Sep 14 to Sep 20, 2026

Shantanu Goel trained Laya for ~24 hours to play FreeDoom E1M1 from text-only perception. Jev played well zero-shot; the Laya base model did not. Honest negative result.

An experiment that pushes the Qwen3.8 models onto the Neural Engine of Apple Silicon rather than the GPU. A long-shot attempt to cut power use for local inference.

Mario Never Dies: a games & real time build posted by appcypher (@theappcypher) that puts Jev, TypeSafe AI's System One decision model, to work, with a demo video on the post.

Skill that audits and cleans up the skill list and instructions Codex hands to Astra, installable through npx skills.

A stability-focused recipe for a single DGX Spark with a KDA decode kernel and EAGLE speculation, fixing KV-loop failures others hit.

Dockerized FastAPI service bundling all three checkpoints behind a language router, with timing-safe API keys, OpenAPI docs, triage and moderation presets and bulk inference.

Unmodified Laya weights, real NES emulation via Stable Retro, text observations from RAM, model-driven jumping. 27.5 s of game time at 16.79 ms median inference.

Experimental first-person co-op Unity game for one to four players about a cleaning crew, fully generated by Astra.

Exposes DSH as an MCP server so Claude Code delegates the reading and typing to cheap DeepSeek Flash workers. A cost-saving pattern posted as a Show HN.

Ports ExLlamaV3 kernels to RDNA 3.5 with grouped MoE, int8 mixers and MTP. The author quantizes frontier models down to fit smaller GPUs.

Plugin that lets opencode use Muse Spark via Meta device login and the monthly Muse Code plan, with no API key needed.

One typed question per frame; the other paddle is three lines of arithmetic as a control. Physics in WebAssembly, decisions in native Laya over POST /decide.

Rishi Raj Jain redacts only the personal details in support messages rather than blanking whole columns, using Jev inside Neon Postgres.

Shane Walker's plugin that routes OpenCode through a Muse Code subscription so Meta's coding agent can be used from a familiar interface.

Measures the decision head's attention drop-off on long inputs across 400 documents, then chunks, screens and aggregates, or locates the one passage that matters.

Agamya Acharya made a browser widget for GitHub PRs that checks CI, diff size and reviews, then lets Jev answer 'should I approve this?' in about 200 ms.

Hota911 let Jev play Puyo Puyo, giving it each candidate placement result and asking four questions per move: which phase of play, then which result is best.

Mario Jankovic made an extension that combines YouTube's AI labels with each video's title and description so Jev can hide AI slop from your feed.

Playable browser horror game, Don't Look Away, with two angels and a flashlight, built with fal, Astra in Codex and Three.js.

Kyle McLaren released JevQL, which lets vanilla Postgres filter rows by meaning using Jev, with no extension, plus a CLI, SDKs and MCP support.

The decision model never writes code; each step it picks the next tool, scores progress, estimates risk and says if the goal is reached. Any OpenAI-compatible LLM executes.

Hermes plugin that classifies each task against a ~300-skill index with Laya, injects the top skill's SKILL.md as per-turn context, and fails open. Cache-safe by design.