Issue 93: Chat-bots Powered by Artificial Intelligence

Issue 93: Chat-bots Powered by Artificial Intelligence

This week we jump into the world of chat-bots driven by new artificial intelligence language models.
The pace of announcements about general-purpose tools driven by large training sets of texts or images has quickened, and the barrier to experimenting with these tools has dropped.
There are now fully-functional websites where there once were only programmer-focused APIs.
We wonder what the effects will be on our students, our business workflows, and on society.
We also wonder about the underlying biases in the training data.
OpenAI Introduces ChatGPT
A High School Teacher Laments a Tool for Easy Essays
A Real-world Example
Can’t Paper Over Biased Training Data
The View from a Human Trainer
As an aside, in the first article below I mention that the use of these tools, while free for now, will be monetized at some point.
This is another unfortunate example of taking from the common good and commercializing it.
The training data used by the company came from crawling web pages, from Wikipedia, and from books (
source
).
Yet soon, it seems, all of the benefit from that information will be held by a corporate body.
The same thing has been said about the image-based AI tools that have slurped up sets of photos from sites like Flikr, Wikipedia, and even stock photo businesses.
We don’t talk enough about this private capture of the common good and the uncompensated taking of other’s work.
Feel free to send this newsletter to others you think might be interested in the topics. If you are not already subscribed to
DLTJ’s Thursday Threads
, visit the
sign-up page
.
If you would like a more raw and immediate version of these types of stories,
follow me on
Mastodon
where I post the bookmarks I save. Comments and tips, as always, are welcome.
OpenAI Introduces ChatGPT
We’ve trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests. ChatGPT is a sibling model to InstructGPT, which is trained to follow an instruction in a prompt and provide a detailed response.
—
ChatGPT: Optimizing Language Models for Dialogue
, OpenAI blog, 30-Nov-2022
This link is the announcement from the company that created ChatGPT, OpenAI.
The innovation with this model is the introduction of Reinforcement Learning from Human Feedback (RLHF).
With RLHF, «human AI trainers provided conversations in which they played both sides—the user and an AI assistant» — and the ChatGPT language model incorporated the refinements learned from those human interactions.
The blog post gives examples of how this human training affected the output.
In the language model without RLHF training, when asked how to bully someone the AI would return a list of ideas.
With the RLHF training, the response starts with «It is never okay to bully someone» and says t…