ChatGPT -- DataCamp learning
Note from DataCamp on
Introduction to ChatGPT
- ChatGPT is generative AI, which is a subset of ML (ML is a subset of AI).
|
Artificial
intelligence
|
-
It generates the new contents by using patterns in the
information that it has learned before (a large language model).
-
What does it do?
-
Prompt (question, asking, request, action) – input
-
Throw it at the large language model (LLM)
-
LLM uses a complex algorithm to determine the pattern
and structure and create the response (generative)
-
Create a response as output
-
Limitation
-
LLM
-
Frequency and order of words (like a building block,
each block represents the word)
-
Many types of data are trained to build LLM
-
To create a model, a complex algorithm was used to understand
the pattern of the training data
-
The model is fine-tuned through responsiveness and
feed-back (I guess by humans? – like thumbs up/down when you see a response
that does/does not make any sense!
-
Potential bias due to the training data that was used
to build the LLM
-
Training data comes from
-
Books
-
Articles
-
Websites
-
As you may see, ChatGPT can track the context, so if we
create a relevant prompt, – the response will not be correctly
-
Thus, to use ChatGPT efficiently, - it is better to
talk about only one topic at a time.
-
Hallucinations occur when the LLM feels confident in
its response to the prompt and gives an answer that goes beyond its training
dataset.
-
Legal/ethical consideration – saying you want to create something with your
own intelligence but you used ChatGPT to generate it – soo who owns the
product? A person who provides the dataset to train GPT or person who used
ChatGPT or OpenAI – kinda gray area
-
Garbage in Garbage out: see older post - https://t-lerksuthirat.blogspot.com/2020/06/machine-learning-for-everyone.html
-
Thus, we need to write a good prompt that will result
in a better (more accurate) response.
-
Prompt engineering – the process of writing prompts to
maximize the quality and relevance of the response
-
Guideline
-
Clear and specific
-
Contain necessary information
-
Specify how long of the response that you would like it
to be - one page, one paragraph, or one sentence
-
Be concise
-
Remove information that does not provide useful context
-
Using correct grammar and spelling in the prompt –
ChatGPT uses grammar when interpreting the task
-
We can provide the sample to ChatGPT in order to
forcibly generate the response in a specific format
-
Use or not use ChatGPT – rule of thumb – use it when
you are able to verify the data generated from the ChatGPT.
Next step – what
drives the improvement
-
LLMs
-
Learning from a huge text dataset
-
It will increase
-
More complex will be learn, for example, sarcasm and
idiom
-
Algorithms detect patterns in text
-
Fine-tune the model by rating responses
-
Users help fine-tune the response - this will make the
model more human-like
-
Bias is the most challenging part – thus the model
should create high quality and balanced data.
-
Law and ethics
-
Misuse in ChatGPT
-
Creating malicious content
-
Misrepresenting AI-generated content (people are being
fooled by AI)
-
It is possible that ChatGPT will be more specialized.
Comments
Post a Comment