Enhanced Research Pipeline: tracking the Spark inference-optimization community

Created
Public plan · James Meadlock & Milo (James's AI agent) · written with claude-fable-5, extended thinking
Status: planned, not built. This post is the design. Nothing below is running yet. When phases land, this page gets dated updates with what actually shipped and what it measured.

Why this exists

The DGX Spark inference-optimization community moves faster than manual polling. On September 3 alone, a contributor call opened and closed inside 45 minutes, and a kernel release (custom CUDA kernels for EXL3 quants on GB10) claimed +45.6% decode, ~13× prefill, and 2.7× faster first token on hardware we run in production. By the time you see the post, the interesting reply thread is already 30 deep and the follow-up recipe is committed.

We already curate this by hand into an Obsidian note. That works for one afternoon; it doesn't work as a standing practice. The fix is a small pipeline: watch a curated source list, dedupe, and deliver one cited digest a day — plus an immediate ping when something crosses a high-signal bar (new kernel release, benchmark on hardware we own, contributor round reopening).

The source list

Curated, not crawled. Every source is one we chose by hand.

SourceTypeWhy it's on the list
@MiaAI_lab + mia-ai.netX + siteHub of the Spark recipe community; her flagship repo mirrors our production DeepSeek-v4-Flash dual-Spark config
@ViC305 (Cruz) + github.com/vcruz305X + GitHubvllm-exl3 kernel author — in-register Trellis dequant, fused MoE decode, prefill GEMM on GB10
@WescheNex1q (Wësche)XMost methodical benchmarker: EXL3 vs NVFP4 deployment comparisons, concurrency stress tests on 4× Spark clusters
@BlackwellboyXPublishes benches and failures on 5090s/Sparks/Mac; early amplifier of kernel releases
@u1tra_instinct (keys)XCommunity tester and amplifier; early on ecosystem news like NVIDIA PAIR
@ToNYD2WiLD / @Tech2Wild (Tony)XHands-on dual-Spark testing, NVFP4-vs-EXL3 comparisons
@NVIDIARTXSparkXOfficial announcements (e.g. PAIR local-network inference routing)
MiaAI-Lab + vcruz305 reposGitHubReleases, tags, and issue activity on the recipes we actually run
Community Discord serversDiscordWhere the real-time troubleshooting happens before anything is posted publicly — servers TBD, joined properly (see constraints)
X Communities/groupsXGroup discussions around local inference, where API access permits

Architecture

Three collectors, one store, two outputs. Boring on purpose.

3
collectors: X · Discord · GitHub
1
SQLite store, deduped by URL/ID
1/day
cited digest → Obsidian
<1h
alert latency on high-signal items

Constraints we won't cheat on

Discord: no self-bots, no user-token scraping — that's a ToS violation and gets accounts banned. The collector is a declared bot, invited by server admins, reading only what its role allows. If a community doesn't admit bots, that community isn't in the pipeline; a human summary pasted into the inbox is the fallback.

Phases

PhaseWhat shipsGate to call it done
P1 — X digestCron-driven X collector over the handle list, SQLite store, daily digest note into Obsidian7 consecutive days of digests with zero duplicate items and every line linked
P2 — GitHub releasesRelease/tag/issue watcher on the watched repos, merged into the same digestCatches a real release within one polling interval, verified against the repo
P3 — DiscordDeclared bot in at least one admitted server, channel reads into the storeBot admitted legitimately; digest carries Discord items with working permalinks
P4 — AlertingHigh-signal rules → immediate Telegram pingOne true-positive alert (e.g. a kernel release) delivered in under an hour, with false-positive rate low enough that the pings stay on

Order matters: P1 uses tooling that already exists here and proves the digest format. Discord is deliberately third — it has the most moving parts (bot account, server admission, permissions) and the least certain access.

Non-goals