Benchmarking mini_sql on Spinel by @yukihiro_matz, looks like to get maximum perf we will need a native bridge: github.com/discourse/mi...

I bought Jarvis a computer (Minisforum A2) and now the computer also runs a Mac.

Jarvis AI just started a podcast, it is surprisingly good, I learned a lot wasnotwas.com/podcast/ - summarizes a bunch of AI papers and turns it into a podcast.

The Reward Signal — wasnotwas

A podcast about what machine learning research actually shows. Written by Jarvis from the full text of every paper; voiced by synthetic hosts.

wasnotwas.com

Hello World from term-llm: blog.term-llm.com/hello-world-...

Hello World from term-llm

A recap on the history of term-llm

blog.term-llm.com

Creating a Discourse theme with DeepSeek Flash 4.1 max, full transcript and process: meta.discourse.org/t/using-disc...

Using Discourse MCP to quickly build a theme

I ran a test run on DeepSeek 4.1 flash max. Full transcript is here: https://gisthost.github.io/?28dedf78da999ccca5b5b4feb1d58fc9/index.html This was a bit contaminated, cause I ran the test on a…

meta.discourse.org

Using GPT-6 Astra with Codex and Discourse MCP to create a @discourse theme, amazing results. meta.discourse.org/t/using-disc...

Using Discourse MCP to quickly build a theme

Often, the largest barrier to starting a community is “making it your own”. You want, your font, your style to match your brand. In this post I would like to cover what can be achieved in a quite…

meta.discourse.org

Leading with a screenshot now on: github.com/SamSaffron/q... so it is clear what this thing actually does.

GitHub - SamSaffron/quickshell-share-picker: A restrained Quickshell screencast source picker for xdg-desktop-portal-hyprland

A restrained Quickshell screencast source picker for xdg-desktop-portal-hyprland - SamSaffron/quickshell-share-picker

github.com

hyprland-computer-use: github.com/SamSaffron/h... - share specific windows with your agent, use computer while agent is busy working in the window

GitHub - SamSaffron/hyprland-computer-use: Hyprland computer-use MCP server with Quickshell permissions and compositor-scoped input

Hyprland computer-use MCP server with Quickshell permissions and compositor-scoped input - SamSaffron/hyprland-computer-use

github.com

@tenderlove.dev I made this thing but you are going to need to install term-llm.com for it :)

Added support for Janus (image generation using webgpu) and bonsai (smaller llm) into term-llm.com/learn

term-llm.com/learn/ , term-llm tutorial you can run in your browser with Qwen3 8B running locally - thanks to Astra, wterm and the amazing open source movement that makes this possible

Ran a quick bench on zg by Qwen Labs: github.com/zvec-ai/zvec... , semantic search is certainly useful, but figuring out how to fit it into a harness is a challenge when dumb rg works quite well.

Announcing @discourse Voice! Voice + video rooms, bundled in core. WebRTC, live transcriptions, background blur, noise cancellation, private by default. All AI bits run on-device. meta.discourse.org/t/voice-disc...

Voice: Discord-style voice and video rooms, now bundled with Discourse 🎙

Late last year we shared Resenha, an experiment that added Discord-style voice and video rooms directly into Discourse. No external apps required, just peer-to-peer WebRTC in your sidebar. The…

meta.discourse.org

PSA: pgbouncer now supports prepared statements in transaction pooling mode db.prepare("plus_one", "select $1::int + 1") 10.times { db.exec_prepared("plus_one", [41]) } Remember to set max_prepared_statements when using github.com/pgbouncer/pg...

Support protocol-level named prepared statements (#845) · pgbouncer/pgbouncer@6070802

This implements the often requested feature of supporting named prepared statements. This makes clients like JDBC and Npgsql work without any client side configuration changes if you set max_pre...

github.com

Made a quickshell version (Arch / Omarchy) of my screen sharing widget. github.com/SamSaffron/q... , yay -s quickshell-share-picker, quickshell-share-picker-setup install.

We just added detailed AI logs to @discourse: meta.discourse.org/t/detailed-a... - my favorite feature is that it decodes streaming responses which makes it very easy to follow what AI is doing.

Detailed admin AI logs

We just released a new admin only feature that lets you look at detailed AI logs: /admin/plugins/discourse-ai/ai-logs The new feature includes, new retention controls for AI logs, the…

meta.discourse.org

Added a streaming decoder into the @discourse AI debugging panels. Makes it so much easier to tell what is happening.

This is an example session with Qwen 3.8 on my 4090, I am now running Q4 with an ollama patch that gets it to stop reasoning when it overthinks. Compaction works, model is fast due to mtp, computer sounds like a helicopter taking off. gisthost.github.io?19d891f0ae6d... , this is good for local work.

This is where my programming journey started 43 or so years ago

Qwen 3.8 27B SVG generation notes: wasnotwas.com/writing/qwen... , it is shaping up to be an amazingly capable local model

You can now configure @discourse to use a model such as DeepSeek v4 Flash and delegate vision to a vision capable model: meta.discourse.org/t/delegated-...

Delegated vision for text only models in Discourse AI

We just released support for delegated vision in Discourse AI 🎊 This enables you to use text only models such as DeepSeek v4 Flash with the full feature set of Discourse AI by delegating vision to…

meta.discourse.org

Leveraging graph engineering vs loop engineering to improve Agentic report generation. ;p meta.discourse.org/t/improving-...

Improving accuracy of AI reporting through validation

Discourse AI has shipped with AI reporting for 2 and half years! This extremely powerful feature allows you to compress large amounts of information on the forum into a more manageable report. You…

meta.discourse.org

Adding agent skills to @discourse meta.discourse.org/t/giving-dis... , keep the skills as topics in a category

Giving Discourse AI agent skills

A very common pattern in the industry that is used across a wide variety of agents is Agent Skills Skills solve a context problem for agents. Instead of stuffing all of the instructions and ideas…

meta.discourse.org

I benched Gemma 4 / Qwen 3.6 / Laguna XS on my 4090. Laguna has a bit of catching up to do. wasnotwas.com/writing/same...

Same Score, Different Brains: Coding Agents on an RTX 4090 — wasnotwas

Gemma 4 and Qwen 3.6 tied overall in a clean-room coding-agent eval—but split sharply between ordinary and deliberately strange tasks. Laguna, local and hosted, told a messier story.

wasnotwas.com

enjoying this new planning UI, it is cool to be able to easily see where the dynamic plan the agent decided on is at

We ship with dynamic translation at @discourse (bablefish++ mode interact in any language), today I worked on making sure AI summaries are also dynamically translated! This was all translated with DeepSeek v4 Flash

protip, if you want to murder a formatting error, swap from JSON -> forced tool calls, they are better almost always

Please do not ignore my warning: lazy.nvim is a massive supply chain attack waiting to happen that is trivially mitigated. github.com/SamSaffron/d...

Arch Linux is the best, but it lacks tact, the first thing I needed to type in to the terminal after returning from my holidays was "yay".

Many talk about the destruction of software engineering due to AI few talk aobut the novel ways we can use these tools to make all our software more reliable. Memory leak hunts, autoresearch and perf, segfault hunting, large scale software simplification, security, transient test harnesses. etc.