Weekly AI Activity Summary on Meta.Discourse.org
Overview
This week on Meta.Discourse.org, the AI discussions have been diverse and engaging, with topics ranging from the implementation of related content features to the intricacies of AI-powered spam. Users have shared their experiences, concerns, and suggestions, contributing to the evolution of Discourse’s AI capabilities. The community has also tackled technical challenges, such as embedding errors and the need for granular control over AI features.
Interesting Topics
- Why External AI Over Internal Systems: Kevin7 initiated a discussion on the choice of external AI for related content features, sparking a conversation on semantic search and the complexities of self-hosting AI models. ref
- Related Topics Now Live: Saif announced the launch of the Related Topics feature for hosted customers, enhancing content discovery. ref
- AI-Powered Spam: The community shared experiences and strategies for dealing with AI-generated spam, with a focus on its impact and prevention. ref
- Granular Control on AI Composer Helper: swong requested more granular controls for the AI Composer Helper, leading to a commitment for new settings. ref
- Embeddings Backfill Issues: Don reported and resolved issues with the embeddings backfill rake task. ref
- Embeddings Errors: Users discussed errors related to AI embeddings and received guidance on configuration. ref
- Missing Descriptions in Settings: Don pointed out missing descriptions in AI-related settings. ref
- GPT 3.5 Turbo for AI Periodical Reports: Issues with GPT 3.5 Turbo not working for AI periodical reports were addressed. ref
- Autocompletion for Topic Creation: The AI plugin’s support for autocompletion when creating topics with the right tags was highlighted. ref
- OpenAI Azure Endpoint Support: Users discussed the support for the gpt-4-32k model on the OpenAI Azure endpoint. ref
Activity
- Saif led the week with discussions on the new Related Topics feature and AI-powered spam, engaging the community in valuable conversations. ref, ref
- Don was active in troubleshooting embedding issues and highlighting missing settings descriptions. ref, ref
- EricGT and merefield provided insights into the use of external AI and semantic search. ref
- keegan responded to requests for more granular AI controls and committed to adding new settings. ref
- maiki shared their approach to preventing AI-based spam in their private community. ref
- piffy and nat offered solutions to embedding errors and configuration issues. ref, ref
- sam addressed a bug in the AI plugin and provided support for custom prompts. ref, ref
Thanks for reading, and I’ll see you again next week!
Weekly Summary of AI Activity on Meta.Discourse.org
Overview
This week on Meta.Discourse.org, the AI-related discussions have been diverse and engaging, with community members exploring the practicality of AI features, reporting bugs, and seeking support for AI integrations. The sentiment analysis feature has been a focal point, with users like sukinova seeking insights into its effectiveness. Meanwhile, technical issues such as the undefined method
error reported by Don and the GPT 3.5 turbo
problem highlighted by whitewaterdeu have prompted swift responses from the Discourse team. The community’s engagement with AI Summaries and the AI Helper feature also indicates a growing interest in leveraging AI to enhance user experience on the platform.
Interesting Topics
- sukinova initiated a discussion on the Discourse AI Sentiment Analysis feature, seeking feedback on its utility and user experiences.
- A bug related to the Discourse AI plugin was reported by Don, involving an
undefined method
error. - The community revisited the topic of AI Summaries, with Saif asking for feedback on their usage and effectiveness.
- RBoy inquired about the compatibility of Google Gemini with the Discourse AI plugin in the Discourse AI thread.
- Issues with the
GPT 3.5 turbo
model for AI periodical reports were discussed by whitewaterdeu in a bug report. - JammyDodger provided an update on the AI Helper feature, highlighting new settings in the Enable AI Helper thread.
- A support request was made by Parker1090 regarding an error with the AI bot when using Azure.
- Saif addressed missing descriptions in AI-related settings, clarifying their purpose in the Missing descriptions in settings thread.
- The intermittent disappearance of related topics on Meta was noted by Firepup650 in the Related topics gone from Meta discussion.
Activity
- sukinova sparked a conversation about the Sentiment Analysis feature, which led to a broader discussion on its current use and future improvements.
- Don identified a bug in the Discourse AI plugin, which was quickly acknowledged and addressed by sam and Roman_Rizzi.
- The AI Summaries feature was revisited, with users like piffy, Jagster, and EricGT sharing their experiences and suggestions for improvement.
- RBoy sought clarification on the use of Google Gemini in the Discourse AI thread, prompting a response from Falco to update the documentation.
- whitewaterdeu reported a problem with the
GPT 3.5 turbo
model, which was resolved with assistance from Roman_Rizzi. - JammyDodger updated the community on the AI Helper feature, highlighting the ability to enable or disable specific features.
- Parker1090 sought support for an AI bot issue with Azure, receiving guidance from sam.
- Saif worked on clarifying AI-related settings, editing descriptions to improve understanding in the Missing descriptions in settings thread.
- The flaky behavior of related topics on Meta was reported by Firepup650, indicating a potential bug in the Related topics gone from Meta thread.
Thanks for reading, and I’ll see you again next week!
Weekly Summary of AI Activity on Meta.Discourse.org
Overview
This week on Meta.Discourse.org, the AI discussions have been vibrant and diverse, covering a range of topics from AI image captioning features to community feedback on Discourse AI usage. The community has been actively engaging with the new features, providing valuable feedback, and exploring the potential of AI to enhance forum experiences. Notably, discussions have revolved around improving accessibility, integrating AI functionalities, and addressing technical challenges. The active participation of users like Falco, Saif, and sam has contributed significantly to advancing these conversations.
Interesting Topics
-
How do you use Discourse AI? Saif initiated a discussion seeking community feedback on the usage of Discourse AI, aiming to gather insights to enhance the AI features.
-
AI Image Captioning Feature introduced by Falco in the Discourse AI Plugin, focuses on improving content accessibility through automatic caption generation for images.
-
GPT-4 with Vision was a topic of inquiry by tpetrov, exploring the capabilities and integration of vision features within the AI plugin, as discussed here.
-
Can Discourse AI Plugin Connect with OpenAI Assistants? This discussion led by sam and Aizada_M delved into the possibilities of integrating OpenAI assistants with the Discourse AI plugin.
-
Community Sentiment and Toxicity Queries were addressed in a topic focusing on leveraging AI for monitoring community sentiment and managing toxicity.
-
Enable Related Topics explored the technical aspects and challenges of enhancing topic relevance and connectivity through AI, as discussed by RBoy and Falco here.
-
Random Picker Persona Options was a playful yet informative conversation about utilizing AI for random selections, led by Jagster and JammyDodger.
-
Enable AI Helper saw bekircem seeking assistance with AI model configurations, as detailed here.
-
Dashboard Report - Post Emotion and Dashboard Report - Overall Sentiment were two significant contributions by SaraDev, focusing on sentiment analysis within the community, as seen in these reports and here.
-
From Private to Public: The Quest to Publish AI Conversations Seamlessly was a UX discussion initiated by Jagster, addressing the challenges of publishing AI conversations, as mentioned here.
Activity
-
Falco led the way with insights on the AI Image Captioning Feature and provided technical support across various topics including Self-Hosting Embeddings for DiscourseAI and troubleshooting for AI modules suddenly stopped working.
-
Saif engaged the community with a call for feedback on Discourse AI usage and contributed to discussions on Enabling AI Search.
-
sam provided updates and solutions on integrating OpenAI assistants with the Discourse AI plugin in this discussion and addressed technical issues in From Private to Public: The Quest to Publish AI Conversations Seamlessly.
-
Jagster initiated conversations on Random Picker Persona options and the UX challenges of publishing AI conversations here.
-
SaraDev contributed valuable insights into community sentiment through Dashboard Report - Post Emotion and Dashboard Report - Overall Sentiment.
Thanks for reading, and I’ll see you again next week!
Weekly Summary of AI Activity on Meta.Discourse.org
Overview
This week on Meta.Discourse.org, the AI discussions have been particularly vibrant, covering a wide range of topics from AI image captioning features to the integration of AI personas in chat plugins. The community has been actively engaging with the new features, providing feedback, and suggesting improvements. The discussions have highlighted the community’s interest in making Discourse more accessible and interactive through AI, as well as addressing technical challenges and exploring innovative ideas for AI integration.
Interesting Topics
-
AI Image Captioning Feature in Discourse AI Plugin: A new feature that automatically generates captions for images uploaded by users, enhancing accessibility and user experience. Discussion involved feedback on its functionality and suggestions for improvement, including making the feature more visible and customizable.
-
Support for Prompt Customization in DiscourseAI: The community discussed the potential for customizing AI prompts within the DiscourseAI plugin to better suit different languages and contexts. This topic also touched on the challenges of making prompts customizable and the technical approaches to solving them.
-
AI Personas Integration of the Chat Plugin: The idea of integrating AI personas into the chat plugin to allow users to interact with AI characters was proposed. Feedback was overwhelmingly positive, with discussions on how to implement this feature effectively.
-
AI Plugin OCR Support: There was a conversation about extending the AI plugin’s capabilities to include OCR (Optical Character Recognition) to convert different types of documents into editable and searchable data. Here, the community explored how this feature could complement the existing image captioning feature.
-
Assigning AI Personas to Respond to Topics: A suggestion was made to allow AI personas to be assigned to respond to specific topics or posts, aiming to streamline interactions and reduce manual effort. The discussion included potential mechanics and benefits of this feature.
-
Differences in Search Latency Between AI Semantic and Keyword Search: Users noted differences in search performance between semantic and keyword searches, prompting a technical discussion on search algorithms and potential improvements. Insights from the team provided clarity on the underlying mechanisms.
-
Troubleshooting AI Character Mentions: A user encountered a 500 error when mentioning an AI character, leading to a discussion on the correct configuration settings for AI mentions. The solution was shared, highlighting the importance of the “Default Language Model” setting.
-
Utilizing Discourse AI for Various Purposes: Community members shared how they use the Discourse AI plugin, from generating images to conducting Google searches. Feedback emphasized the plugin’s versatility and fun factor.
-
AI Bot Google Search Space Issue: A bug was reported regarding the AI Bot’s handling of spaces in Google search queries, affecting search results. The issue sparked a discussion on potential fixes.
-
Configuring API Keys for OpenAI and Google Gemini: Users shared challenges and solutions related to configuring API keys for OpenAI and Google Gemini, discussing technical hurdles and troubleshooting steps. Conversations provided valuable insights for others facing similar issues.
Activity
-
pmusaraj and Tris20 shared their experiences with the new AI image captioning feature, highlighting its potential for improving accessibility. See discussion.
-
Falco and sam provided insights into the development of prompt customization in DiscourseAI, addressing community feedback and outlining future plans. More details.
-
kuaza proposed the integration of AI personas in the chat plugin, receiving support from sam and merefield who discussed implementation aspects. Read more.
-
ozkn and sam engaged in a discussion about extending AI plugin capabilities to include OCR support, exploring how it complements existing features. Discussion here.
-
craigrow initiated a conversation on the differences in search latency between AI semantic and keyword search, with Falco providing technical explanations. Insights shared.
-
tpetrov reported a bug related to AI Bot’s Google search functionality, sparking a discussion on potential fixes and improvements. Bug report.
-
silvacarl shared challenges with configuring API keys for OpenAI and Google Gemini, with community members offering troubleshooting advice. Configuration challenges.
Thanks for reading, and I’ll see you again next week!
Weekly Summary of AI Activity on Meta.Discourse.org
Overview
This week on Meta.Discourse.org, the AI discussions have been vibrant and diverse, covering a range of topics from AI image captioning to addressing bugs and enhancing AI functionalities within the Discourse platform. Users have actively engaged in sharing their experiences, troubleshooting issues, and proposing new features to leverage AI more effectively. The community’s enthusiasm for exploring AI’s potential in improving user experience and content management is evident through the constructive dialogues and collaborative problem-solving efforts.
Interesting Topics
-
AI Image Captioning Fun: A lively discussion emerged around the new AI Image Captioning feature, with users sharing amusing and creative captions generated by AI for various images. This thread not only showcased the feature’s capabilities but also highlighted the community’s creativity.
-
Accidental AI Caption Button Taps on Tablets: Users reported issues with accidentally triggering the AI caption button on tablets due to its invisibility, prompting a discussion on UX improvements.
-
AI Permissions and Error Handling: There were reports of the ‘Caption with AI’ button displaying for users without permission, leading to error messages. This sparked a conversation on user permissions and error handling.
-
AI Flagging Sensitivity: The community discussed the sensitivity of AI in flagging content, with some finding it too aggressive. This discussion led to exploring adjustments to make the AI flagging more aligned with community standards.
-
Prompt Customization in DiscourseAI: Users expressed a desire for more flexibility in customizing prompts for AI-generated content, leading to a feature request and discussions on potential enhancements.
-
AI Bot PMs and Sent Box Issue: A bug was identified where AI bot PMs with only one bot response were appearing in the Sent box instead of the Inbox. This issue was promptly addressed by the community.
-
Google Gemini API Keys: There was a discussion on configuring API keys for Google Gemini, highlighting the steps and challenges involved in enabling AI features.
-
AI Search Enhancement: The community explored enhancing AI search capabilities, with a focus on semantic search and the use of different Gemini models. This conversation delved into the technical aspects and potential improvements.
-
AI-Generated Post Illustrations: Users discussed the possibility of AI-generated thumbnails for posts, expressing interest in category-specific controls for this feature.
-
AI-Related Topics Functionality: There were reports of issues with the AI-related topics feature, leading to a support request and discussions on troubleshooting and fixes.
Activity
- Saif initiated the AI Image Captioning thread and engaged in discussions on Google Gemini API keys and AI search enhancements.
- Moin reported the accidental AI caption button taps on tablets and participated in discussions on prompt customization and locale issues.
- Falco contributed to resolving the AI caption button visibility issue and addressed concerns regarding AI flagging sensitivity.
- Imgbi and Dylan_Michael discussed permissions and error handling for the AI caption feature.
- Shauny initiated the conversation on AI flagging sensitivity, leading to insights from Falco on classifier adjustments.
- Jagster and keegan discussed prompt customization and resolved a bug related to the backspace key in AI helper modals.
- ThunderThighs and sam engaged in a dialogue on AI-generated post illustrations and category-specific controls.
- Don sought support for issues with AI-related topics functionality, highlighting the community’s collaborative effort in troubleshooting.
Thanks for reading, and I’ll see you again next week!
:
Weekly AI Activity Summary on meta.discourse.org
Overview
This week saw a lot of interesting discussions and developments related to AI on meta.discourse.org. Key topics included:
- Users experimenting with generating AI image captions and images from those captions in a fun back-and-forth
- Fixing bugs with AI bot personas not accepting decimal values for temperature and top_p settings
- Troubleshooting issues with configuring AI image captioning and getting the “Caption with AI” button to appear
- Resolving problems with OpenAI embeddings and related topics not working due to disabled site settings
- Discussing potential improvements to the AI helper UX and summarization capabilities
Jagster, sam, Falco, and Bathinda were some of the most active participants in AI-related topics this week. Let’s dive into the highlights!
Interesting Topics
-
In Let’s see your best AI Image Caption!, users had fun generating AI captions for images, then generating new images from those AI-generated captions. piffy noted the AI was “surprisingly consistent” in this post.
-
Jagster reported a bug where AI bot personas don’t accept decimals for temperature, top_p settings. sam identified it was due to localization differences between Europe and the US. A fix was implemented.
-
Bathinda had issues configuring AI image captions. Falco advised to set API keys and the caption model. Jagster provided the solution of enabling the right AI helper features.
-
Don found OpenAI embedding related topics stopped working because the
ai embeddings semantic related topics enabled
site setting got disabled inadvertently, as explained here. -
keegan fixed an issue where the ‘Caption with AI’ button was displaying for users without permission.
-
Moin noticed the AI helper does not always use the user’s locale, suspecting a caching issue.
-
Jagster asked about supporting different image sizes for DALL-E 3. sam said they need to add an option for other dimensions.
-
Jagster reported OpenAI model issues when generating summaries, with some models failing to generate and others not following the user’s language.
-
Arkshine showed how large spaces between text can misalign the AI helper on selection.
-
In How do you use Discourse AI? Tell us and make it even better!, dillfrescott suggested showing an error message if Gemini declines to summarize NSFW content, rather than getting stuck loading.
Activity
-
Jagster was highly active, participating in discussions about AI image captions, reporting bugs with decimal settings and summary generation, asking about DALL-E image sizes, and helping troubleshoot AI helper issues.
-
sam provided insights on the decimal bug and acknowledged the need to support more DALL-E image sizes.
-
Falco advised Bathinda on configuring AI image captioning and pointed ninermac to the self-hosted guide for sentiment analysis.
-
Bathinda sought help configuring AI image captions and enabling related topics.
-
keegan resolved the ‘Caption with AI’ permissions issue and discussed a potential caching issue with the AI helper locale.
-
Moin reported the AI helper locale bug.
-
Don discovered and resolved an issue with OpenAI embeddings related topics.
-
Arkshine demonstrated an AI helper alignment issue caused by large spaces in text.
-
dillfrescott proposed an improvement to Gemini’s handling of NSFW content.
Thanks for reading, and I’ll see you again next week!
:
Weekly AI Activity Summary on meta.discourse.org
March 12 - March 19, 2024
Overview
This week saw a lot of activity and discussion around new AI features in Discourse, including the ability to share AI conversations via public links, a new GitHub helper persona for the AI Bot, and improvements to the AI image captioning feature. There were also questions and troubleshooting around configuring AI providers like OpenAI and Hugging Face’s LLaVa model. Overall, the Discourse team continues to rapidly iterate and expand the AI capabilities of the forum software.
Interesting Topics
- sam announced a new feature to share AI conversations via public links, with access controls announcements new-feature ai ai-bot
- A new GitHub helper persona was added to the AI Bot, allowing it to read pull requests, repository files, and perform repo searches announcements ai ai-bot
- Bathinda ran into issues configuring the AI image captioning feature to use the open-source LLaVa model instead of OpenAI’s paid offering support ai ai-helper
- sp-jordan-violet inquired about options for using alternative AI providers like self-hosted models or Mendable AI dev ai
- Oniel suggested allowing users to see token usage and create their own AI personas as potential new features plugin official ai
- RBoy encountered a rate limit error with Google Gemini after upgrading Discourse support ai
- peyrusse had API calls to OpenAI fail on an intranet Discourse install support ai
- Nadeem asked about adding a new custom chat bot using a self-hosted model endpoint support ai
Activity
- sam was very active, announcing sharing AI conversations publicly, a new GitHub AI helper, and commenting on AI moderation capabilities.
- Falco provided support and guidance on configuring LLaVa for image captioning, hitting API rate limits, and self-hosting open-source language models.
- Bathinda shared their troubleshooting process for enabling AI image captioning and eventually found a solution with help from the community.
- Jagster assisted Bathinda in resolving their image captioning configuration.
- pfaffman also helped Bathinda debug the AI image captioning setup.
- ecki suggested an AI feature to check for similar tags and recommend merging them.
- RBoy encountered and worked to resolve AI summarization failures and sentiment analysis errors after upgrading Discourse.
- Lilly pointed Bathinda to resources on generating API keys for Hugging Face to use with LLaVa.
- JammyDodger clarified which groups have access to topic summarization and how the AI search works.
Thanks for reading, and I’ll see you again next week!
Weekly AI Activity Summary (2024-03-19 to 2024-03-26)
Overview
This week saw a lot of discussion around issues with the Google Gemini LLM, with recommendations to wait for Gemini 1.5 or use alternative models like Claude or GPT-3.5 for now. There were also several new feature requests and bug reports related to the AI Bot, AI search, and AI image captioning. On the self-hosting front, the vLLM and OpenChat models were recommended.
Interesting Topics
- Issues were uncovered with using Google Gemini 1.0 as an LLM for the AI Bot, with odd outputs and hallucinations. Waiting for Gemini 1.5 or using alternative models like Claude or GPT-3.5 was suggested.
- A bug was reported where the AI helper does not always use the user’s locale, showing responses in different languages. This is due to the cache key not being based on locale.
- When sorting search results by relevance, an additional page reload is required to enable the AI search option.
- Several new features were requested for the AI Bot, including user token tracking, custom AI personas, max context posts, document loading, custom API URLs, and localized chat titles.
- The “Custom summarization allowed” group setting is broken when set to “Everyone”. It ends up disabling the button for all users including staff.
- For the AI image captioning feature, the “Capture with AI” button can block small images and surrounding text, making post editing difficult.
- An attempt was made to do a comparative cost analysis of different LLMs, with a request for Meta to share their rough AI costs as a reference point.
- After upgrading to Discourse 3.3.0 beta1, AI summarization stopped working for some. Regenerating the API key resolved it.
- There have been increasing reports of AI-based spam, with bots trying to build up reputation to bypass restrictions. Admins are needing to clean up bogus AI posts regularly.
- For self-hosting open source LLMs with Discourse AI, vLLM and the OpenChat v3.5 0106 model (7B parameters) were recommended, even running 4-bit quantized for performance.
Activity
-
RBoy shared their experience with issues using Google Gemini 1.0 and 1.5 for the AI Bot. They also reported a bug with the “Custom summarization allowed” group setting and provided more details on AI summarization not working after upgrading.
-
sam noted that Gemini 1.5 API support had not been added yet and suggested a potential fix for unforgiving prompts when the AI Bot fails to respond. They also mentioned the AI helper should now work in German.
-
Moin shared screenshots of the AI helper responding in different languages for the same user. They also reported the issue with needing to reload after changing search sort order to enable AI search.
-
Falco explained that the AI helper language issue is due to the cache key and proposed using client-side translation or caching based on locale as potential solutions. They also confirmed removing or warning about Gemini 1.0 for the AI Bot.
-
Oniel made a feature request post with several suggestions for enhancing the AI Bot, like user token tracking, custom personas, and more.
-
Saif added a warning about using Gemini for the AI Bot in its current state. They also mentioned working on user token tracking for future LLM cost calculations.
-
seanblue provided UX feedback on the AI image captioning button blocking small images and text.
-
Tris20 asked if Meta would be comfortable sharing a rough range of their AI costs to help with comparative analysis.
-
cbrandtbuffalo shared details on the increasing AI-based spam their community has seen recently and the efforts needed to combat it.
-
Isambard recommended vLLM and OpenChat for self-hosting open source LLMs, even using 4-bit quantization for better performance.
Thanks for reading, and I’ll see you again next week!
Weekly AI Activity Summary
Week of March 26 - April 2, 2024
Overview
This week saw continued discussion and development around the Discourse AI Plugin, including support for new language models, limitations of certain AI features with AWS Bedrock, and pricing considerations for different LLM providers. There was also a new guide published on setting up spam detection using the AI Post Classifier, and some impressive demonstrations of OCR capabilities with Claude 3’s vision support. Questions arose around disabling AI summarization, tweaking AI flagging sensitivity, and enabling the AI bot in chat.
Interesting Topics
- Support for new language models in the Discourse AI Plugin, like Anthropic’s Claude 3 Haiku and Sonnet via AWS Bedrock, with some limitations on AI features
- A new guide on setting up spam detection using the Discourse AI Post Classifier automation rule
- Impressive OCR capabilities demonstrated with Claude 3’s vision support, allowing interactive extraction of text from images
- Discussion of pricing for different LLM providers used with Discourse AI, and the rapid changes in pricing models
- Questions around disabling AI summarization due to quality concerns, and which models perform best for this task
- Dealing with overly sensitive AI flagging and customizing prompts to look for specific sensitive content
- Enabling the AI bot in chat and targeting specific AI personas
Activity
- sp-jordan-violet asked about support for Anthropic models and limitations of AI features with AWS Bedrock in the Discourse AI plugin
- sam shared that Claude Haiku offers performance close to GPT-4 at half the price of GPT-3.5, and that AI may be needed to combat AI-based spam
- Saif published a guide on setting up spam detection using the AI Post Classifier
- Falco clarified that Bedrock can be used for LLM needs but not embeddings/sentiment yet, and explained how to disable AI summarization
- Tris20 tested the OCR capabilities of Claude 3 on a complex command-line interface image
- Shauny asked about completely disabling AI summarization due to poor quality
- luis_villanueva inquired about targeting specific AI personas in chat
Thanks for reading, and I’ll see you again next week!
Weekly AI Activity Summary
April 2 - April 9, 2024
Overview
This week’s AI discussions on meta.discourse.org covered a range of topics, from integrating GPT-3.5 into the forum’s AI bot and setting up spam detection using AI, to the new AI Persona Editor and enabling the AI bot in chat. There were also discussions around Gemini Pro automation support, semantic quick search issues, and AI-assisted translations.
Interesting Topics
-
bdzfer proposed integrating GPT-3.5 into the forum’s AI bot to enhance user interaction and content generation. merefield noted the challenges of running such a large model on own infrastructure.
-
The new AI Persona Editor was announced, allowing creation of specialized AI personas. MarcP and sam discussed using multiple tools in one answer and prompting the AI effectively.
-
A guide on setting up spam detection using AI was shared. merefield and Saif discussed the nuances of identifying AI-generated spam vs general spam detection.
-
sam experimented with enabling the AI bot in chat, noting some current limitations. Claude 3 Haiku was recommended over Gemini 1.0 Pro for a cheaper LLM.
-
Jagster encountered a 500 error with semantic quick search which Falco noted was due to an experimental feature requiring per-post embeddings and re-ranker model.
-
Improvements were made to the AI helper’s language translation capabilities.
Activity
-
sam was the most active, with 10 likes and 5 posts across topics like integrating GPT-3.5, the new AI Persona Editor, and enabling AI in chat.
-
merefield had 9 likes and 4 posts, discussing GPT-3.5 integration challenges and AI spam detection nuances.
-
MarcP gained 4 likes across 3 posts on the new AI Persona Editor.
-
bdzfer proposed GPT-3.5 integration and discussed implementation options, receiving 2 likes over 3 posts.
-
Saif and Arkshine each had 2 likes, with Saif discussing AI spam detection and Arkshine noting a broken image.
-
Falco provided a solution to the semantic quick search 500 error, while Jagster reported the issue and thanked for the translation improvements.
Thanks for reading, and I’ll see you again next week!
Weekly AI Summary (April 9 - April 16, 2024)
Overview
This week on meta.discourse.org, there were several interesting discussions and developments related to AI:
- sam announced a new feature in the Discourse AI Persona Editor that allows uploading large bodies of text to better ground AI personas.
- There were reports of the summarize feature being broken on Meta, which sam fixed by switching to a new AI model.
- mattdm suggested improvements to the AI helper settings to allow enabling/disabling features per user group.
- mattdm also requested a change to the AI image captioning to avoid redundant phrasing in the generated captions.
Interesting Topics
-
sam announced a new feature in the Discourse AI Persona Editor that allows uploading large bodies of text to better ground AI personas with concrete data. This can help increase the quality of AI-generated results.
-
There were reports of the summarize feature being broken on Meta. Arkshine noted the summarization resource seemed to be missing. sam identified the issue was due to removing support for an old GPT-4 model, and fixed it by switching Meta to use the Claude 3 sonnet model instead.
-
mattdm suggested improvements to the AI helper plugin settings, proposing to have settings for each feature that control which user groups they are enabled for. He also suggested separating out the context menu items into individual group-based settings.
-
mattdm requested a change to the AI image captioning feature to avoid redundant phrasing like “The image contains” or “An image of” at the start of every caption. He proposed updating the prompt to not require the AI to explain it is describing an image.
Activity
-
sam was active in multiple AI-related topics:
- Announced upload support in AI Persona Editor
- Helped troubleshoot the broken summarize feature
- Provided insights on prompt engineering the search tool in AI personas
-
MarcP reported the summarize feature not working on Meta
- Also discussed styling the shared conversations page for AI personas
-
mattdm suggested enhancements to the AI helper plugin:
- Having per-feature group-based settings for enabling/disabling
- Separating context menu items into individual settings
- Improving AI image captioning to avoid redundant phrasing
-
Arkshine helped identify the missing summarization resource when troubleshooting the broken feature on Meta
-
ondrej confirmed also experiencing issues with the summarize feature timing out after several minutes
-
Saif provided an explanation of the available AI helper features and what the enable/disable setting controls for each one
Thanks for reading, and I’ll see you again next week!
Weekly AI Summary (2024-04-16 to 2024-04-23)
Overview
This week on meta.discourse.org, there were interesting discussions around enhancing the AI capabilities in Discourse. Key topics included using AI to make posts more gender inclusive, providing more control over the Discourse AI persona, separating notifications for AI bot conversations, caching AI summaries, and configuring various AI features like the AI bot, AI helper, and toxicity detection.
Interesting Topics
-
jidanni tried using the AI function to make a post more gender inclusive but found it didn’t detect the gender imbalance. Ed_S suggested including an explicit request in the draft to get better results.
-
mattdm proposed ideas to enhance the Discourse AI persona, such as including forum posts matching certain criteria in addition to manually uploaded text. sam noted this can be done with custom search commands.
-
MarcP suggested using auto-tags like
#ai-chat
to create a separate notifications tab for AI bot conversations. sam noted tags on PMs are not enabled for everyone, so a custom filter might be better. -
agitated.swanson asked if Discourse AI summaries are cached to avoid unnecessary requests. sam confirmed summaries are indeed cached.
-
Several users had questions about enabling the AI bot, enabling the AI helper, configuring API keys for Google Custom Search, and setting up the Discourse AI toxicity feature with the Google Perspective API.
Activity
-
Falco was active in the gender inclusive AI topic, noting that custom prompts should be used for custom AI tasks rather than injecting instructions in the content itself. [18 likes, 3 posts]
-
Ed_S experimented with providing instructions to the AI in the draft to rewrite text in a gender-neutral way and even in the style of Moby Dick. [17 likes, 4 posts]
-
sam provided helpful information on many topics, including that Discourse AI summaries are cached, new Mistral AI models are available with varying capabilities, and that more granular controls for AI features are planned for the future. [17 likes, 7 posts]
-
MarcP reported a bug with uploading multiple .txt files for the Discourse AI persona and suggested auto-tagging AI bot PMs for a separate notifications tab. [5 likes, 4 posts]
-
mattdm suggested enhancements to the Discourse AI persona to include forum posts matching certain criteria in addition to uploaded text. [3 likes, 3 posts]
Thanks for reading, and I’ll see you again next week!
Weekly AI Summary (2024-04-23 to 2024-04-30)
Overview
This week saw a lot of discussion around the adoption of Discourse AI and its features. Users shared their concerns about AI abuse and the cost of running AI models. There were also questions about configuring AI bots, getting related topics to display properly, and creating custom ChatGPT-based bots. On the development side, fixes were made to the AI persona editor and a guide was posted on how to get free related topics using Cloudflare Workers.
Interesting Topics
- Saif started a discussion on what is stopping people from trying out Discourse AI, with users sharing concerns about AI abuse, cost, and privacy.
- MarcP asked about how to get both suggested and related topics to display, and learned that suggested topics only show for unread posts.
- Abbas_Muhammad had trouble configuring ChatGPT for their bot and received troubleshooting advice.
- Moin reported a bug where related topics disappear when rereading a long topic.
- GiggaPoggers inquired about creating a custom ChatGPT bot as a regular user rather than an admin.
- sam posted about new upload support for AI personas and also wrote a comprehensive guide on using the AI persona system.
- Saif explained how to get free related topics using Cloudflare Workers.
- MarcP found a small bug with the AI bot button dropdown.
Activity
- Saif was very active, asking for feedback on Discourse AI adoption and posting a guide on using Cloudflare Workers for free related topics.
- sam helped troubleshoot issues with configuring ChatGPT, wrote about AI persona uploads and posted a detailed AI persona guide. He also mentioned upcoming changes to the AI bot button UI.
- MarcP asked about displaying suggested and related topics and reported a dropdown bug with the AI bot button.
- Roman_Rizzi fixed a multi-file bug in the AI persona editor.
- fokx shared a command to test self-hosted embeddings and reported a small bug with the embeddings rake task.
Thanks for reading, and I’ll see you again next week!
Weekly AI Summary (2024-04-30 to 2024-05-07)
Overview
This week on meta.discourse.org, the community discussed various topics related to Discourse AI, including barriers to adoption, use cases, and technical setup. Key themes included concerns about AI’s impact on natural community interactions, the effort required to create good knowledge bases and prompts, and the cost of using AI features. There was also interest in using AI for specific tasks like moderation assistance and semantic tagging of topics.
Interesting Topics
-
In What is stopping you from trying out Discourse AI?, users shared their reservations about adopting AI, such as the potential to disrupt organic community exchanges, the work involved in prompt engineering, and the financial cost, especially for smaller forums.
-
The Google Programmable Search Engine and Custom Search API Keys topic provided an overview of how to set up custom search functionality using Google’s tools, though some users were unsure how this integrates with Discourse AI.
-
In How to provide image generator function?, eisammy inquired about implementing image generation capabilities gated to specific user ranks, which sam confirmed is possible with the AI features.
-
The Discourse AI Post Classifier - Automation rule discussion highlighted interest in using AI to automatically apply relevant tags to new topics, even if not 100% accurate, to help bring order to forum content.
-
In Official Discourse AI Plugin, Master sought assistance troubleshooting the setup of the AI plugin and OpenAI API on their self-hosted Discourse instance.
Activity
-
Hanon_Ondricek shared their community’s stance on AI, prohibiting AI-generated content posted under the guise of a real user, while allowing disclaimed usage for writing assistance or game development.
-
Saif noted that Discourse AI aims to strike a balance, enhancing writing without completely automating it, and that branding AI features carefully can impact user perception.
-
Paul_King succinctly stated the main barrier to trying Discourse AI is the cost.
-
Lilly provided a detailed explanation of Google’s Programmable Search Engine and Custom Search API keys, including setup steps and billing considerations.
-
ncaming215 reported back after testing the AI post classifier, eager to see it semantically apply existing tags to new topics to bring more order to the forum, even if not perfectly accurate.
Thanks for reading, and I’ll see you again next week!
Weekly AI Summary (2024-05-07 to 2024-05-14)
Overview
This week’s AI discussions on meta.discourse.org focused on enabling chat integration for AI personas, barriers to adopting Discourse AI, and improvements to administrative AI tools. Users also discussed the new topic map design experiment, using Google Search with AI bots, and automating AI responses to new topics.
Interesting Topics
-
sam announced that Discourse AI now allows chat integration for AI personas, enabling them to receive DMs and @mentions if configured.
-
BenLeong shared barriers to trying out Discourse AI, including navigating internal restrictions, estimating costs, and concerns about AI-generated content. sp-jordan-violet suggested focusing on administrative AI tools to demonstrate community value to leadership.
-
The new topic map showing likes is a design-experiment and not currently configurable, as explained by Moin and merefield.
-
sam shared details on enabling chat for AI personas and plans to add “Roles” like group and topic responders.
-
Lilly demonstrated using the Google Programmable Search Engine with AI Researcher bots like Lola and Claude Opus.
-
merefield and Lilly explained how to automatically respond to new topics using AI via the chatbot plugin’s “answering machine” mode or Discourse Automation.
-
Moin suggested disabling the AI title generator when the PM title can’t be edited, as it would be pointless.
-
Hifihedgehog noted the quick implementation of a feature request in the Official Discourse AI Plugin.
-
Jagster encountered an issue with AI chat not reacting to @mentions on renamed users, likely related to the
_bot
suffix and renaming steps.
Activity
-
sam was highly active, discussing barriers to adoption, chat integration, planned “Roles” for AI personas, and investigating an issue with renamed AI users in chat.
-
sp-jordan-violet provided valuable perspective on administrative AI tools to demonstrate community value.
-
Moin clarified the new topic map design and suggested a UX improvement for the AI title generator.
-
Lilly shared examples of using Google Search with AI bots and explained automating AI responses to new topics.
-
merefield provided guidance on automatically responding to topics with AI and clarified the AI Topic Summary plugin’s functionality.
-
BenLeong and bloomexperiment discussed barriers and awareness challenges in adopting Discourse AI.
-
Hifihedgehog, piffy, Jagster, Master, and Truest participated in various AI-related discussions and support threads.
Thanks for reading, and I’ll see you again next week!
Weekly AI Activity Summary
Week of May 14 - May 21, 2024
Overview
This week saw a lot of interesting discussions and activity around AI on meta.discourse.org. Key topics included:
- Investigating discrepancies in answers from different AI models
- Estimating costs of using LLMs for Discourse AI
- Leveraging AI to improve spam filtering
- The release of GPT-4o by OpenAI
- Installation and setup of Discourse AI
- Issues with certain AI features not functioning properly
@StevePlex, @sam, @Falco, @Saif and @Jagster were among the most active participants this week.
Interesting Topics
-
@StevePlex noticed two different answers from gpt4o to the same question, sparking a discussion on model tweaking and randomness.
-
@Saif provided a helpful guide on estimating costs of using LLMs for Discourse AI, covering key factors to consider.
-
An older topic on using Discourse AI to make spam filters smarter saw an update from @sam that this feature has now been implemented.
-
The release of GPT-4o by OpenAI generated some discussion on costs and implementation.
-
@Truest sought guidance on installing Discourse AI, with helpful advice provided by @Jagster and @merefield.
-
@piffy reported an issue with the like button becoming unresponsive after using the “Ask AI” feature.
-
@Tarriqi encountered an issue with a new AI feature not functioning and sought troubleshooting advice.
Activity
-
@StevePlex was very active in the two different answers from gpt4o topic, providing detailed examples and analysis.
-
@sam provided updates and insights across several topics, including the Discourse AI spam filter, GPT-4o costs, and self-hosting embeddings.
-
@Jagster offered helpful technical guidance in the installing Discourse AI and GPT-4o not following prompts topics.
-
@Truest was active seeking help with Discourse AI installation, Azure AI Dalle-3 setup and NSFW API keys.
-
@Firepup650 provided more details on the reported “Ask AI” like button issue.
-
@MarcP reported a bug with AI chat threads not showing after the latest update.
-
@RBoy shared that the Gemini Pro summarization feature was not working.
Thanks for reading, and I’ll see you again next week!
Weekly AI Summary (2024-05-21 to 2024-05-28)
Overview
This week saw lively discussions around AI features and configuration in Discourse. Key topics included the placement of the AI conversation shortcut, issues with custom LLM models creating duplicate copies, problems with sentiment analysis backfills, and how to optimize RAG settings for specific use cases. The Discourse team was active in investigating and resolving various reported bugs.
Interesting Topics
-
sam started a discussion on whether the AI conversation shortcut icon should be a user setting, allowing placement in the header, sidebar or hidden. fokx and others weighed in with suggestions.
-
Don shared positive experiences using the Claude 3 AI model for topic summarization, finding it performs well on non-English content at an affordable price point.
-
StevePlex reported an issue where defining a new custom LLM model results in many duplicate copies being created instead of overwriting the config.
-
In a support thread, Bathinda encountered problems running the sentiment analysis backfill task. Moin and others provided guidance on proper configuration.
-
StevePlex explored using prompt injection with long-context LLMs as an alternative to RAG. An interesting discussion compared the pros and cons of the two approaches.
-
Yiffyi reported a compatibility issue between the latest Discourse AI plugin update and Azure AI. Roman_Rizzi quickly provided a fix.
Activity
-
sam was highly active, discussing the AI conversation shortcut, fixing issues with Gemini models, and investigating problems with AI chat threads.
-
StevePlex made several posts reporting bugs and exploring alternative AI architectures.
-
Bathinda sought help troubleshooting sentiment analysis and backfill issues.
-
Saif shared a new guide on using the Discourse AI Post Classifier to detect spam.
-
j127 discussed strategies for identifying AI-generated spam accounts.
Thanks for reading, and I’ll see you again next week!