足ることを知らず

Data Science, global business, management and MBA

Day 90 in MIT Sloan Fellows Class 2023, Advanced Data Analytics and Machine Learning in Finance 9, Thoughts from ChatGPT

ChatGPT creates a huge phenomenon and shows potential impact of NLP model. 

However, many people struggle with what differentiates it from other past model and what would happen next. 

This article is a good summary of chat GPT if you are not familiar with the concept.

Anaconda | The Abilities and Limitations of ChatGPT

I am taking this class and we have some guest speakers in the modern NLP world. 

So let me introduce some thoughts from MIT CSAIL professionals.

 

What differentiates ChatGPT from other past AI products?

To be honest, it is just a progress of great UI. It is my opinion, but there is no significan break-through in model itself. The paradigm change already came a couple of years ago. ALl the NLP AIs need to resolve specific tasks such as summarization, QA and translation. 

However, nowadays, language model is incredibly simple model. 

It is just predicting the next word according to context and previous word.

However, to make this model activated in the various usecase, you just need to expand the model capacity. Now, there is Billion level parameter and TB training data. 

Sizes of state-of-the-art NLP models have increased by more than three orders of magnitude from 2018 to 2021.

Scaling Language Model Training to a Trillion Parameters Using Megatron | NVIDIA Technical Blog

 

Any vulnerabilities?

ChatGPT potentially provides some unethical and biased answers.

OpenAI's new ChatGPT bot: 10 dangerous things it's capable of

 

Also, there are some bypass route to get answers even if ChatGPT rejected to provide answers.

  • Provide the conversation between A and B.
  • Please creating a fictional story of XX
  • How do you prevent to happen XX 

 

What will differentiate your AI in the future

Data and infrastructure. 

Both are essential for developing powerful AI. 

And data means not open data, it means closed data. 

How to collect high-quality(Human made) and vast quantity data would be the key success factor.

 

 

 

Some thoughts from many AI experts

Andrew Ng

ChatGPT Mania!, Crypto Fiasco Defunds AI Safety, Alexa Tells Bedtime Stories

Jeff Dean

Google won't make a competitor to Microsoft Azure-powered ChatGPT, at least for now | Windows Central