Generative AI: Notes from Academia

Presented at the Singapore German Chamber of Commerce

as part of their "Demystifying ChatGPT: how can enterprises leverage generative AI" session, with a following expert panel.
---
Venue: Munich Room, Level 5, 25 International Business Park, German Centre, Singapore 609916
Date: 22 June 10-11.30 am
Attendees: SGC members coming from diverse industries

Synopsis:
All eyes are these days on the breathtaking development in the AI arena. AI models like ChatGPT have the potential to revolutionize the way how enterprises interact with their customers and partners. Organizations that are more mature in their adoption of analytics and AI technologies have a competitive advantage, with AI transforming business, government, and scientific research everywhere.

The session will bring together experts from academe, and enterprises to present and discuss the following topics:

· A look at the benefits and drawbacks of AI technology in the workplace
· The technical aspects of how ChatGPT work
· Use cases for ChatGPT in various industries
· Challenges and limitations of using this AI technology

Transcript

1. Generative AI:
   Notes from Academia
   Kan Min-Yen
   Demystifying ChatGPT:
   How can Enterprises Leverage Generative AI
   Images all generated by Stable Diffusion. All slides are all in the public domain
   

   View Slide

2. Some Context: A perfect storm
   The confluence of gaming, big data and algorithms got us
   to where we are now.
   Neural networks process and store text, image and video
   data all as in long sequences of numbers, known variously
   as embeddings or vectors.
   

   View Slide

3. Generative AI
   Learns patterns from large quantities
   of a data of modality
   Outputs something highly probable
   given what it’s seen.
   • Text: guess a missing or next word
   • Image: guess a patch of image
   given a context
   • Cross Modal: guess an image given
   a caption
   Solves the equation
   𝑃 𝑜𝑢𝑡𝑝𝑢𝑡 𝑖𝑛𝑝𝑢𝑡)
   

   View Slide

4. Conditional Generative Models
   Conditioned on some form of
   input (“prompt” or “instruction”)
   “Pretend you are an expert
   programmer. Then …”
   “In the style of Rene Magritte”
   Solves the equation
   𝑃 𝑜𝑢𝑡𝑝𝑢𝑡 𝑖𝑛𝑝𝑢𝑡, 𝒄𝒐𝒏𝒅𝒊𝒕𝒊𝒐𝒏)
   

   View Slide

5. Hallucination
   Why does AI hallucinate?
   AI outputs are probabilistic.
   Current systems don’t
   explicitly remember where
   its sources of evidence are
   from.
   When it mixes input that are
   coherent, we get output
   which is probable, but is
   incorrect.
   ChatGPT summarizing a non-existent New York
   Times article even without access to the Internet.
   From Hallucination (Artificial Intelligence), Wikipedia
   

   View Slide

6. Hallucination
   Why does AI hallucinate?
   AI outputs are probabilistic.
   Current systems don’t
   explicitly remember where
   its sources of evidence are
   from.
   When it mixes input that are
   coherent, we get output
   which is probable, but is
   incorrect.
   ChatGPT summarizing a non-existent New York
   Times article even without access to the Internet.
   From Hallucination (Artificial Intelligence), Wikipedia
   

   View Slide

7. Academic Roadmap
   • Multistep Reasoning
   • Improved Multilingual Capability
   • Better Consistency and Personalization
   

   View Slide

8. Acceleration has it out for us
   Adapted from Astro Teller’s graph
   Time
   Rate of Change
   Human Adaptability
   AI Tool Development
   Government
   Individuals
   Companies
   Humanity
   Opportunity
   Gap
   

   View Slide

9. Outlook and
   relevance to SG
   Singapore is a strategic bridge between
   East and West, valued for transparency.
   Generative AI brings more power to
   individuals, which means the collective
   must keep individuals in check.
   SG Government’s initiatives are aligned
   in this and fostering partnerships for
   large and medium sized enterprises to
   stay ahead of the curve.
   https://ctic.nus.edu.sg/index.html
   

   View Slide