Ryan Spice · technical blog
A SvelteKit-first blog project staged inside the AI Wiki, with repair logs, debugging notes, research comparisons, and a lightweight dev log that stays grounded in the actual workflow.
Quick links
A compact map of the public archive, feed, and build surface.
Latest articles
All currently published public technical notes, newest first.
technical analysis
Fable 5, GPT-5.6, Kimi K3 and Qwen show why the durable AI advantage is shifting from a single best model to the fleet, harness and execution layer around many models.
build log
I rebuilt my Fugu/Fusion coding harness around the models that are actually available: GPT-5.6 Sol conducts, Fugu executes through bounded specialist lanes, and a three-model remote Fusion panel reviews important work before the conductor decides.
daily dev log
Today's Pulse Pro focus is Lab routes: prove the composition-spike path, crosscheck gates, and keep parked engine-eval rows from becoming false P0 work.
field note
OpenAI has moved GPT-5.6 from limited preview to general availability across ChatGPT, Codex, and the API. The release raises the capability ceiling for coding agents while making version hygiene, access controls, and network boundaries more important.
daily dev log
A repo-truth-first Pulse Pro briefing for the rotated Composition-spike renderer focus: prove the PixiJS/Corsair composition spine and owner-GPU frame budget before new renderer stages.
case study
A case study on treating QBO Mail Dashboard as an internal operations workflow that can sell consulting/productization work without pretending to be a public SaaS product.
case study
A practical case study on narrowing PixelBoats from a broad game prototype into a smaller Android-first animated visual product lane with ASO metadata, proof assets, and conservative sales boundaries.
case study
A self-case study for Keyword Astro: using the ASO workflow, screenshot proof, source-labelled metadata, and RC/final release split to make the product more saleable without overclaiming rank or popularity metrics.
product strategy
A practical product strategy for turning the old Astro email/calendar idea into a 2026 Windows-and-Android-first command center, with local-first privacy, staged launch discipline, and source-backed platform constraints.
daily dev log
A local Pulse-style briefing that turns the morning fusion run and no-edit scale/collision inventory into a concrete P0 intake queue.
short news update
Anthropic says Fable 5 returned on July 1, 2026; the operational story is access, safeguards, fresh limit resets, a temporary weekly allowance, and developer anxiety over whether the relaunch feels like the same model or a permissioned version.
quick technical note
A practical Windows PowerShell setup for running Claude Code against DeepSeek's Anthropic-compatible endpoint, with scoped environment variables, a safer launcher script, and a low-risk smoke test.
daily dev log
A local Pulse-style captain's log for the P0 boat footprint work, with the June 26 and July 2 follow-up pulses folded in as update notes instead of duplicate articles.
build log
The archived June 2026 version of my local Fusion/Fugu harness, preserved as the historical setup that preceded the GPT-5.6 Sol conductor and three-model remote Fusion panel.
field note
A field note on wiring GLM-5.2 into Hermes through Cloudflare Workers AI, hitting quota pressure quickly, and landing on a safer lead/delegate architecture for expensive reasoning models.
setup guide
A real setup for driving NVIDIA Nemotron 3 Ultra (550B-A55B) behind a Hermes agent, including the stale-stream timeout fix, thinking-mode tool-use settings, sampling defaults, and the practical 256K vs 1M context-window tradeoff.
model research analysis
A practical look at GLM-5.2, why its 1M-context claim matters, how IndexShare changes sparse-attention search, and what builders should do differently when exposing sources to coding agents.
technical note
A short note on bringing Canopy Digital into the same publishing fold as the personal technical blog without duplicating the content system.
agent architecture
A practical note on what Claude Fable 5's sudden departure and policy fight changes for serious coding and source-grounding workflows.
technical recovery guide
A backup-first Windows recovery guide for DeepSeek GUI conversations that appear to disappear after an app update changes the local data directory.
ai news analysis
OpenJarvis is trending because it reframes the personal AI assistant as a local-first, measurable, editable agent stack instead of another cloud chat wrapper. Here is what matters, what is hype, and whether it can run on a normal developer PC.
agent architecture
Part 1 of a practical series on mixing DeepSeek V4 Pro, DeepSeek V4 Flash, and local Gemma4-style agents without turning a coding workflow into an expensive committee.
research analysis
A practical comparison of ChatGPT Deep Research and DeepSeek-style reasoning APIs, focused on workflow, retrieval, transparency, and what builders should actually take away.
product strategy
A Windows-focused adaptation of the mobile app factory playbook for the Microsoft Store.
technical decision log
A detailed comparison of Phaser 4.1 and PixiJS 8.18 for a browser-based multiplayer roguelike sailing game with custom water rendering. Why the default choice (Phaser) turned out to be wrong for this project, and how PixiJS's lower-level rendering primitives map more directly to what I'm actually building.
technical blog post
A companion debugging article about GIMP 3 Python plug-in failures on Windows, the misleading libgraphite2/_Unwind_Resume symptom, Pango/GI failures, PATH/DLL pollution, and the Windhawk hook layer.
technical blog post
A real-world repair log and workflow writeup covering a broken GIMP 3.2 install, Windows DLL/runtime conflicts, Windhawk suspicion, PhotoGIMP, G'MIC-Qt, a Fluent-ish theme, and a PixelBoats-ready asset kit.