Weekly Summary of AI topics

[Discourse]

Overview

This week’s AI discussions on Meta centered on quality, transparency, and UX polish across Discourse AI features—especially Ask Discourse, AI Translation / Content Localization, and inline AI helpers in the composer. A few admins noticed shifting answer quality in Ask Discourse and learned the service recently moved to a locally hosted open-weights model (Ask discourse moins performant, model change note, rationale + guidance).

On the localization front, several threads highlighted how translation progress can be confusing at scale, due to caching and multi-step processing, sparking concrete requests for improved reporting and settings (including better progress visibility and a date picker / fixed cutoff) (AI Translation Progress Graph, caching + two-step explanation, date picker request, topics vs posts pacing). Relatedly, admins asked about language fallbacks (e.g., “serve English when the user’s locale isn’t supported”), and staff confirmed this isn’t supported yet beyond the site default locale behavior (Can I force English for unsupported languages, fallback locales not supported, current fallback behavior).

A few practical/operational issues also surfaced: debugging internal LLM endpoints and allowed host settings (How to use internal AI endpoints?, allowed host suggestion), enabling the AI plugin after it got “stuck” disabled (plus clarification that on free tier it stays enabled for key functionality like spam detection) (Can’t get the AI plugin enabled, fix in progress + free-tier note), and a small but telling edge case where HTML comments get included in AI summaries (ultimately marked #wontfix, with a prompt workaround) (HTML comments are also summarized by AI, wontfix + workaround).

Finally, there was continued attention on UI integration quality for AI assist buttons in editing flows—overlapping icons and absolute-positioned buttons causing layout issues—tying back to the broader inline AI composer integration work (Topic edit interface shows overlapping icons, linked to inline integration thread, AI title suggest icon placement issue, inline integration of AI Suggestions).

---

Interesting Topics

• Ask Discourse answer quality concerns after a model change (ai, ask-discourse, Support)
  gilles reported Ask Discourse felt “less performant” lately (Ask discourse moins performant). nat confirmed a recent switch to a locally hosted DeepSeek v4 flash model (model change note), and Falco clarified the goal is fast, doc-grounded troubleshooting—not dev assistance—recommending dev harnesses like dv instead (explanation + dv recommendation, Using dv (Discourse Vibe) to configure Discourse AI in development).

• Translation progress graph accuracy, caching, and “what does progress even mean?” (translation, ai, #Data-&-reporting)
  LotusJeff found the progress graph misleading on a large site (AI Translation Progress Graph). Falco explained the page is cached to avoid timeouts and that translation involves language detection followed by translation into other locales—with early-stage detection progress no longer shown, making the initial experience feel “awful” (caching + 2-step pipeline). LotusJeff proposed a clearer “eligible vs translated” reporting model and started sharing Data Explorer SQL prototypes (reporting proposal, SQL examples), and nat noted improvements are planned (staff follow-up).

• Debugging internal/self-hosted LLM endpoints in Discourse AI (ai, Support)
  satonotdead struggled to use an internal endpoint URL that worked inside the container, but failed the Discourse AI UI test (problem report + stack trace). Falco suggested ensuring DISCOURSE_ALLOWED_INTERNAL_HOSTS contains the hostname (not only an IP range) (allowed host suggestion). Context referenced earlier Ollama-local setups too (Getting Discourse AI to work with Ollama locally).

• Locale fallback policy: forcing English for unsupported languages (translation, ai, #Feature)
  Jagster asked if unsupported locales could fall back to English without changing a Finnish-heavy site’s default locale (question). Falco confirmed “fallback locales” have been requested but aren’t supported (no fallback locales yet), and nat clarified current behavior: disabling fallback shows original language; using English fallback implies setting default locale to en (behavior details).

• AI Summarization includes HTML comments (and why that’s #wontfix) (ai, ai-summarize, #Feature)
  ばこん noticed that HTML comments (which readers can’t see) are still being summarized (report, example). Falco marked it #wontfix, suggesting a prompt tweak to instruct the summary agent to ignore comments if it matters for a given instance (decision + workaround).

• AI plugin stuck disabled + free-tier “AI plugin can’t be disabled” clarification (ai, Support)
  ondrej couldn’t re-enable AI after disabling it earlier (issue). keegan confirmed a fix is needed for the “stuck disabled” state and clarified that on free tier the AI plugin remains enabled because it powers critical features like spam detection, while individual AI features can be toggled off (fix + policy). The request also referenced existing setup documentation (documentation excerpt source).

• AI UI polish: overlapping icons and absolute-positioned AI buttons (ai, ux)
  Moin reported stacked/overlapping icons when editing topic info (overlap report), and Falco linked it to the ongoing inline AI integration thread (cross-reference, Inline integration of AI Suggestions (in Composer)). In a related UX issue, Moin noted the AI title suggestion button can stay fixed while the title field shifts—especially in translated-title editing—and that suggestions appeared in English (title icon placement); chapoi later called out the position: absolute button as a recurring cause of issues (follow-up).

• Why topics translate faster than posts (and what to tweak) (ai, content-localization, Support)
  LotusJeff observed topic data translating months back while posts lagged behind (question). Falco explained batches are similar sizes, but topics naturally complete faster, especially with many-post topics, and suggested adjusting max age or backfill rate to help posts catch up (explanation + knobs).

• Feature request: date picker (fixed cutoff date) for AI Translation settings (ai, dynaloc, content-localization, #Feature)
  In the date-picker discussion, mcwumbly noted the setting might need to be “translate all posts after {date}” rather than “days backfill” (framing). LotusJeff argued rolling windows can cause older edited content to drift out of scope, and also highlighted the “topics translated but posts not yet” mismatch as a user-facing problem (rolling window concern, additional rationale).

• How sentiment/emotion classification is assigned in the Sentiment dashboard (ai, ai-sentiment, Support)
  fzngagan summarized which models power sentiment and emotion classification and explained the difference between model-based classification and agent-based strategies via site settings (sentiment classification overview), pointing back to the referenced documentation snippet about current models (models reference) and noting admins can create new agents at /admin/plugins/discourse-ai/ai-agents rather than editing defaults (agent approach context).

---

Activity

• Falco clarified product positioning and tradeoffs across multiple AI surfaces: the Ask Discourse model swap and why dv is better for dev work (Ask discourse moins performant, Using dv (Discourse Vibe) to configure Discourse AI in development); translation progress caching and the two-step pipeline (AI Translation Progress Graph); internal LLM endpoint host allowlisting (How to use internal AI endpoints?); locale fallback limitations (Can I force English for unsupported languages); summarization HTML comments tradeoff + workaround (HTML comments are also summarized by AI); composer/edit UI overlap triage by linking it back to ongoing work (Topic edit interface shows overlapping icons, Inline integration of AI Suggestions (in Composer)); and why topics translate faster than posts with tuning suggestions (Why topics translate faster than posts).

• nat confirmed key operational details and next steps: Ask Discourse moved to a locally hosted DeepSeek model and would be checked (Ask discourse moins performant); translation reporting improvements are on the radar (AI Translation Progress Graph); locale fallback behavior depends on site default and supported locales (Can I force English for unsupported languages); and troubleshooting a repeated-bot-reply issue by recommending updates and migration from Automation to the newer workflow plugin due to idempotency concerns (使用 ai 角色对帖子进行分类…, suspicious timings discussion).

• LotusJeff drove several translation deep-dives: reporting confusion and caching implications in the progress graph (AI Translation Progress Graph, follow-up proposal); shared early Data Explorer SQL to quantify eligible vs translated content (SQL examples); asked why topic translations race ahead of post translations (Why topics translate faster than posts); and reinforced why a fixed cutoff date would reduce “translated topic / untranslated posts” periods (Request for a date picker in AI Translate settings, rolling window concern).

• gilles raised the initial signal that Ask Discourse output quality felt worse, prompting staff to confirm the underlying model change and intended use case (Ask discourse moins performant, follow-up reaction).

• satonotdead documented a tricky internal endpoint integration bug where the endpoint worked inside the container but failed via the Discourse AI UI test, sharing stack traces and config attempts (How to use internal AI endpoints?, continued debugging request), and referenced prior Ollama-local guidance (Getting Discourse AI to work with Ollama locally).

• Jagster prompted a clear statement of current limitations around fallback locales—highlighting a real-world “Finnish-default site, but serve English to unsupported languages” scenario (Can I force English for unsupported languages).

• ばこん identified an AI summarization edge case: HTML comments being summarized even though they aren’t visible in the rendered post (HTML comments are also summarized by AI, example clarification).

• ondrej reported being unable to re-enable the AI plugin after previously disabling it, and questioned the “critical on free plan” reasoning after disabling spam detection on a small private forum (Can’t get the AI plugin enabled, follow-up question).

• keegan confirmed the AI plugin “stuck disabled” behavior needs a fix and explained the free-tier constraint: AI plugin remains enabled to power key functions like spam detection, while allowing admins to disable other AI features in the UI (Can’t get the AI plugin enabled).

• Moin flagged two UX regressions around AI buttons in editing flows: stacked icons in topic edit UI (Topic edit interface shows overlapping icons) and the AI title-suggest button’s awkward positioning during translated title edits (including English suggestions in a German editing context) (AI title suggest icon is placed above the title input field).

• chapoi echoed that the absolutely positioned AI button has been a recurring source of layout issues and indicated they’ll explore fixes when time allows (AI title suggest icon is placed above the title input field).

• mcwumbly helped refine the translation-date-picker feature framing by pointing out that “days backfill” is inherently rolling, implying a separate “after date” setting might be clearer (Request for a date picker in AI Translate settings).

• fzngagan provided a practical breakdown of how sentiment/emotion labels are computed, including which Hugging Face models are currently used and how to switch to agent-based classification via settings (Classification in the Discourse Sentiment Dashboard, models reference).

• singi2016cn reported a case where an AI-driven automation unexpectedly replied many times in quick succession, leading to suspicion of retries/idempotency problems and guidance to update and consider migrating from Automation to Workflows (使用 ai 角色对帖子进行分类…, follow-up evidence, solution recommending update + workflows).

Thanks for reading, and I’ll see you again next week!