Weekly Summary of AI topics

News and Events Forum summaries
113

Overview

This week’s AI conversations on Meta centered on making Discourse AI clearer to users and easier to operate at scale. On the product side, there was strong momentum to rename “AI Persona” to the more widely understood “AI Agent” (with translation workflow implications) in Renaming AI Persona → AI Agent and follow-ups like Renaming AI Persona → AI Agent. Admin experience also got attention: sites with AI disabled were still seeing AI dashboards/reports, which was confirmed as a bug and routed into broader reporting work in Don’t show AI reports if AI is not enabled and the related umbrella thread Admin Reporting & Analysis: Incremental Changes.

Operationally, the community dug into cost/performance controls and scaling pain points: Discourse rolled out OpenAI/Azure provider service tiers in Service tiers on Open AI providers, while a large self-hosted instance reported severe load when turning on semantic embeddings search in Enabling AI search crippled my server. There was also continued refinement around AI-assisted UX—especially where AI touches localization and editor UI—in Saving translations by AI-helper as content localization and The title suggester :star: button is placed outside of the title field when editing a translated title.

Finally, AI adjacent ecosystem work continued with MCP tooling: a practical setup guide landed for Codex CLI in Discourse MCP Setup in OpenAI Codex CLI, and was cross-linked back into the canonical announcement thread Discourse MCP is here!.

Interesting Topics

Activity

In total (last 7 days): 6 new topics and 25 posts, with the heaviest engagement around naming/UX polish and practical scaling/cost controls for embeddings and OpenAI usage—see Renaming AI Persona → AI Agent, Service tiers on Open AI providers, and Enabling AI search crippled my server.

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

114

Overview

Over the past week (2026-03-09 → 2026-03-16), Meta’s ai discussions clustered around product polish, reliability, and “real world” operations.

On the product side, Discourse moved closer to standardizing terminology by implementing the rename from AI Persona to AI Agent (Renaming AI Persona → AI Agent). On the infrastructure side, Discourse significantly expanded capacity for its hosted LLM offering—raising limits across all tiers and improving model quality and latency characteristics (Unlock All Discourse AI Features with Our Hosted LLM).

Meanwhile, operators focused on how AI fits into community rhythms: a request to delay AI Agent replies (so they feel less like a chatbot and more like a participant) surfaced both as a new Support topic (Adding a configurable delay to AI Agent responses) and as a follow-up in the longer-running “Agents” guide thread, where Discourse staff indicated that delayed responses would likely belong in a future automation overhaul rather than ai itself (AI bot - Agents).

Integration conversations had a notable bump too: Google’s Programmable Search / Custom Search constraints and deprecations are forcing a rethink of web search tooling, with Discourse exploring alternative providers and even “native search tools” from LLM vendors (Google Search for Discourse AI - Programmable Search Engine and Custom Search API). In parallel, community guides continued to expand around the Discourse MCP ecosystem, including a newly posted OpenCode CLI setup walkthrough (Discourse MCP Setup in OpenCode CLI).

Finally, practical admin workflows came up repeatedly: improving observability for AI spam detection via direct database queries (Discourse AI - Spam detection), questions about sentiment analysis backfilling and debugging (Problems setting up Sentiment), and GDPR-oriented concerns about sentiment processing depending on provider/configuration (Introducing Discourse AI Sentiment Analysis: New Admin Report Available). There was also an open Support thread (in Chinese) on tool-call timeouts, still in the “need more details” stage (Discourse ai 的工具调用超时如何解决？是否可以调整discourse超时时间，如何调整？).

Interesting Topics

Activity

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

115

Weekly AI Summary for meta.discourse.org (2026-03-16 → 2026-03-23)

Overview

AI discussions this week clustered around practical UX and cost-control improvements, especially for translation workflows and summarization placement. On the translation side, Shauny proposed a smoother per-post “translate” affordance plus a way to save/cached translated output to avoid repeated API spend (Translate post with AI and save translation), with Moin linking the idea to earlier localization thinking (Translate post with AI and save translation, Saving translations by AI Helper as content localization).

On the summarization UI front, Ivan_Rapekas shipped a theme component that adds the AI summary action into the topic header / sidebar timeline area, and tied it back to longstanding requests about summary button placement (AI summary in topic header, Feedback: Move summarize button at the top of the topic, Summarize button placement on mobile views).

Several threads focused on polish and reliability in AI admin settings: wording glitches like the repeated “Default LLM” error label were acknowledged and queued for fixing (Why is ‘Default LLM’ repeated…, Why is ‘Default LLM’ repeated…), and i18n layout issues in the LLM cost configuration UI (German) continued to be refined (Field alignment issues… in German, Field alignment issues… in German).

Meanwhile, the community revisited agent safety boundaries (notably concerns around AI acting “as a user” without admin oversight) (Discourse官方会出个官方的openclaw skill么？, openclaw plugin for discourse integration), and tackled integration constraints like tool-calling timeouts and connecting Discourse AI to self-hosted RAG/knowledge bases (Discourse ai 的工具调用超时如何解决？, Discourse ai 如何引入自建知识库RAG？). There was also a small but notable question about whether Discourse MCP can access PDF attachments via the protocol (Discourse MCP is here!).

Interesting Topics

Activity

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

116

Overview

This week’s AI activity on Meta Discourse centered on making AI-powered localization more accurate and predictable, especially for small-but-important UI surfaces like tags and categories. Moin surfaced several “LLM-without-context” translation failures in AI-generated tag translations do not work perfectly, prompting nat to consider prompt improvements and adding extra grounding context such as tag descriptions (reply), while Falco explored tool-assisted approaches like letting the agent read relevant sources (idea, follow-up). Related “keep translations in sync” feedback also landed as feature requests for category name and category description updates (category names, category descriptions).

On the configuration side, a troubleshooting thread revealed confusion around what kinds of PMs get translated and how the UI communicates that. In Help me troubleshoot why AI is not translating PMs on my site, Moin clarified the current limitation (group PMs vs 1:1 PMs) (details), while Falco proposed a clearer multi-choice setting (proposal) and nat hinted that upcoming “translate these categories” controls could reshape the settings UX (plan).

Finally, there were incremental improvements and ecosystem enhancements: clearer messaging for semantic vs exact search results (search clarification), interest in refining AI persona behaviors to reduce noise (mention-only request), and continued adoption of AI summaries in UI via a theme component (feedback).

Interesting Topics

Activity

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

117

Overview

This week (2026-03-30 → 2026-04-06) on meta.discourse.org saw Discourse AI discussions cluster around three big themes:

  1. MCP momentum and agent capabilities: Discourse AI doubled down on the Model Context Protocol with the announcement of client-side MCP support—letting Discourse AI agents call out to external MCP tool servers (Bring your own MCP!) and a full admin guide (AI Bot – Bring Your Own MCP Server). In parallel, the server-side MCP tooling kept evolving, including adding an edit tool so LLMs can update existing posts/wiki content via MCP (Discourse MCP is here!).

  2. Moderation and privacy boundaries in AI automation: A practical moderation question—whether AI triage can scan private messages (DMs)—ended up being a UI/configuration gotcha rather than a hard limitation, and sparked follow-up ideas for clearer controls in the automation UI (Does AI triage automation scan DMs between regular users?, solution).

  3. Model-specific quirks in localization and embeddings: Multiple threads highlighted that “AI features” are often “model behavior + integration details.” Translation issues ranged from German “AI commentary / thinking text” leakage that was fixed quickly (AI Commentary on German Translations) to missing images when translating via Mistral Small, which was mitigated by switching models (Images missing in translated posts when using Mistral as translation model). On the embeddings side, Mistral’s API mismatch (dimensions vs output_dimension) surfaced in configuration (Use Mistral for embeddings). There were also real-world admin bumps caused by deprecated Gemini model IDs in AI bot setups (Issue with AI bots forum bots).

Interesting Topics

  • Discourse AI agents can now connect to any MCP server (“Bring your own MCP”) (ai, #Announcements)
    sam announced that Discourse AI agents can register external MCP server URLs (GitHub, Notion, Linear, search providers, etc.) and then use discovered tools directly from the LLM agent (Bring your own MCP!). The companion how-to explains setup, tool discovery, and how this differs from JS-based custom tools (AI Bot – Bring Your Own MCP Server).

  • MCP usability: request for “remote/web MCP” + adding the ability to edit existing posts (ai, mcp, blog)
    In ongoing MCP feedback, pacharanero explored how MCP could be made more accessible to non-CLI users via a web-published endpoint (Discourse MCP is here!). jrgong highlighted a KB/docs use-case needing edits to existing topics/posts (ref), and Falco confirmed an edit tool was added (“just update to latest”) (ref).

  • AI triage moderation + DM scanning: “Include personal messages” works, but ‘All topics’ caused confusion (automation, ai, Support)
    Denis_Kovalenko tested “Triage posts using AI” and found PMs between regular users weren’t being scanned (Does AI triage automation scan DMs between regular users?, test details). RGJ confirmed the PMs weren’t reaching audit logs and identified the workaround: leave “Topic Type” empty rather than “All topics” (ref). The fix worked immediately (ref), and the thread turned into a UX discussion about clearer options (ref, ref).

  • Translated German posts included “AI commentary/thought process” text—quickly fixed (ai, content-localization, bug, fixed)
    putty reported German translations leaking “thinking/translate” commentary into output (AI Commentary on German Translations). nat shipped an update to tighten formatting and cleaned up affected content (ref), with user confirmation afterward (ref).

  • Mistral translations dropped images in translated views (upload:// links), resolved by upgrading model (ai, content-localization, Support)
    Denis_Kovalenko found that switching the translation model from OpenAI to Mistral caused translated versions to render text but omit images (Images missing in translated posts when using Mistral as translation model, behavior details). RGJ suggested prompt hardening and/or trying a better model (ref), and switching from Mistral Small → Mistral Large fixed it (ref). Later, Falco asked for clarification on which “Mistral Small” was meant and recommended using stronger small-class models if needed (ref).

  • Embeddings with Mistral: OpenAI-compat config breaks on dimensions parameter naming (ai, #Feature)
    RGJ documented that configuring Mistral embeddings through an OpenAI-shaped integration fails if Discourse sends dimensions, because Mistral expects output_dimension (Use Mistral for embeddings). Removing the parameter makes the test succeed, suggesting a compatibility layer or provider-specific mapping may be needed (ref).

  • AI bot errors traced to deprecated Gemini model IDs + guidance for image generation models (ai, ai-bot, Support)
    ice.d ran into “Not found” errors with legacy bot configuration (Issue with AI bots forum bots). Lilly pointed out likely deprecation of gemini-2.5-flash-pre and suggested updating model URL/ID (including an image-capable option) (ref, config example), with NateDhaliwal sanity-checking whether any LLMs were configured (ref).

  • Should AI personas reply only to @mentions? Team leans toward workflows rather than niche toggles (ai, ai-bot, #Feature)
    In an existing feature request, sam questioned whether “reply only to @mentions” is better as a default than as another setting (Allow AI Persona/Agent to respond only to @mentions…). Falco argued that edge cases are better served by upcoming project workflows—e.g., a mention-trigger workflow can handle the behavior without adding more switches (ref).

  • Agent response delay: workflows are expected to cover timing controls (ai, Support)
    sam noted that configurable delays for AI agent responses are the type of thing workflows should support, though not immediately; otherwise, the API path requires custom dev (Adding a configurable delay to AI Agent responses).

  • User-level control over AI (“disable AI nudges”) and PM translation settings migration (ai, ai-summarize, content-localization, ux/#Feature)
    paco argued that a per-user equivalent to discourse_ai_enabled could help people opt out of AI UI nudges without disabling AI site-wide (User Interface Preferences: include setting to disable AI nudges). Separately, translation settings changes continued to evolve around personal messages: nat linked a migration PR and described how prior “public content only” settings map into new category + PM targeting controls (AI translation of all PMs).

Activity

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

118

Overview

This week’s AI-focused activity on Meta (covering 2026-04-06 → 2026-04-13) centered on practical integration details—especially around AI discoverability files, provider/model choice for GDPR-sensitive deployments, and translation robustness.

On the “AI discoverability” front, the community dug into a real-world conflict between the community ai #Plugin for generating llms.txt and Discourse core’s newer (and currently limited) native llms.txt routing: pacharanero reported the override behavior in :robot: Discourse llms.txt Generator Plugin, Ivan_Rapekas confirmed the breakage in the same thread, and kaktak committed to an update to restore plugin behavior in their follow-up. Related context was cross-posted to the core discussion on native support in enabling native llms-txt support in Discourse.

In parallel, there was continued emphasis on model/provider selection for embeddings and translation, especially for communities needing strong EU/GDPR alignment. In Use Mistral for embeddings, Falco shared a working configuration and suggested considering stronger embedding models; and in Images missing in translated posts when using Mistral as translation model, provider options and “zero data retention” surfaced as part of deciding what’s acceptable for compliance and risk.

Finally, translation quality issues got very “hands-on”: a new bug report described a cooked/markup error after translation, and Moin traced it to Markdown table formatting—fixing the source table resolved the translated output in Cooked error after translate and was confirmed by cuo_wu in the resolution.

On the product-usage side, admins continued exploring AI Personas/Agents behavior controls—specifically how to stop agents replying instantly and how to constrain them to “mention-only” patterns. That discussion linked together Allow AI Persona/Agent to respond only to @mentions, not to replies to its posts and a workaround shared in Adding a configurable delay to AI Agent responses.

Interesting Topics

Activity

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

119

Overview (2026-04-13 → 2026-04-20)

AI-related discussion on meta.discourse.org this week centered on translation reliability and localization workflows, with most of the action in bug and Support threads using the ai, dynaloc, and content-localization tags. The biggest theme was intermittent, hard-to-reproduce translation failures—including languages being skipped at random and backend errors—prompting debugging suggestions like enabling hidden verbose logging and inspecting /logs (see AI Translation skips Portuguese (pt) locale, follow-up debugging, and backend error report).

There was also a practical support thread on language detection and manual overrides when posts are mixed-language (German + English titles), and how translation can appear “broken” due to external configuration issues like outdated API keys (see Post not being detected as German and the resolution). Separately, an admin-only locale-switching error turned out to be caused by a stale theme preview query parameter in Chrome (see Error when switching locale and the fix).

On the “AI platform” side, there was renewed interest in Discourse MCP connectivity (including Claude connectors and HTTP availability) (see Discourse MCP is here!, and confirmation that HTTP is supported). Finally, the long-running AI agents how-to thread received a new question about custom agent skills for tailored scenarios (see AI bot - Agents).

Trendline: most “AI issues” this week weren’t about output quality—they were about operational robustness (job behavior, retries, backend availability, and configuration visibility) (e.g., skipped translations, verbose logging, and retry behavior questions).

Interesting Topics

  • AI translations intermittently skip locales (initially observed as Portuguese missing) in bug
    Denis_Kovalenko reported that enabling many locales could lead to Portuguese not being generated (and later: any locale being skipped randomly), with titles and bodies translating inconsistently (see the original report: AI Translation skips Portuguese (pt) locale, clarifying settings: supported locales question, and the “randomly skipped locale” update: inconsistent results).
    Debugging moved toward logs and deeper internals: nat suggested checking /logs and enabling the hidden ai_translation_verbose_logs setting (see hidden verbose logs suggestion), while RGJ later surfaced backend failures (503 unreachable_backend) affecting tags/topics/posts (see error output). The thread also raised implementation questions about why translation jobs are configured with retry: false (see retry question).

  • Using hidden settings to troubleshoot AI translation logging
    When Denis_Kovalenko couldn’t find ai_translation_verbose_logs in admin search (see can’t find setting), Moin explained this is expected because it’s a hidden site setting, pointing to the documentation on hidden settings (see pointer to hidden settings doc and the referenced guide: Using hidden site settings). Shortly after, RGJ enabled the setting to aid investigation (see enabled verbose logs).

  • Mixed-language posts can confuse detection; manual language selection does force detection in Support
    putty shared a case where a German post wasn’t being translated, asking whether selecting German forces the language (see problem report). Falco confirmed that selecting a language does exactly that, and noted the post was mixed English/German with English titles influencing detection (see confirmation + explanation).

  • Translation “not working” traced to configuration (API key / provider) rather than the feature itself
    In the same thread, putty initially saw no translation populate even after forcing it (see forcing translation didn’t help) and later noticed an error about the translated title being missing (see title missing error). Ultimately, the issue resolved when they corrected their translator setup (an old API key during a Claude plan switch) and switched back to CDCK’s LLM—after which title translation worked (see solution).

  • Composer UX change: locale selector moved into the composer toolbar
    Moin clarified that the language dropdown was moved into the composer toolbar, linking it to a core change (see before/after screenshots + PR reference). This came up while discussing translation workflows and manual entry (see follow-up preference discussion).

  • Admin-only “topic doesn’t exist / preview theme” error when switching locale is caused by a stale preview_theme_id
    Denis_Kovalenko reported an admin-only issue: switching interface language in a topic showed a persistent error about previewing a theme that doesn’t exist (see report). pmusaraj diagnosed it as a stuck ?preview_theme_id=ID parameter in Chrome (see diagnosis), and removing it resolved the issue (see solved confirmation).

  • Translation quality & limits: post size/context window, and model recommendations
    While debugging sporadic translation gaps, nat mentioned a separate scenario where titles translated but bodies were skipped due to body size, and suggested checking the LLM context window settings; they also strongly advised against using “GPT mini” for translations based on customer feedback and early testing (see model + size/context notes). Denis_Kovalenko confirmed they had a very large context window configured (see context window detail).

  • Discourse MCP connectivity: request for Claude.ai connector support; HTTP already supported
    In the blog thread about MCP, putty asked whether an HTTP/SSE streaming version of the Discourse MCP server might be released to use as a connector in Claude.ai Chat (see question). Falco replied that HTTP support already exists and pointed back to earlier replies in the announcement thread (see HTTP supported response).

  • AI Agents extensibility: request for custom skills in AI bot agents
    赤丸的小烧酒 asked (in Chinese) whether agents can add custom skills for different scenario replies, seeking the ability to customize their own AI agent behavior (see custom skills request).

Activity

  • Denis_Kovalenko drove two localization/AI troubleshooting threads this week:

  • pmusaraj focused on diagnosis and narrowing down configuration causes:

  • nat provided feature-level debugging guidance and model caveats:

    • Suggested checking /logs and enabling hidden verbose logging in AI Translation skips Portuguese (pt) locale, and discussed size/context-window considerations plus model recommendations in this deeper dive.
    • In Post not being detected as German, challenged an unrelated-commit hypothesis, asked about the LLM in use, and flagged Gemini deprecations contextually (within the same reply).
    • Requested a new topic for a separate issue report once the German-detection thread was solved (see request).

  • RGJ helped operationalize debugging and surfaced concrete failure signals:

  • Moin pointed to docs and clarified UI changes affecting localization workflows:

  • putty contributed heavily across translation support and MCP discussion:

  • Falco answered usage questions and clarified MCP capabilities:

  • canbekcan explored translation workflow issues and hypotheses around recent changes:

    • Suggested a “select language first, then add title/content” workflow and described needing to recreate language options in Post not being detected as German.
    • Investigated a “missing title” problem, initially suspecting theme-related behavior in this reply, then reported they could reproduce errors and referenced recent code changes in this post.
    • Clarified they don’t use AI translation (academic requirements) and closed out their participation after UI clarification in this note.

  • 赤丸的小烧酒 added an AI agents product-direction question by asking about agent extensibility through custom skills in AI bot - Agents.

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

120

Overview

This week’s AI-related discussion on meta.discourse.org clustered around making Discourse AI features more reliable, more automatable, and cheaper to integrate with external LLM workflows. On the reliability front, admins dug into why translations get skipped or appear “stuck”—including transient provider failures and rate limits—plus practical debugging steps like enabling verbose logs and checking audit/log tables (AI Translation skips Portuguese (pt) locale, What happens to translations when LLM changes?). On the automation side, there was a new how-to guide for auto-tagging topics using AI triage + Discourse Automation (Tag topics using AI) and a deeper dive into how agent “Examples” are interpreted (and how that can accidentally cause over-flagging) (AI triage examples not sent properly?). Finally, integration and efficiency got a boost with a new plugin that serves cooked content as Markdown, reducing token costs for downstream LLM use and potentially pairing well with API/MCP usage (Discourse to Markdown Plugin, Discourse to Markdown Plugin).

Interesting Topics

Activity

Extra reference links mentioned this week (context for the above threads)

These were directly referenced inside the week’s discussions and help explain the surrounding ecosystem: Discourse AI plugin, Discourse Automation, LLM settings guide, AI bot personas / Agents, Discourse MCP is here, Discourse Chatbot now smarter than ChatGPT, and the referenced AI search 500 thread (All AI functions are working ok but AI search gives 500 error).

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

121

Overview

AI-related discussion on Meta this week centered on polishing new UX and automation workflows, and tightening correctness guarantees in AI-assisted editing and translation.

The biggest thread was a hands-on bug hunt of the new docked AI composer: Lilly documented issues around editing, quoting, mobile scrolling, and file uploads in New ai docked composer, with keegan iterating quickly on fixes and ultimately gating the feature while it’s refined (update). In parallel, the AI Helper proofreading flow saw discussion about preserving quoted text—especially important for sensitive or exact citations—with confirmation that a fix landed and additional configuration suggestions followed (Proofread breaks quotes, examples guidance).

On the operations side, admins compared notes about translation jobs getting “stuck” due to LLM rate limits and configuration questions in What happens to translations when LLM changes?. And on the enablement front, Discourse published a new how-to showing how to auto-categorize topics by combining Discourse AI triage with Discourse Automation (Auto-categorize topics using AI), while a plugin discussion explored serving text/markdown to make AI/MCP consumers happier (Discourse to Markdown Plugin).

Interesting Topics

  • Docked AI composer regressions + rapid fixes (bug, ai, composer): Lilly reported that the new docked composer could block editing actions, behave oddly when quoting, and feel inconsistent vs the regular composer—especially around line breaks (report, Shift+Enter feedback). keegan shipped multiple fixes and follow-ups while explaining intended behavior and next steps (fixes summary, gating announcement).

  • RTE-first design decision in the docked composer (Markdown supported, but no preview): There was clarification that the docked composer is intended to be primarily RTE, while Markdown remains available but without a preview due to space constraints (design explanation, confirmation).

  • Quoting + sidebar + navigation edge cases when interacting with the bot UI: Quoting the bot was implicated in sidebar gaps/disappearing UI and even trapping users in the bot conversation, which then improved after subsequent fixes (initial behavior, later status).

  • File uploads failing after the first post in the docked composer: After other issues improved, Lilly narrowed remaining problems to uploads failing after the first post, plus intermittent quoting issues that later cleared after a rebuild (bug report, triage update, maintainer response).

  • AI proofreading should not “improve” quoted text (bug, ai-helper): bksubhuti highlighted the risk of AI altering quoted religious/source text and argued that quotes must be preserved verbatim (concern). Falco pointed out the issue was fixed and suggested trying a better model if it still reproduces (fix reference).

  • Configuring proofreader agents with examples + specialized personas: bksubhuti shared a specialized Pāḷi-oriented persona prompt and asked about engine choices (persona details), while Falco asked whether they were using examples, noting the default proofreader ships with multiple examples to help ground quote-handling (examples suggestion).

  • Translation jobs stalled by rate limits + confusion over “thinking” settings (Support, ai): In a translation troubleshooting thread, Falco suggested disabling “thinking,” while RBoy asked what that meant in the Discourse AI UI and shared an error showing token-per-day rate limiting causing repeated failures (suggestion, rate limit error, UI question).

  • Serving Markdown for better AI/MCP consumption (#Plugin, markdown, ai): The Discourse-to-Markdown plugin thread explored “content negotiation” as a clean path for AI clients: try Accept: text/markdown against canonical URLs, then fall back to JSON API behavior if unsupported (proposal, follow-up). The same discussion explicitly connected this to MCP usage (see also Discourse MCP is here).

  • AI-generated images improving in quality (and prompt-sharing interest): In a long-running support-bot discussion, 37Rb noted a big jump in image generation quality compared to prior attempts (experience), and EricGT encouraged sharing prompts and tips more broadly (request).

  • New how-to: auto-categorize topics using AI triage + Discourse Automation (#Site_Management, automation, ai): Discourse published a guide detailing prerequisites (Discourse AI, Automation, configured LLM, and an Agent/persona) and the overall workflow for using AI to decide whether a topic belongs in a different category (guide; see prerequisite references to Discourse AI, Discourse Automation, the LLM settings guide, and AI bot personas).

Activity

  • Lilly led a detailed QA pass on the docked AI composer, documenting initial breakages (New ai docked composer), acknowledging ongoing fixes (follow-up), and then narrowing down remaining issues like quoting and uploads (status, rebuild result). She also flagged the “upload after first post” regression as the main remaining bug (report).

  • sam acknowledged the docked composer feedback loop and relayed that fixes were actively underway, pointing to ongoing work by keegan (response).

  • keegan implemented and coordinated fixes for the docked composer, explained the intended RTE-first UX and Markdown tradeoffs (explanation), and later gated the feature behind upcoming changes while polishing continued (update).

  • bksubhuti raised a correctness/ethics angle: AI proofreading must preserve quoted blocks, especially for exact religious/source citations (concern). After updating, they confirmed behavior and continued experimenting, including sharing a custom proofreader persona and asking for model suggestions (confirmation, testing, persona prompt).

  • Falco provided targeted troubleshooting across two areas: he pointed to a shipped fix for proofread/quote handling and recommended trying a better model if issues persisted (Proofread breaks quotes), asked about using examples to ground the agent (examples), and suggested disabling “thinking” to address translation behavior (What happens to translations when LLM changes?).

  • RBoy brought real-world translation ops pain: they shared that translation attempts were repeatedly failing due to token-per-day rate limits and asked what “thinking” refers to in the Discourse AI configuration UI (error report, clarification question).

  • benword expanded on how the Discourse-to-Markdown plugin could support AI/MCP consumers via HTTP content negotiation, outlining a pragmatic “try Markdown then fall back to JSON” strategy (Discourse to Markdown Plugin) and tying it to MCP integration possibilities (related: Discourse MCP is here).

  • jrgong confirmed that the suggested “content negotiation then fallback” approach was essentially what another LLM (Claude) implemented for them already (reply).

  • 37Rb shared positive field feedback that AI image generation had improved substantially, citing fewer artifacts compared to earlier attempts (support bot discussion).

  • EricGT amplified community learning by asking for prompt-sharing and tips based on 37Rb’s results (request).

  • Discourse published a new admin-facing guide for using AI triage to auto-categorize topics, explicitly documenting prerequisites and setup references (Auto-categorize topics using AI; related docs: Discourse AI, Discourse Automation, LLM settings guide, AI bot personas).

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

122

Overview

AI discussions on meta.discourse.org this week (2026-05-04 → 2026-05-11) centered on operational reliability of ai translations—especially how “thinking/reasoning” models can break locale detection, exhaust completion budgets, and leave translation jobs stuck or failing in confusing ways (see AI translations errors and What happens to translations when LLM changes?). A second thread of debugging focused on token-limit surprises (request-size vs response-size, TPM/TPD rate limits) and how those interact with Discourse AI’s LLM configuration (see AI exceeds LLM token thresholds randomly and unpredictably).

On the UX/configuration side, there were smaller but practical updates: proofreading prompt customization for quote-handling (see Proofread breaks quotes), a PM/Agent follow-up image upload regression that’s already fixed upstream (see Can’t post image in Agent follow-up via PM), and a new user setup question about Google Gemini model configuration on the LLM settings page (see Discourse AI - Large Language Model (LLM) settings page).

Overall: 26 new posts across 3 new topics, with most activity coming from Falco and RBoy, largely in bug and Support threads tagged ai (e.g., AI translations errors, AI exceeds LLM token thresholds randomly and unpredictably, and Can’t post image in Agent follow-up via PM).

Interesting Topics

  • Translation failures caused by “thinking” output consuming the completion budget
    RBoy reported translation errors like Validation failed: Raw can't be blank, Cooked can't be blank in AI translations errors, and Falco identified that reasoning tokens can “eat all the tokens” under max_tokens, leading to empty/invalid outputs (analysis in context). The debugging discussion also touched on why Discourse avoids reasoning for translations (same thread) and referenced prior history around large-scale translation failures (related discussion).

  • Translations “stuck” after an LLM change: locale detection is fragile with thinking models
    In What happens to translations when LLM changes?, RBoy described translations showing as incomplete with no new logs or progress (stuck symptoms). Falco explained the underlying issue: you can’t use a thinking model for locale detection, because “thinking blocks” break parsing and without locale detection translations don’t proceed (root-cause explanation; conclusion acknowledged in follow-up).

  • Structured outputs requirements for translation models (e.g., json_schema)
    After switching to a non-reasoning model, RBoy hit a 400 indicating the selected model does not support response_format: json_schema (error report). Falco clarified that translations need a model supporting structured outputs—“basically every SotA model released recently” (guidance).

  • Practical translation debugging: use /p/POST_ID and audit logs, but don’t filter on response_tokens
    Falco advised checking the failing post via /p/120 and inspecting ai_api_audit_logs (debug approach). When RBoy didn’t see matching audit rows (query + mismatch), Falco recommended removing the response_tokens clause from the SQL filter (fix). The thread also clarified the /p/ vs /t/ difference during investigation (follow-up).

  • Token-limit confusion: 413 errors are request size, not “max output tokens”
    RBoy reported seemingly random token-limit overruns despite lowering output token caps (initial report). Falco emphasized that 413 indicates the request is too large (not the requested response), and suggested focusing on the LLM “context window” configuration while also noting 8k is unusually small by modern standards (clarification). RBoy replied with their configured context window and the provider’s limit, questioning why Discourse would exceed configured bounds (details).

  • Rate-limit pressure (TPD/TPM) as an upstream contributor to translation instability
    In the same token thread, RBoy noted the translation pipeline initially stalled under daily token rate limits (429), then later failed with 413 request-too-large errors after resuming (sequence of failures). This paired with ongoing translation troubleshooting in What happens to translations when LLM changes? and AI translations errors.

  • Proofreader customization: where to find the built-in examples to adjust quote behavior (ai-helper)
    bksubhuti asked how to find examples so they could tweak their custom proofreading personality to avoid breaking quotes (question). Falco pointed them to the Proofreader agent examples inside the admin UI (admin/plugins/discourse-ai/ai-agents/-22/edit) (direction), and bksubhuti confirmed they found the example set and that it outputs JSON (confirmation).

  • PM Agent follow-up couldn’t post images (fixed upstream)
    Ethsim2 reported they couldn’t post an image when following up with an Agent via PM on 2026.5.0 (bug report). Lilly replied that it was already fixed (and referenced a related incomplete report) (response), and Ethsim2 identified the missing upstream commit they needed (follow-up). (Related: New AI docked composer.)

  • New LLM configuration question: Gemini model id / provider URL confusion
    danhanghai asked for help configuring gemini-3.1-flash-lite via the Google provider on the LLM settings page, sharing their model id and endpoint URL (question). For broader context, that question sits within the long-running reference topic Discourse AI - Large Language Model (LLM) settings page (how-to ai).

Activity

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

123

Overview

This week’s ai discussions on meta.discourse.org clustered around practical reliability fixes—from language/locale detection and translation credit usage to minor UX paper-cuts and self-hosting setup issues.

On the localization side, thomasjsn reported that Norwegian content was being detected as no and then translated into nb_NO, leading to near-duplicate text and wasted credits in Norwegian is identified as no by locale detector agent, content localization supported locales is nb_NO. The thread quickly evolved into a prompt-level workaround (post 5), and then a core/default-agent improvement confirmed by nat (post 7).

Meanwhile, the Discourse AI UI got a small-but-noticeable polish item: RBoy found that changing the default LLM didn’t immediately update agent labels until refresh in Minor UI bug changing default LLM, and awesomerobot followed up with a fix PR (post 2).

Self-hosters also had a useful troubleshooting moment: NotAnonymous hit a Docker + Hugging Face 404 while setting up sentiment, and Falco provided a working workaround in Self-Hosting Sentiment and Emotion for DiscourseAI (see post 15 and confirmation in post 16).

Finally, a couple of smaller check-ins rounded out the week: a follow-up on Gemini “thinking budget” testing limitations in Thinking budget for Gemini Pro, error when using 0 or -1, renewed interest in “me too”/acknowledgement patterns with pointers to the new solved improvements in Option to hide ‘me too’ replies, and a new report in Chinese about AI plugin hyperlinks not responding in 社区官方的ai插件中的超链接未能正常跳转，点击后无反应.

Interesting Topics

Activity

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

124

Overview

This week’s AI-related activity on meta.discourse.org (2026-05-18 → 2026-05-25) was mostly about making AI bot conversations smoother and easier to manage, plus one ongoing operational concern around localization in AI-assisted email workflows.

On the product side, Discourse shipped two small-but-impactful UX improvements for AI chat: the ability to star AI conversations so important chats stay pinned at the top (Star common AI conversations, also referenced in Star common AI conversations), and a docked composer that keeps the input box persistently available in AI bot topics to reduce “reply friction” (Introducing a docked composer for AI bot conversations, see also Introducing a docked composer for AI bot conversations).

Meanwhile, an older #Feature thread resurfaced with a request for updates: translation support is still a pain point—especially when moderation pipeline remarks and other custom text aren’t translated for users receiving emails in their configured language (Use translated posts when emailing users with their user language set, again in Use translated posts when emailing users with their user language set).

Interesting Topics

Activity

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

125

Overview

AI-related discussion on meta.discourse.org this week (2026-05-25 → 2026-06-01) clustered around three themes:

  1. AI integrations & automation — There’s growing interest in event-driven AI features, like triggering external automation when AI Artifact data changes. The request for artifact update webhooks in Add webhook / event support for Discourse AI artifact key-value updates (or allow admins to disable sandboxing) was met with encouragement to revisit after Discourse’s upcoming Automation direction (“Workflows”) (reply), plus agreement that scoped server-side webhooks would be valuable (follow-up).

  2. AI UX and content fidelity — Multiple threads highlighted places where AI-adjacent UX still needs polish: the docked AI bot composer’s “Enter to send” behavior (discussion, more), translation output corrupting Markdown quote syntax (bug report), and ongoing desire for full Markdown parity in bot/chat contexts (request continuation).

  3. Localization & multi-model support — Several posts focused on language and model flexibility: request to localize AI prompts to Chinese (request, reply), progress on using translations in user-facing messaging (including a fix merged for translated “flag reasons” appearing correctly) in Use translated posts when emailing users with their user language set, and expanding sentiment/emotion analysis to more LLM providers like Gemini (question, update).

Interesting Topics

Activity

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

126

Overview

This week’s AI-related discussions on meta.discourse.org clustered around three practical themes:

  1. AI-assisted composition and moderation UX: admins want more granular control over what the AI Helper generates (titles vs tags vs categories) and how it behaves in constrained category setups. This surfaced both a feature request for per-button toggles in the composer (Feature request for ai title generator: toggle title/tag/category) and a bug where tag suggestions didn’t respect category tag restrictions (AI Helper can suggest tags not allowed in category).

  2. Clarifying “AI” vs “not actually LLM” functionality: a key point was reiterated that tag/category suggestions are embedding-based rather than LLM-driven, which affects how much “prompting” can steer outcomes (How to customize AI tag and category suggestions).

  3. AI + localization + tooling edge cases: there were reports of language mis-detection and translation freshness confusion in localized AI features (Weird: reply edit not displaying, and indicated as written in French when it was in English), plus a developer-facing issue where the AI Tools Test Runner appears to negotiate SSL against a plain http:// internal endpoint (Der AI Tools Test Runner macht bei http-URLs intern SSL).

Interesting Topics

  • AI localization mis-detects language and hides edits behind an outdated translation (#site-feedback, ai, dynaloc, content-localization)
    stephtara reported that a post written in English was flagged as French and that a subsequent edit “didn’t show” in the rendered view (report). Moin explained the mis-detection can be triggered by a single word like “French”, and noted the “missing edit” was likely due to viewing an out-of-date translation rather than the source post (diagnosis + workaround). The thread ended with the original poster confirming the explanation and manually correcting the detected language (confirmation), plus a memorable reminder about the limits of language detection:

    “Artificial intelligence isn’t intelligence” (quote).

  • Bug: AI Helper suggests tags that the category does not allow (bug, ai)
    thgl found that the AI Helper could recommend restricted tags and even allow selecting them, only to block submission later—while manual tag entry correctly enforced restrictions (bug report). zogstrip acknowledged the issue quickly (ack) and followed up with a fix already prepared via a PR (fix status). The reporter confirmed the turnaround was fast (thanks).

  • Feature request: toggles to enable AI tagging/categorization but disable AI title generation (#feature, ai)
    Frully asked for separate toggles for title, tag, and category suggestion buttons—wanting to keep AI help for classification while requiring users to craft their own titles (request + rationale). NateDhaliwal suggested a pragmatic interim approach: hide only the “suggest titles” button via CSS (CSS workaround), which the requester accepted as workable (follow-up).

  • How to customize AI tag/category suggestions: clarification that it’s embeddings-based, not LLM-prompted (Support, ai, ai-helper, Solved)
    Frully wanted to “teach” the helper how their community organizes content (e.g., consistent #meetings patterns and required tags) and noted the system prompt examples focus on titles, not tags/categories (question). Falco clarified why: tag/category suggestion doesn’t use LLM prompts at all, relying on embeddings of the draft against existing topics instead (answer / solution).

  • AI Tools Test Runner: http.get() appears to attempt SSL for internal http:// endpoints (Support, rest-api, ai)
    Tobias1 shared a reproducible script showing http.get("http://stable-diffusion:7860/") resolving correctly to an internal IP while failing with an SSL handshake error—suggesting the runner or its HTTP client layer is attempting TLS anyway (details + error output).

Activity

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