Generative AI by Andrew Ng (coursera-DeepLearning.AI) - Note

Link to coursera: https://www.coursera.org/learn/generative-ai-for-everyone

AI

  • Supervised learning

    • Labeling the output

    • Performance depends on the amount of data

  • Unsupervised learning

  • Gnerative AI

  • Reinforcement learning


Large language models (LLM)

  • Using the supervised learning to generate the text (learning from the repetitive sentences)

  • LLM – > hundreds of billions of words – > that’s why the model can create a good performance

  • Language is quite repetitive – > that is why it can generate the text


LLM

  • A new way to find information

    • It can give you information

    • Sometimes, it hallucinate

  • Writing partner

  • Finding information through LLM (but better to double checks) – perhaps web search might be better

Generative AI

  • General technology (for now)

  • Useful for lots of things

  • Work well with unstructured data (non-tubular data, for example, table form)

  • Generating the image – > diffusion model by which labelling the text together with the image and repetitively do it step-by-step until getting the clear image – > by creating such model, that is why we can generate images based on text!


To improve the generative AI performance

  • Retrieval augmented generation (RAG)

    • Giving LLM access to external data sources

    • Have 3 steps

      • Finding the external data sources

      • Incorporate retrieval text into an updated prompt

      • Generate output from the new prompt with additional context

  • Fine-tune models

    • Adapt model to the specific task, like finance, law, or a medical issue

    • Give the LLM a modest number of inputs and outputs to adjust the model

    • Thus, the output might be more accurate and more precise

  • Pretrained models

    • Train model from scratch

    • Suitable for building specific application

      • It is better to start with general LLM first; otherwise, it will take a lot of resources - time and money

Choosing model

  • Model size

    • Depending on the parameters used to build model; more parameters mean complex model and take time to deploy

    • It is really depending on the tasks; less complex pick the smaller one, less parameters –  it will execute faster


Evaluation of AI potential – for application

  • Technical feasibility

    • Can AI do it?

  • Cost (Business value)

    • Cost saving?

    • How much time is needed to get this task done?


Artificial General Intelligence

  • AI that can do any intellectual task that a human can

  • Human are multiskilled but AI is more specific (good at one particular task)


Job + tasks breakdown – > help to decide which task could be automated using generative AI

https://www.onetonline.org/


The economic potential of generative AI: The next productivity frontier

https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier#/





Comments

Popular posts from this blog

Useful links (updated: 2024-10-23)

Odd ratio - อัตราส่วนของความต่าง

Note: A Road to Real World Impact (new MU-President and Team) - update 12 Sep 2024