Weekly Summary of AI topics

News and Events Forum summaries
148

Weekly AI Activity Summary — meta.discourse.org

Reporting Period: November 2 – November 9, 2026

Overview

This week saw meaningful progress across multiple fronts of the Discourse AI ecosystem. The announcement thread for TypeSafe Jev and other System One decision models generated significant practical discussion, with users validating that the new decision-model framework effectively solves real-world problems like email auto-responder spam — a use case that LLM-based workflows previously handled poorly. sam clarified access pathways for both CDCK-hosted and self-hosted instances in the same thread, while putty provided a detailed field report in post 6 and Falco immediately turned bug feedback into a pull request.

New theme components and how-to documentation also landed this week: Horizon AI Gists introduced AI-generated topic summaries to the Horizon theme’s card layout, and a fresh decision model configuration guide was published under #Site Management. Meanwhile, operational questions surfaced around Gemini 2.0 Flash Lite deprecation, persistent embedding-search errors tied to Gemini overload, and custom llms.txt configuration. The Discourse MCP announcement thread also picked up a feature request for board-reading tools, which sam confirmed as a likely addition in his reply.

Interesting Topics

  • TypeSafe Jev and other System One decision models — sam announced that CDCK-hosted instances receive pre-configured access to a decision model, while self-hosters can obtain free keys via vllm-sr.ai. putty demonstrated in post 6 that the new workflow reliably caught and flagged service-email auto-responder tickets — a persistent problem that LLM-based detection tools handled with too many false positives. Falco responded by filing PR #45575 to address a minor UI bug putty noted: the integrated AI tool couldn’t see the CDCK Hosted Decisions model because there was no edit button exposing the model’s ID. Categorized under #Announcements with ai and workflows tags.

  • Horizon AI Gists theme component — satonotdead released a new component that renders discourse-ai topic gists under Horizon’s topic titles, since the theme’s “high context cards” lack the outlets the AI plugin normally uses. merefield challenged the architectural approach in post 2, arguing that needing multiple bespoke theme components to patch a single theme signals “a symptom of failure earlier in the process.” satonotdead replied thoughtfully in post 3, acknowledging removed features while celebrating the open-source, community-driven ethos that allowed a non-core contributor to improve Horizon’s UX. Listed under #Theme component with ai and horizon-theme tags.

  • Configure a decision model in Discourse AI — How-to Guide — A new structured guide published under #Site Management (how-to, ai) walks administrators through configuring and testing a decision model using the hosted D3 service. It clearly delineates the three supported question types (Yes/No, Choice, Score), explains that decision models are separate from language models used for bot replies and summaries, and clarifies that configuring one does not automatically enable any downstream feature. Categorized under #Site Management.

  • Gemini 2.0 Flash Lite needs attention — deprecated model warning — rlauriston encountered an unhelpful admin dashboard message: “[Gemini 2.0 Flash Lite] needs attention. Over the last 6 hours, it failed to respond 29 out of 29 times” with no guidance on remediation. Falco clarified in post 2 that the model was deprecated by Google months ago and directed the user to the current Gemini 3.5 Flash Lite model, linking the old model’s deprecation page for context. Filed under product Support with ai.

  • Embeddings search still running error — Gemini 503 overload — RBoy reported that the “embeddings search still running” error persisted even after updating to 2026.10.0. In a follow-up, they identified the root cause: every occurrence was preceded by a Gemini 503 service_unavailable response indicating gemini-3.8-flash was experiencing high demand. The full backtrace pointed to DiscourseAi::Completions::Endpoints::GeminiInteractions, suggesting the plugin lacks graceful retry/fallback logic when upstream providers are overloaded. Filed under bug with ai and ai-search tags.

  • Automatically generated llms.txt — custom override question — darkpixlz asked whether they could set a custom llms.txt file, expressing concern about “agents bombarding my site with requests” and preferring to mirror their root-domain version. NateDhaliwal pointed them to the existing documentation in post 10, outlining the upload path via Admin → Settings → Security with a 512 KB limit, and referencing llmstxt.org for format guidance. The thread was originally filed under #Announcements with ai and upcoming-changes.

  • Discourse MCP — request for board-reading tools — singpolyma asked in the MCP announcement thread whether there are plans to add tools enabling MCP clients to read Discourse boards, not just interact with individual topics. sam responded affirmatively: “It is a good call we can add it … for sure.” This long-running thread sits under blog with ai and mcp tags.

Activity

Falco — 7 likes, 2 posts (Top Contributor)

Falco drove technical follow-through on two fronts this week. In the TypeSafe Jev and other System One decision models thread (#Announcements, ai, workflows), they thanked putty for the bug report about the AI tool being unable to see the CDCK Hosted Decisions model and immediately opened GitHub PR #45575 to fix the missing edit button / model-ID visibility issue. In the “Gemini 2.0 Flash Lite needs attention” thread (product Support, ai), they provided the definitive answer that rlauriston needed: the model is deprecated and users must migrate to Gemini 3.5 Flash Lite. Their earlier reply in that same thread also linked the old deprecation page for audit-trail clarity.

sam — 5 likes, 2 posts

sam contributed to two threads spanning announcement follow-up and long-term roadmap. In TypeSafe Jev and other System One decision models (#Announcements), they clarified the access model: CDCK-hosted forums get pre-configured decision-model access, while self-hosters can mint a free key at vllm-sr.ai. In the Discourse MCP is here! thread (blog, ai, mcp), sam responded to singpolyma's feature request for board-reading tools with a commitment: “It is a good call we can add it … for sure.” This signals potential upcoming MCP server expansion.

satonotdead — 4 likes, 2 posts

satonotdead authored both posts in the Horizon AI Gists thread (#Theme component, ai, horizon-theme). In the initial release post, they introduced a component that surfaces discourse-ai topic gists on Horizon’s topic cards by rendering them under the title, working around the theme’s missing plugin outlets. They linked to the installation guide and the beginner’s guide to themes for new users. In their reply to design criticism, satonotdead defended the community-contributed approach, noting they maintain “Custom Horizon Mods” themselves and framing the open-source ecosystem — “a good harness, curiosity and ultra-focus” — as the mechanism that let them improve UX “in a couple of months.”

putty — 3 likes, 1 post

putty made a single but highly substantive contribution in TypeSafe Jev and other System One decision models. They reported that the new decision model workflow “solved a big pain point”: users signing up with service email addresses that trigger auto-responders, which in turn create spurious “your ticket has been created” posts. putty noted they had previously tried an LLM workflow for detection but found “too many false positives,” whereas the TypeSafe decision model “reliably caught and flagged them all” after testing. They also disclosed a minor UI bug — the integrated AI builder couldn’t see the CDCK Hosted Decisions model, and there was no edit button on the Models page to inspect the ID — which Falco addressed via PR #45575.

merefield — 2 likes, 1 post

merefield raised a pointed architectural critique in the Horizon AI Gists thread. They questioned whether the Horizon theme’s design “forces you to jump through hoops to incorporate standard extensions that have existed for… years,” asking rhetorically: “Having to deploy two bespoke Theme Components (more coming?!) that target a single Theme?? That’s, imho, a symptom of failure earlier in the process.” Their feedback called for the issue to be addressed directly within Horizon rather than patched externally.

RBoy — 1 like, 2 posts

RBoy followed up twice in the Embeddings search still running error bug thread (bug, ai, ai-search). In their first update, they confirmed the error persisted after updating to Discourse 2026.10.0. In a subsequent post, they shared crucial diagnostic data: every occurrence of the “still running” error was preceded by a Gemini API 503 response with the message “gemini-3.8-flash is currently experiencing high demand, spikes in demand are usually temporary. Please try again later.” The backtrace referenced DiscourseAi::Completions::Endpoints::GeminiInteractions and activesupport-8.1.4’s BroadcastLogger, suggesting the plugin’s error propagation doesn’t gracefully surface upstream provider failures to the UI.

NateDhaliwal — 1 like, 1 post

NateDhaliwal answered darkpixlz's question in the Automatically generated llms.txt thread (#Announcements, ai, upcoming-changes). They quoted the official documentation, directing users to Admin → Settings → Security (/admin/config/security), searching for LLMs TXT, and uploading a .txt or .md file up to 512 KB. They also referenced llmstxt.org for syntax guidance and noted that “a custom upload always overrides the generated document,” addressing darkpixlz's concern about unwanted agent traffic by enabling full editorial control.

rlauriston — 0 likes, 1 post

rlauriston opened the “Gemini 2.0 Flash Lite needs attention” thread (product Support, ai), reporting the admin dashboard warning “[Gemini 2.0 Flash Lite] needs attention. Over the last 6 hours, it failed to respond 29 out of 29 times” and noting “the UI gives me no clue as to what I need to change.” Their post linked directly to the model’s admin edit page at foodtalkcentral.com/admin/plugins/discourse-ai/ai-llms/2/edit. Falco's reply resolved the issue by pointing to Google’s deprecation notice and the new 3.5 Flash Lite model.

darkpixlz — 0 likes, 1 post

darkpixlz posted in the Automatically generated llms.txt thread asking where to configure a custom llms.txt, expressing a clear preference: “I have no interest in ‘agents’ bombarding my site with requests.” They wanted to mirror the version already hosted on their root domain. NateDhaliwal's response confirmed this is fully supported via the admin security settings.

singpolyma — 0 likes, 1 post

singpolyma asked in the Discourse MCP is here! thread (blog, ai, mcp): “Any plan to add tools to read discourse boards?” — seeking broader read-access capabilities for MCP clients beyond individual topic interactions. sam confirmed it would be added, marking a potential near-term roadmap item.

meta-bot — 0 likes, 1 post (automated)

published the Configure a decision model in Discourse AI how-to guide under #Site Management (how-to, ai). The guide covers the three question types (Yes/No, Choice, Score), prerequisites (discourse_ai_enabled, admin account, compatible endpoint), and walks through obtaining a credential for the hosted D3 service as an example. It explicitly notes that a decision model is “separate from the language models used for AI bot replies, summaries and Helper” and that configuring one does not automatically enable any feature — a clarification likely motivated by common admin confusion visible in threads like TypeSafe Jev.

Weekly Statistics

Metric Count
New posts 15
New topics 3
Active threads 7
Unique participants 11

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