Transformers

A type of neural network that uses a technique called “self-attention” to identify which parts of an input are most essential

  • Diffusion models

    Gradually transform a random noise starting image into a target image, through a series of small, randomly determined steps

Reinforcement Learning from Human Feedback

This approach uses signals from human evaluators (such as upvotes or downvotes) to improve the performance of an Al model.

Prompt engineering

  • need to be specific, nuanced and very well articulated

  • provide right context and guidance
  • needs trials and iterations

Zero-shot, one-shot, and few-shot learning refer to situations where Al is able to perform new tasks with very limited learning, even zero learning-basically no data.

Workflow automation

  • a good place to start with AI
  • creating, integrating, analyzing, summarizing reports - and facilitate that end-2-end experience

Use AI for Data and Analytics

  • Integrating data and reports
  • Creating new data sets
  • Generating synthetic data
  • Summarizing large data sets

Speed up innovation

Create adaptive content based on AI tutors.

Getting started

  • Define objective
  • Evaluate current capabilities
  • Identify data needed
  • Select the AI systems
  • Monitor, refine and assess results

Identify the right AI systems

  • Course Notes AI landscape map
  • What formats of data do I need?
  • How often does it need to be refreshed?
  • What source of high quality data do I have at scale?

Implementation How easy or difficult will it be to implement this type of Al in your company?

Scalability Can the model scale with your business?

Transparency Are you able to understand how the model making its decisions?

Vendors

  • Provide you with an established Al system
  • Help with data management and integration
  • Offer technical support and expertise

You can’t manage what you don’t understand

Ask yourself, given my objectives and current state, and by thinking Al-first, what can I do to create business value in a differentiated way?

Action Create a lightweight sandbox capability to help teams experiment in a safe environment

Limitations of AI today

as of 2023-7-23

An oversimplified view of your objective

The most important thing for any project, especially an Al project, is setting a clear objective and plan

High computational costs

Generating new content often requires significant computing space, power, and time

Algorithm hallucination

Al doesn’t know how to say “I don’t know”

  • ensure to have guardrailes in place

Staleness

Generative Al models can be limited by the data they have been trained on

Restrictiveness

Generative Al models don’t do well when it comes to access to basic information

Interpretability

Many generative Al models tend to be black box systems, meaning it can be difficult to understand why they generate a given output

Token constraints

A token refers to a unit that the generative Al system is able to process at one time

Ability to keep state

Keeping state or memory is essentially the ability to remember past inputs and build off them in order to generate outputs

Data quality and availability

We discussed the importance of “feeding” your Al with high-scale, high-quality data

Ethical concerns

Generative Al can be used to create fake content or for malicious purposes

Ethics & Data privacy

Linked with [[Ethics in AI]]

  • transparency
    • types of data used
    • explaining decisions made
  • data bias
  • data privacy
    • collect, stores, use
    • know what sensitive data you have
    • establish clear policies
    • educate your employees
  • Intellectual Property (IP) violations
    • understand the laws around that

Key Steps

  • Establish reporting and feedback mechanisms
  • Review prompts and output
  • Educate users of Al content about their responsibilities

Future

  • Advancements in algorithms
  • Reduced demand for computing power
  • Allowance of more tokens
  • Computational resources

Linking

Enjoy this post? Buy me a Coffee at ko-fi.com

Notes mentioning this note