HOME GITR3KT Morning Tech Brief — No. 54 · Friday, October 23, 2026 Friday · #ff5c8a ← All briefs
Edition No. 54 hero art — robot arm pouring coffee over a stack of GPUs, agent bots climbing a circuit-board ladder over a circuit city at dawn
Friday, October 23, 2026

The Morning Brief — Edition No. 54

Today's tour: your SKILL.md format joins the Linux Foundation, NVIDIA's 64GB mini-box says it can run your Qwen 3.8, AMD admits its next GPU price hike is aimed at NVIDIA's prices, KeePassXC finally fixes Auto-Type on Wayland — and Modern Warfare 4 drops today. Plus the repos everyone is starring.

🧠AI tools

the day the skills format grew up

Agent Skills, Open Responses, OpenEnv, OpenSandbox join the Linux Foundation's Agentic AI Foundation

Four donations in one day, announced Oct 22: Anthropic → Agent Skills (the SKILL.md package format — metadata + instructions + scripts/references, loaded on demand), OpenAI → Open Responses (model-interface interop), Hugging Face → OpenEnv (training/eval environments), Alibaba → OpenSandbox (secure execution). AAIF goes from 3 hosted projects at its Dec 2025 launch to 10. Anthropic's David Soria Parra explicitly pointed at the MCP precedent as the model. This is the format your harness already speaks becoming neutrally-governed shared infrastructure. Caveat from the reporting itself: neutral governance on paper doesn't stop the biggest deployers steering in practice. Watch what the spec does to your skills — not the press release.

Ground Truth — Oct 22

NVIDIA's GTC Berlin local-AI push: llama.cpp spec-decode gains on Qwen 3.8, day-zero LFM2.5

NVIDIA name-checked your exact stack: llama.cpp spec-decode updates claimed to make end-to-end agentic tasks up to 1.4× faster for Qwen 3.8 Flash Next (RTX PRO 6000), and vLLM's FlashInfer prefill kernels push Qwen3.8 27B up to 1.3× on DGX Spark — "alongside the llama.cpp and vLLM communities," per the blog. Also: day-zero support for Liquid AI's LFM2.5-24B-A2B (your pending shootout candidate), new FP4-Local checkpoints for single-machine VRAM, and eight 64GB DGX Sparks clustering to 1TB unified memory. Vendor numbers on NVIDIA hardware — steal them as test ideas, not promises. The llama.cpp spec-decode PRs are the part that might actually land on your box.

SiliconReport — Oct 22 (announcement Oct 21–22)

Krafton drops Raon-1 mini preview: 28.8B/4.3B MoE, 262K context, Apache 2.0, DFlash2 drafter included

A game publisher is now a model vendor: Krafton (PUBG) shipped an MoE — 28.8B total / 4.3B active per token, 18 linear-attention + 7 full-attention layers, 262,144 context, tool calling, Apache 2.0, llama.cpp/vLLM/MLX support. MXFP8 and NVFP4 quants; the NVFP4 build reportedly runs on a single RTX 5090 (32 GB) — squarely in your dual-16GB territory at that precision. Best part: they shipped a DFlash2 draft model with it, claiming up to 2.18× throughput over autoregressive on B300. It's a preview, so specs may wobble — but an Apache-2.0 MoE with a bundled drafter aimed at 32GB-class machines reads like it was specced against your box. Try the GGUF when it lands on the hub.

AI Brief — Oct 21 (primary: Hugging Face)

Recursive's Flint: Apache 2.0 laptop-sized MoE coding model, built (mostly) by AI, 131 tok/s on a MacBook Pro

40B total / 2.5B active per token, Apache 2.0, free on Hugging Face — and it claims ~131 tok/s on a MacBook Pro M4 Pro, roughly double peers like Gemma 4 (58–69 tok/s). Leads models under 50B on SWE-Bench Verified, and on Recursive's own six-test average only Alibaba's Qwen3.8 27B scores higher — arguably the most interesting sentence in the press release, since that 27B is already your champion. Built by Recursive's in-house "Eureka Machine" AI system in under four months; sparser shape + low-bit training shrank it 71% "with almost no loss in quality." Benchmark framing is Recursive's own, but a permissively-licensed 40B coding MoE hunting your exact bracket is worth the shootout once the GGUFs drop.

TechCrunch — Oct 21

Spec-decode speedup claims this week — vendor numbers, not promises

EAGLE-4 (Qwen3.8-27B, 4×B300, SGLang)
3.42× — holds to 512 concurrent (eagle-hq/EAGLE-4)
Krafton Raon-1 + DFlash2 (B300)
2.18× over autoregressive
llama.cpp spec-decode (Qwen3.8 Flash Next, agentic)
1.4× — NVIDIA claim, RTX PRO 6000
vLLM FlashInfer prefill (Qwen3.8 27B, DGX Spark)
1.3× — NVIDIA claim

Four numbers, four different stacks — don't compare them directly. The useful read: spec-decode is the week's shared obsession, and EAGLE-4's code (source-available, not OSI — see repos) is the one you can actually inspect for your own stack.

Does it fit your box?

Params × bits-per-param ÷ 8, plus KV/context and runtime overhead. Your standing rule: base + max KV growth must clear 29.4GB (92% of the 32GB pool) with margin.

0 GB
Bar scaled to 32GB pooled dual-A770 · red line = your 29.4GB (92%) standing rule · a single A770 = 16GB

💽Hardware

the memory shortage is now priced into everything

NVIDIA ships 64GB DGX Spark ($4,999) at GTC Berlin; 8-unit clusters hit 1TB unified memory

The local-inference story of the day, rolling out today: 64GB unified memory, same GB10 Grace Blackwell Superchip, starts at $4,999 — via partners only (Acer, ASUS, Dell, Gigabyte, HP, MSI), no NVIDIA-branded unit this time. NVIDIA says the 64GB runs "local models such as Qwen3.8 27B, Gemma 4 26B-A4B and Nemotron 3.5 Bolt" — your exact model family, name-checked by a vendor. Claimed 1.3× on agentic Qwen3.8 27B via faster vLLM spec-decode + FlashInfer prefill. The price irony: the 64GB costs more than the original 128GB Founders Edition ($3,999); the memory shortage pushed the 128GB to $6,950. Eight 64GB units ≈ $40k for 512GB; eight 128GB = 1TB, clustered over ConnectX-7 with manual config for now. NVIDIA Sync gets 8-unit support plus a "Model Launcher" (downloads model + container, wires to OpenCode) later this month. Not a buy recommendation — but it reframes what a 64GB box is worth, which is useful context for the B70 watch.

SFFGeek — Oct 22 (GTC Berlin)

AMD Radeon partners prep a second GPU price surge — deliberately closing the gap with NVIDIA

The "affordable" brand officially gave up: after January's 5–10% hikes, the next wave is expected Feb–Mar, driven by DRAM constraints and a reported ~10% TSMC supply-price pass-through for Q4 2026. AMD already raised GDDR6 kit prices for partners (Sapphire, ASUS, XFX, Vastarmor) effective Oct 1. The strategy per TechOlam is deliberate alignment with NVIDIA's inflated levels, not just cost coverage — analysts see prices not peaking until late 2027. The RX 9070 XT range already sits €816–930 in Germany. This validates your dual-A770-at-sane-prices posture twice over: you bought VRAM before the tax, and every new card is now taxed by GDDR-per-board. Nothing to do here except not buy GPUs.

TechOlam — Oct 22

NVIDIA denies RTX 5090 discontinuation rumors as the flagship vanishes from shelves

GeForce VP Jeff Fisher in Taipei: production "continues," cards go to distributors — while admitting supply chains are "heavily overloaded." The rumor was that top chips are being redirected to datacenter AI (more profit); Fisher says no. The reality on the shelf: the 5090 has disappeared from NVIDIA's official US store, third-party prices are "absurd," and a Reddit photo of a 5090 box on a store shelf "causes a stir." Schrödinger's graphics card. Context: the 50 series is up ~66% YoY, the 5090 up 132% (TechTimes, Oct 12). Your standing rule — games land on slot-1, never on whatever the market is screaming about — continues to be the cheapest performance decision in the room.

iXBT.games — Oct 22

NVIDIA driver 621.01 + Microsoft Advanced Shader Delivery: precompiled shaders kill first-launch stutter

GeForce Game Ready 621.01 (Oct 21) brings Advanced Shader Delivery support across all RTX GPUs (20/30/40/50 series) + RTX Spark on Win 11 24H2+ — precompiled shader blobs delivered before you play, so no first-launch compilation hitch. First confirmed games: Gears of War: E-Day and Call of Duty: Modern Warfare 4 (out today). Microsoft claims MW4's initial shader prep drops "from several minutes to seconds" (vendor claim, unverified). Also lands on AMD (Adrenalin 26.6.1+) and Qualcomm; Steam and Epic support coming. The footnote that matters to you: Intel's list names Arc B570/B580 + Core Ultra (driver 32.0.101.9150+) — your A770 is Alchemist, not named. The stutter-kill doesn't reach your gaming card yet. On Linux, shader pre-caching is still the game's job, so your stack keeps playing by the old rules.

TechRepublic — Oct 21 (Microsoft announcement)

GPU price inflation — the shortage in one glance

RTX 5090, YoY
+132% (TechTimes, Oct 12)
RTX 5060 Ti 16GB vs $429 MSRP
~+70% US street (~$730)
RTX 50 series, YoY
+66% (TechTimes, Oct 12)
5060 Ti 16GB, Japan, 30 days
+26.4% (~$871, ¥137,800)
AMD Radeon partner hike #2 (coming)
~+5–10%, next wave Feb–Mar

Memory-heavy cards absorb the most of the shortage tax. Your 32GB pooled pool — bought at pre-tax prices — is the cheapest VRAM in the house, and nothing on this chart argues for buying GPUs right now.

🐧Software

releases that actually touch your machine

KeePassXC 2.8.0 lands: Qt 6 port, Wayland Auto-Type, Quick Unlock

First stable of the 2.8 series (Oct 21): the Qt5→Qt6 migration plus the thing Wayland said was impossible — Auto-Type on Wayland via XDG Desktop Portals (the AppImage does it with systemd-run on GNOME/KDE). Also: Quick Unlock via Polkit on Linux (fingerprint/device auth instead of the master password every session), configurable keyboard shortcuts, remote database sync via external tools (WebDAV/SFTP), and SSH-agent key generation (RSA, ECDSA, Ed25519). Fixes five vulnerabilities, one with a CVE (CVE-2026-69150) — and the release notes dryly add that "none of the vulnerabilities originated from AI-generated code, although AI-assisted analysis helped identify them." On Omarchy, this is a genuine quality-of-life update. Update and re-enroll Auto-Type.

KeePassXC project blog — Oct 21

Steam client update fixes Vulkan alt-tab freeze on Linux

Oct 21, stable channel: Linux changelog is two lines — "Fixed Vulkan games hanging after alt-tab" plus a Steam updater UI fix. It sat in beta since Oct 16. One caveat worth honoring: a follow-up report claims the deadlock persisted on the fixed build, so frame this as mostly fixed, not victory. The sleeper in the same update: a certificate-revocation-list refresh after the eMudhra compromised-CA incident (8 compromised intermediate certs — an attacker holding any of those private keys could impersonate any website). Fedora 45 even slipped its release over one of those certs sitting in its trust store. Update the client, test alt-tab on your Vulkan titles, and keep an eye on it.

GamingOnLinux — Oct 21 (official changelog)

DankMaterialShell rebrands to Dux Shell with 2.0 "Heisenberg" release

The Quickshell Wayland desktop shell just got its biggest release and a new name: DankMaterialShell → Dux Shell, parent org → Dux Labs ("Dux" echoes Tux; the dms command and abbreviation survive, so configs don't break). 2.0 "Heisenberg" brings a Material 3 Expressive redesign, a plugin-able Dash dashboard, a lock screen with a visual widget editor, the Dux Lyrical Engine (synced lyrics, word/syllable-level, via LRCLIB/YouTube Music/KuGou), and the standout: Dux Files — a brand-new standalone file manager with tabs, split views, previews, and undo. Supports Hyprland, Niri, Sway and friends. Hermes's #1 queue item is still Strata proving itself with a Miller-columns screenshot; Dux Files just entered the arena as another contender to watch, not to touch yet.

Linuxiac / It's FOSS — Oct 22

🎮Games

launch day, a patch for mapmakers, and one very important bullet

Call of Duty: Modern Warfare 4 launches today — including on Switch 2

Global launch today, Oct 23: $69.99, PS5, Xbox Series X|S, PC, and Switch 2 (Xbox Play Anywhere) — the first CoD to launch day-and-date on a Nintendo console in the Switch 2 era, which is the industry story of the week. Digital pre-orders got campaign early access from Oct 16 and DMZ extraction from Oct 20. Two threads to watch: Microsoft says Advanced Shader Delivery cuts MW4's initial shader prep "from several minutes to seconds" (vendor claim), and the "Call of Duty Tax" pricing debate an Activision exec kicked off Oct 15 keeps simmering. If it lands on your box: the standing rule holds — prove the process lands on 03:00.0, or the test didn't happen.

Fairgate Report / Xbox Wire — Oct 23

CS2's Oct 22 update is a love letter to mapmakers

Every line is tooling: hostage scripting in map scripts, asset precaching, a new GameRules export (scripts control warmup, scores, round endings), a top-level Console export that can set convars without sv_cheats enabled, and script-placed radar elements now working in every game mode (previously Rush only). Competitive matchmaking, weapons, and maps untouched. Second patch in eight days. The Console-export detail is the one to sit with: that's a meaningful change in what shipped maps can do without the cheat flag — seeds for a community-map renaissance, and a reminder that the most consequential patches are often the ones with no balance notes.

Talkesport — Oct 22

The Blood of Dawnwalker Update 1.1 — first major post-launch patch

Out Oct 22 on PC/PS5/Xbox Series X|S with 1,100+ improvements and fixes: a skill-respec potion — the "Draught of Mind Rebirth", relearn Swordmastery/Witchcraft/Vampirism, meaningful because ability learning spends time from the 30-day calendar — a new Story difficulty below the old Story (renamed Forgiving), larger subtitle fonts, full-line subtitles, improved camera behavior. And per the patch notes: "Animal petting: Coen can now pet dogs and cats across Vale Sangora" — frankly the most important bullet of the entire edition. Dark fantasy from the ex-Witcher 3 leads (Rebel Wolves), seven weeks post-launch, doing the classic "we heard you" accessibility pass.

Keengamer — Oct 22

⭐GitHub repos

spec-decode week — click a column header to sort
RepoStarsCreatedLicenseLanguageAngle
eagle-hq/EAGLE-4923Oct 20Source-available ⚠️PythonOfficial EAGLE-4 inference code — your spec-decode watch item
deepseek-ai/DeepInfer20.6kOct 16Apache 2.0Python/C++/CUDADeepSeek's production serving engine, open-sourced
xerj-org/xerj13.3kmatureApache 2.0RustLocal-first search+MCP for agents — 2.7× fewer tokens vs grep
cloudflare/security-audit-skill40.9kJun 18MITJavaScript6-phase agent security-audit skill — REA money lane
Gathin23/cleanyourtv26.4kOct 17unclear ⚠️JavaDebloat any Android TV over Wi-Fi via ADB — agentic-built

eagle-hq/EAGLE-4 — official spec-decode code drops

923 stars since Oct 20: EAGLE-4's official inference code — "one flexible drafter for all batch sizes and hardware," harnesses for SGLang, vLLM, and TokenSpeed, auto-calibrating draft-tree depth, Docker one-liners, paper and support matrix in-repo. Measured 3.42× on Qwen3.8-27B single request, holding to 512 concurrent. The license gotcha: NOT OSI open source — a source-available license with a trademark policy and a $20M-revenue commercial-competing-use carve-out. So read it as research code you can learn from, not a dependency. It feeds your DLite/EAGLE-4 watch item and the vLLM-on-Arc spike — the spec-decode lineage you've been tracking is now inspectable.

GitHub — verified today · source-available license

deepseek-ai/DeepInfer — DeepSeek's production inference engine, open-sourced

20.6k stars since Oct 16: the engine that powers DeepSeek's large-scale serving, released as offline inference — Apache 2.0, so the license fits your driver. FlashMLA/DeepGEMM/DeepEP/TileLang kernels, MTP-style "DSpark" draft decoding, batch-invariant + deterministic sampling (the reproducibility virtues your ringer runs care about), SWA bounded-replay prefill. Two caveats: it supports only DeepSeek-V4.1-Flash at launch and targets NVIDIA SM100 + Huawei Ascend 950 — zero Arc relevance today; and it's an engine core, not a full server (KV cache system, disaggregation, load balancing not included). Built on vLLM with SGLang/TensorRT-LLM inspiration, upstream acknowledged. Second-stack candidate thinking after the SGLang kill — watch it grow hardware support.

GitHub — verified today · Apache 2.0

xerj-org/xerj — a local-first search engine that makes agents cheaper

13.3k stars, Apache 2.0, one Rust static binary: indexes code/docs/logs/JSONL/PDFs (BM25 + vector + hybrid) behind an Elasticsearch-compatible API, MCP-native. The measured claim: 2.7× fewer output tokens vs grep (26× vs memory) at the same solve rate. The core idea — "reference coding," where the agent indexes nearby OSS repos and retrieves how they solved it before writing code — ran 16/16 solves at $1.58 vs $11.18 from memory in their study. xerj init wires it into Claude Code in one command; xerj gain counts actual tokens saved. README even has an honest limits section (wins on private/post-cutoff code, neutral on memorized public libs). Local-first, cost-aware, single binary — right in the Headroom/RTK/Memora lane you're adopting, and it rhymes with the graphify direction.

GitHub — verified today · Apache 2.0

cloudflare/security-audit-skill — the adversarial auditor

40.9k stars, MIT: a six-phase structured audit — recon → coverage-led hunting → candidate validation → structured output → independent record verification → target-neutral reporting. The standout design rule: the agent that checks a finding is never the agent that found it. Only confirmed boundary failures get "confirmed" verdicts; multiple runs are additive. Fourteen attack-class playbooks included (memory-safety/binary, AI-and-LLM prompt injection, supply-chain, cloud), zero-dependency validators, npx skills add to install. Honest caveat: a single run finds roughly half the vulns — repeat runs needed. This is the seed of Cloudflare's own vuln-discovery harness, and your REA money ranking puts audit competitions and direct audits at #1 ($150k–$1M+/yr). Point Hermes at a target codebase.

GitHub — verified today · MIT

Gathin23/cleanyourtv — the viral agent-built TV debloater

26.4k stars in six days: debloat any Android/Google TV over Wi-Fi via ADB and swap the ad-filled home screen for a clean Apple-TV-style launcher — driven by a coding-assistant prompt (PROMPT.md), no root, no cable. Born as a viral tweet Oct 17 (7.4M views): the author's parents' TV cleaned by Claude Code over Wi-Fi; 170+ TV models across 200+ contributors in 4 days (TCL 57, Sony 31, Xiaomi 28… plus projectors). Mechanism: pm disable-user on telemetry/ad services — nothing uninstalled, one-line restore — plus a tiny native launcher; the assistant discovers per-brand bloat and asks before touching anything. Topics include claude-code and vibe-coding: the whole repo is agentic-built. License is unclear (README says "Licence," the fetch says NOASSERTION) — check before deploying. Pure real-life-Jarvis energy: agents doing physical-world work in your living room.

GitHub — verified today · license unclear

💛

The Wedding Countdown

…
days to go — Sat, Nov 27, 2027

Edition No. 54 lands on a round number: 400 days to go. Four hundred mornings from now, the lodge does its cozy-late-November thing. The record still reads: the lodge, weekend package Fri Nov 26 – Sun Nov 28, 2027, the whole weekend under $10k · Rip City Grill holding #1 at $14/person · vendor quotes not yet requested. Nothing booked, nothing rushed — everything still open, which is a feature, not a bug. See you there. 💛