Invited ngawangtrinley@gmail.com gade@dharmaduta.in kevin@indrajala.tw passang9814@gmail.com
Attachments Strategy & Roadmap Review- AI Tools WG
Meeting records Transcript
Summary
Ngawang Trinley, པ་སངས་ཚེ་རིང (Passang Tsering), and Kevin Shakya discussed the workflow and development plan for Buddhist AI tools, focusing on defining core features and prioritizing a “wish list” of updates, including updating the homepage to reflect “Buddhist AI tools.” The participants agreed on the scope of AI tools, which encompass translation and adaptation, and reviewed working group roles, with Ngawang Trinley proposing Kevin Shakya to start as the meeting coordinator/facilitator for the adaptation/translation working group. པ་སངས་ཚེ་རིང (Passang Tsering) demonstrated the translation tool and highlighted key issues, including the model’s failure to maintain stanza context due to line-by-line translation and the need for better resource integration, more user control over the prompt and model selection, and the integration of multiple AI assistants/agents for tasks like proofreading.
Details
Notes Length: Standard
-
Workflow and AI Tools Ngawang Trinley initiated the meeting by refreshing the workflow and discussing the current focus on AI tools, specifically mentioning the need to update the homepage and calendar to reflect “Buddhist AI tools”. They outlined the development plan, which involves finalizing a “wish list” of updates and new features, prioritizing them on the GitHub board, and then writing user stories (00:18:03). The core goal is to determine four to six core features for the quarter (January, February, March), with each feature comprising five to ten stories (00:19:24).
-
Tool Scope and Definitions The meeting addressed the scope of the translation and adaptation tools. Ngawang Trinley defined the AI tools as those used to publish, upload, edit, translate, edit, align, format, and adapt Buddhist texts (00:13:15). A discussion with པ་སངས་ཚེ་རིང (Passang Tsering) confirmed that translation and adaptation should likely be housed in the same tool, as paraphrasing (a form of adaptation) is considered a type of translation, despite different usage (00:14:57).
-
Development Planning and Timeline Ngawang Trinley emphasized that by the end of the current week (week zero), the team needs to finalize the wish list and define the four to six core features, as well as the main stories for each (00:18:03) (00:20:52). This prioritization will allow the tech team to decide their work for the next two-week sprint starting on Monday (00:22:18). The core features are expected to take two to six weeks each to develop (00:19:24).
-
Working Group Roles and Facilitation Roles within the working group were reviewed, with Ngawang Trinley noting that པ་སངས་ཚེ་རིང (Passang Tsering) is involved in translation/adaptation, and Kevin Shakya merged themself with the “voting advisors” list, eliminating the “non voting advisors” role (00:11:01). Ngawang Trinley suggested the need for a meeting coordinator/facilitator to ensure decisions are taken and meetings run on time, proposing Kevin Shakya to start in that role for the adaptation/translation working group (00:27:13).
-
User Demonstration and Feedback པ་སངས་ཚེ་རིང (Passang Tsering) shared their screen to demonstrate their use of the translation tool, explaining their process of inputting the original Tibetan text and letting the AI model (currently 2.5 GI pro or 2.5 flash) perform the translation (00:32:11). When reviewing translations, they look at the commentary from Peta tools and mark stanzas if the AI fails to translate the context, which they attribute to the model attempting line-by-line translation. པ་སངས་ཚེ་རིང (Passang Tsering) also noted that they have given feedback on this issue twice or thrice in the community hub and suggested that the main problem needing attention is the prompt being used (00:33:18) (00:36:09).
-
Core Features for Improvement The discussion identified key features needing improvement based on པ་སངས་ཚེ་རིང (Passang Tsering)'s feedback. A major issue is the lack of context and resources within the tool, requiring them to access commentaries and dictionaries externally (00:36:52) (00:44:47). Key improvements include: allowing users more control over the prompt and model selection (00:34:22) (00:40:50); displaying linked resources (like commentaries and UCCC) for specific verses on the left side, with the ability to select which resources go into the prompt (00:36:52) (00:39:53); and integrating dictionary functionality, preferably on the left side with the option to hide it (00:44:47).
-
User Interface and Experience པ་སངས་ཚེ་རིང (Passang Tsering) indicated that the current left/right UI layout for source and target text is “quite good” for comparison (00:42:46). However, they pointed out issues with manual functions not working (specifically the resources panel) (00:39:53) and excessive empty spaces/line breaks in the text, requiring manual correction (00:47:02). They also expressed that while they can edit the translation text directly (00:53:57), they spend most of their time editing in the middle of the screen (00:52:44), and asked about placing model selection and prompt options on top of the editing area (00:49:01).
-
Prompt and Context Issues A primary concern for པ་སངས་ཚེ་རིང (Passang Tsering) is the AI model’s inability to maintain the context of the whole stanza due to translating sentence by sentence, suggesting the model may not be using the intended resources like UCCC commentary (00:39:53). They noted that attempts to update the prompt in the settings only changed part of it and had no effect on the result (00:36:52). Ngawang Trinley clarified that the UI should distinguish between general context and context for a specific verse (00:42:46).
-
Editing Preference and Workflow པ་སངས་ཚེ་རིང (Passang Tsering) preferred to edit translations directly on the text rather than in a separate box, finding the current method quite good. They described their process of bringing the AI translation response into the main translation column and then performing the review there, noting that having two places to edit is “a little bit weird” (00:56:09). པ་སངས་ཚེ་རིང (Passang Tsering) also explained that editing in the translation column allows them to compare results from different models by selecting the text and using another model (00:57:54).
-
Segmentation Issue: Line-by-Line Translation Ngawang Trinley and པ་སངས་ཚེ་རིང (Passang Tsering) identified a major problem where the system translates line by line instead of verse by verse, which པ་སངས་ཚེ་རིང (Passang Tsering) noted severely impacts the quality of the translation and context (00:58:58). Ngawang Trinley confirmed that this line-by-line translation, which occurs even when selecting a stanza, is a “big problem” (00:59:52) (01:06:03). པ་སངས་ཚེ་རིང (Passang Tsering) suggested that not performing segmentation made the translation much better because it translated things together (01:07:00).
-
System Instability and Model Selection པ་སངས་ཚེ་རིང (Passang Tsering) reported issues with the model not working at all, having to contact “Kang” multiple times about the instability of the models. Ngawang Trinley stated that they need to be able to select which model to use and what prompt to apply, including saving and changing different prompts based on the section being translated (01:00:57). པ་སངས་ཚེ་རིང (Passang Tsering) agreed with the need for prompt customization, though they noted it would only work for users who know about prompting (01:01:48).
-
User Interface and Resources The participants discussed the user interface layout, with པ་སངས་ཚེ་རིང (Passang Tsering) confirming that having resources like reference texts open on the side of the screen is “quite good” for them (01:01:48). Ngawang Trinley pointed out that there is an icon for discussion comments intended for collaborative work on a translation, and པ་སངས་ཚེ་རིང (Passang Tsering) confirmed its utility for multiple people working on the same text (01:03:03).
-
Need for Multiple AI Assistants/Agents Kevin Shakya and Ngawang Trinley introduced the idea of having different AI agents, such as a proofreader, to assist in the translation process (01:10:09). Ngawang Trinley argued that having several assistants is important, noting that currently there is only one for translation (01:14:06). པ་སངས་ཚེ་རིང (Passang Tsering) confirmed that they currently check commentaries outside of the tool to ensure context is correct, and agreed that an agent could perform checks against commentaries, Sanskrit, or identify mistranslations, saving them from line-by-line proofreading (01:09:07) (01:10:55).
-
Defining AI Translation Workflow Ngawang Trinley proposed that པ་སངས་ཚེ་རིང (Passang Tsering) and Kevin Shakya document the specific tasks they perform with AI outside of the current tool, such as getting direct translation, double-checking against commentaries, and vocabulary standardization (01:12:12). Ngawang Trinley suggested that once these steps and AI usages are listed, they can be integrated into the tool to make the process easier (01:13:03). The participants agreed to this testing and review of the workflow for both adaptation and translation (01:14:06).
-
Next Steps and Follow-up Meeting The group decided to meet again tomorrow morning in India at 9:00, with Ngawang Trinley suggesting 11:00 their time (01:15:13). Ngawang Trinley tasked Kevin Shakya with conducting research on AI-powered translation workflows, focusing on how people use AI at different stages and the types of AI assistants used. This research and the workflow analysis from པ་སངས་ཚེ་རིང (Passang Tsering) will inform the finalization of features and the assignment of tasks (01:16:14).
Suggested next steps
-
Ngawang Trinley will add Kevin Shakya as the meeting coordinator/facilitator for the adaptation/translation working group.
-
Ngawang Trinley will ensure that the dictionary being used by པ་སངས་ཚེ་རིང (Passang Tsering) is integrated into the translation tool on the side with an option to hide it.
-
Ngawang Trinley will also work on the workflow on his side.
-
Kevin Shakya will do research on the main AI-powered translation workflows and the different AI assistants people use.
-
Ngawang Trinley and པ་སངས་ཚེ་རིང (Passang Tsering) will meet again tomorrow morning at 9 in India to discuss the main things that need to be changed based on the workflow tests.