2-Pane Combined
Full Summaries Sorted

Could I make myself smarter with a Thinking Partner persona as a Machine Learning/Natural Language Processing expert?

Here are some thoughts (and a few links) that may provide some more context for what I’m starting to do with Now Comment and the related tech details, which are largely beyond me.
In short, I was starting to use the notion of a Thinking Partner persona as ML/NLP expert to see if I could make myself smarter on a technical subject where I have no experience or aptitude beyond my curiosity and conviction that these tools and the inductive way of thinking they promote (for the moment) are going to be become how we evolve and adapt to the ever-expanding information environment we are clearly living in now…
As a non-technical person, I’m trying to understand ML/NLP concepts and system-level characteristics as design principles, in order that they might might be applied in familiar workflows and content generation I can recognize, which could then inform a chatbot that coaches for student voice specifically, say, and understands how to develop a writer v. improve the text according to rote categories…A steep hill to climb w/ or w/o AI.
Some of this has involved researching the open source models and the benefits for fine-tuning for folks with their own interest areas and access to deep domain expertise. I don’t know how to, say, organize a fine-tuning training set…I’d like to understand that. So I’m banging away at dense texts like a human, anthropomorphizing as I go. Predictably, I’m finding that so far it’s a less than straightforward or effective process. Interesting yes.
I can get definitional information out, but the detail about how to design something, how to run through 5K of student writing samples, for example, is tricky beyond generic categories when working with the open prompting tools. I need to get much, much better at prompting differently and narrow the focus of my questions in technical ways to see if the system might write the way an AI company would create a how-to for developing a training set — here’s OpenAI’s help file on how to create a training set…It’s like the Character Book on Open AI — full of step and settings and formats — but for a highly technical outcome.
One can see how to create question/answer formatting, and with that all the python-ready code or CSV converters that would spit out JSON-ready information to correctly inform the creation of a fine-tuning set. How to get from “make sure all the Ads on my website are correct” to “annotate all the ways this piece student writing revearls what we can think of her writing “voice” and provide comments on how she might go further, is well beyond me…
The formatting for JSON, etc reminds me of the early consumer internet/CD rom days when we were learning how to do SGML so we could render text on a rudimentary web page. The querying against personal word choice, the expression of values and tightly held beliefs, etc…this feels quite far. But finding language and workflow to turn this latter category into a repeatable parsing process is motivating.

Related: in my effort to find people with deep ML/NLP expertise, I mentioned I managed to speak (albeit briefly) with a CS / NLP professor who had overseen some education-related work to develop a teaching and learning dataset to support a chatbot tutor. She had a few things to offer in relation to that work:
1) She would have liked to develop that project with live teachers and students — what a slice of the NWP network might be moved by, given the proper framing and support. She and her grad student had relied on Mechanical Turk input for role-play and content generation sequences. The results and the engagement would have been more useful, I gathered, had real people been involved in the design and content generation. Additionally, she said their methodology now would not be really, or usefully, repeatable…
2) She added not got bandwidth right now, too much admin, and she’s no longer on education-related work, but she deeply appreciated the potential and the value in what’s going on with all the edu engagement.
3) She offered to send me things that might be related, then she forwarded a link to an upcoming Computational Linguistics conference and a call for participants working on reliable tutors and chatbots for education – Shared Task on Generating AI Teacher Responses in Educational Dialogues. It’s based on academics doing structured prompt and fine-tuning work in a Kaggle / Open Challenge model, sending in JSON examples of their work, contributing it to a curated evaluation process….
All this speaks to some of the ideas Elyse voiced initially: having chatbots that train/support the writer v. correct the text. In principle, Thinking Partners is all about this notion with the added benefit/complexity of providing a “peer” or expert review function that’s situated deeply in a narrative and contextual context and query.
I’m experimenting with my questions on all of this in the TP system, and others, in the way I used to wander the stacks at our local library when I was a kid…with about the same amount of wonder (a lot) and about the same amount of actionable take-away (not a lot, but many ideas and very energized).
Eager for your thoughts on any of this. I’ve essentially written out a one-way conversation here. Let me know what if any of this feels useful to you in your thinking about NC & TP and in your curiosity generally.

DMU Timestamp: June 02, 2023 17:50

0 comments, 0 areas
add area
add comment
change display
add comment

Quickstart: Commenting and Sharing

How to Comment
  • Click icons on the left to see existing comments.
  • Desktop/Laptop: double-click any text, highlight a section of an image, or add a comment while a video is playing to start a new conversation.
    Tablet/Phone: single click then click on the "Start One" link (look right or below).
  • Click "Reply" on a comment to join the conversation.
How to Share Documents
  1. "Upload" a new document.
  2. "Invite" others to it.

Logging in, please wait... Blue_on_grey_spinner