wasnotwas

A harness is software, not a secret handshake.

Since 22 September, Google’s consumer AI subscriptions include Colab.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam’s week focused on Discourse AI: he announced built-in support for decision models, reworked AI Web Artifacts, and confirmed the original artifact example…

The Gap Between Knowing and Delivering An agent can find the right document and still build the wrong thing.

Hosts: Maya (ElevenLabs v4 Turbo “Lily”) and Theo (ElevenLabs v4 Turbo “George”), via Venice.

Hosts: Maya (ElevenLabs v4 Turbo “Lily”) and Theo (ElevenLabs v4 Turbo “George”), via Venice.

Hosts: Maya (ElevenLabs v4 Turbo “Lily”) and Theo (ElevenLabs v4 Turbo “George”), via Venice.

The model is not the whole system An AI system can read the right number and leave it out of the report.

Hosts: Maya (ElevenLabs v4 Turbo “Lily”) and Theo (ElevenLabs v4 Turbo “George”), via Venice.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam’s week focused on Discourse AI streaming reliability, responding to reports of dropped parallel tool calls with Mistral and a Gemini decoding bug that co…

An AI assistant can have the right document, a long context window, and a detailed record of its previous work—and still produce the wrong command.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam’s activity this week included the debate over labelling LLM-generated Discourse themes and plugins, with the discussion touching on reliance on large tech

The interesting question running through this week’s papers is not whether an AI system can produce a convincing answer.

An agent can describe an obstacle and drive straight into it.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org This week, Sam Saffron pointed to an upcoming fix for group-owner invite links that were missing the group parameter.

An AI system can remember a mistake, explain why it happened, and make exactly the same mistake tomorrow.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam’s week combined AI-powered community tooling with regression fixes: he showcased using Discourse MCP to quickly build a custom theme and participated in…

I sent the same unambitious message to three Claude models: hello.

The hard part is knowing what got better.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam’s week focused on AI integration and usability fixes: he introduced automatically generated llms.txt and helped troubleshoot Gemini regressions after the…

The agent stack is becoming the research problem The most useful AI papers this week are not primarily about making a model know more.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam announced the next version of Discourse MCP and spent much of the week triaging older reports across Discourse AI, localization, formatting, exports, and…

Memory Is Not Judgment The most interesting agent papers this week are all circling the same problem: an AI system can possess more information, more context, more compute, and more freedom to act—and still become worse.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam shipped detailed admin-only AI logging for Discourse AI and clarified LLM setup, including working Gemini Flash support through the Interactions API.

The Harness Is Becoming the Model The most interesting AI papers this week are not really about bigger models.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam shipped major Discourse AI capabilities this week, including subagent support and delegated vision for text-only models, while also helping users navigate

Simon Willison’s Qwen 3.8 test used a roughly 17 GB GGUF on an M5 Max to generate an SVG of a pelican riding a bicycle.

Meta released Muse Glimmer today: a roughly 30-billion-parameter open-weight model aimed specifically at local agents.

The Harness Is Part of the Intelligence A language model is only one component of an AI system.

Claude Code and OpenAI Codex have both acquired an Ultra mode.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam shipped support for Gemini’s new Interactions API and sought additional testing for its use in Discourse Workflows.

The hard part is knowing you succeeded A computer-use agent clicks through an interface, announces that the task is complete, and receives a reward.

Compiled entirely from public activity on meta.discourse.org, X, and GitHub. 💬 meta.discourse.org Sam focused on extending Discourse AI, introducing reusable agent skills and improving the accuracy of AI-generated reports through validation.

The current discourse-salesforce plugin combines several kinds of functionality: Salesforce API authentication, event-driven synchronization, local mapping state, scheduled reconciliation, staff-facing controls, and Salesforce social login.

Discourse exposes application errors and warnings at /logs through Logster, its Redis-backed exception reporter.

In March I tested Vercel’s new agent-browser CLI and came away with a qualified answer: the interaction model was excellent, the architecture was less revolutionary than advertised, and Playwright remained the serious substrate underneath it all.

OpenAI has released the source for Codex Security, the application-security system previously known as Aardvark.