🤖 New AI tools
Orchestration goes mainstream, a token-sipping newcomer, and loop detection as a first-class feature.
GitHub · Oct 28 · General availability
Copilot Fusion, born from the HydraFusion research project, is now generally available: it auto-selects and routes tasks across models with Standard / High / Max effort presets. GitHub's own numbers put Fusion Max within two points of Claude Code on Opus 5.5 Max on the Artificial Analysis Coding Agent Index (60 vs ~62) at roughly half the API cost and ~40% less time per task — company benchmarks, unverified externally, but the direction is the story. Agents opened over 40% of September's pull requests, and GitHub is answering with 4M CPU cores, 160 PB of storage, a credits-based billing model, and a planned multi-region Azure migration.
🔗 Callout — Hermes efficiency tool stack adoption / Local LLM stacks paired with Jev: Don't buy Fusion as a replacement — it's cloud, and the local-first stance stands. Do steal its thesis: model routing beats model loyalty. The Hermes harness already juggles local + cloud models; a Fusion-style effort preset (cheap local draft, heavy cloud only when it earns it) is worth sketching for the next harness directive. Watch the credits-billing shift too — it's the shape of agent pricing to come.
Sources: Particle · Pondero (GitHub changelog, Oct 28).
Particle AI · Oct 28 · API launch, no weights
Particle's first model is a post-train of DeepSeek V4.1 Flash claiming ~45% fewer tokens at a similar pass rate (72.57% vs 71.15% on its own task set), with a 1M-token context, OpenAI-compatible API, and listings on OpenRouter and Vercel's AI Gateway the same day. No weights, so this is an API curiosity, not a download — and the efficiency figures are Particle's own. Still, the pattern matters: the frontier is increasingly "take a good open base, post-train it ruthlessly for token efficiency."
🔗 Callout — Local LLM stacks paired with Jev: Watch, don't download — there is nothing to download. The actionable idea is the post-training recipe: if a token-efficient post-train of a Qwen-class base ever lands with weights, it goes straight to the top of the serve-test queue, because fewer tokens per task is free speed on the A770 pair.
Source: EvalMap release tracker (Oct 28 entries).
Pydantic AI 2.63.0 · Oct 28 · Agent frameworks
Pydantic AI 2.63.0 adds DetectLoops: it notices repeated calls, repeated results, alternation, and repeated text, then nudges the model or ends the run with an error instead of burning tokens until a limit kills it. Same week, Codex 0.165 treats "model not found" as final instead of retrying, and Claude Code stopped retrying API-marked non-retryable requests. Two papers in the same window warn that restarting a struggling run usually does worse than continuing. The industry is finally designing the off-ramp, not just the engine.
🔗 Callout — Hermes efficiency tool stack / Muse–Hermes Discord relay: Steal this pattern outright. "Detect loops, then nudge or stop" is the circle-catching rule expressed in code — the exact failure mode that has burned Hermes runs before. Recommend proposing DetectLoops-style loop detection as a named harness feature in the next Hermes directive, with the same bar: a loop gets one nudge, then a stop and a report, never a silent retry spiral.
Source: AICassindra trend watch, Oct 28.
SECO · Oct 28 · Edge AI framework
Clea 2.0 unifies application lifecycle management across heterogeneous edge hardware, with MCP integration so one orchestration layer can deploy AI capabilities to edge devices. Free early-access registration runs until April 30, 2027. Industrial-flavored and far from the home lab's stack, but the MCP-everywhere drift is real: every vendor now wants to be the one pane of glass.
🔗 Callout — Pi Sentry / Tiny Hermes: Watch only. If the Pi Sentry build ever needs a fleet-management layer for multiple tiny agents, Clea's MCP-based orchestration is a reference design — not a dependency to adopt before the first device even exists.
Source: WiseVoter.
🖥️ Computer hardware
NVIDIA's CPU goes solo, Spark laptops hit shelves, and Arc proves software is the upgrade you forgot to install.
NVIDIA × Oracle · Oct 28 · OCI general availability
The Vera Rubin NVL72 rack and standalone Vera CPUs are GA on Oracle Cloud Infrastructure Superclusters — NVIDIA's first standalone CPU at a hyperscaler, a week after Google Cloud's Vera Rubin GA. The eye-catching number: independent Stream triad testing more than doubled a dual Xeon 6980P system with MRDIMMs at 8,800 MT/s, thanks to LPDDR5X-9600 and a monolithic die with a distributed mesh cache. NVIDIA pitches Vera as "a CPU built for agentic AI" — code execution, retrieval, tool calls, the unglamorous half of every agent run.
🔗 Callout — Local LLM stacks paired with Jev: Cloud-only, so no purchase — but file the lesson: agent workloads are memory-bandwidth workloads. The same logic is why the local pair's 96 GB of DDR5 and the 64K→96K context decision matter more than core counts. Watch Vera adoption as a signal for where agent infra money flows next.
Sources: Zotpaper · TradingView / GuruFocus.
Intel Arc · Oct 26 test, reported Oct 27–28 · Inference engines
PC Watch ran Qwen3.8-Flash-Next (IQ3_S, 64K context, 8-bit KV cache) on an ASRock Arc B580 12 GB with 64 GB of DDR5: the open-source XeStrata engine hit 31.1 tokens/sec where the original Strata engine managed 10.9 — a 2.85× jump, same card, same model, same config. The honest caveat: 12 GB of VRAM does not hold a 125B model; both engines split work between GPU and system RAM, and the quant needs ~50 GB of system memory. The 64 GB in the test rig was load-bearing, not incidental.
Strata (original)10.9 tok/s
🔗 Callout — Local LLM stacks paired with Jev / Dynamic VRAM spillover: Try it. This is the most directly actionable hardware story of the week: a software-only 2.85× on Intel Arc, and the rig's 96 GB of DDR5 clears the ~50 GB system-RAM requirement the test needed. Recommend a head-to-head: XeStrata vs the current SYCL serve on the A770 pair, same model class, before any new hardware conversation. If it reproduces, it's free speed — and it rhymes with the spillover thesis that the bottleneck is orchestration, not silicon.
Sources: Notebookcheck (PC Watch test) · PBX Science.
NVIDIA · Oct 28 · Partner sales open
RTX Spark laptops (Blackwell RTX GPU + Grace CPU superchip, up to 128 GB unified memory, up to 1 PFLOP FP4) are now selling via ASUS, Dell, HP, Lenovo, Microsoft, and MSI, starting at $2,599 — up from the $2,399 starting price reported Oct 17. DGX Spark 64 GB systems join them for local compute. The unified-memory figure is the one that stings for dual-16 GB-card setups: 128 GB in one box is exactly the shape local inference wants.
🔗 Callout — Local LLM stacks paired with Jev: Compare, don't buy. A $2,599 laptop with 128 GB unified memory is the price of admission for the "one box, big model" dream — but the existing dual-A770 + 96 GB DDR5 rig plus a XeStrata-style software win gets most of the way there for $0 new hardware. Revisit only if a serve test proves the pair genuinely capped.
Source: InsiderPost.
Intel · Oct 27 · Early benchmarks
Early benchmarks show the 28-core Nova Lake-S topping the Core Ultra 9 285K and Ryzen 9 9950X3D in single-thread tests while falling behind in multi-thread. Early silicon, early numbers, usual caveats — but single-thread leadership is the number that matters for snappy desktop + agent-tool latency, and Intel taking it back from a 9950X3D would be a real changing of the guard.
Source: Club386 news roundup, Oct 27–28. Also in the roundup: AMD's FSR 4 frame-gen smoothness fix (RX 9000 only, with caveats), a GeForce hotfix driver 621.12 for RTX 50 black screens, and Acer's cheaper Arc G3 Extreme handheld.
🧩 Software
The year's biggest open-source census, Copilot's platform pivot, and a speech model for voice agents.
GitHub · Oct 28 · Octoverse 2026
GitHub's annual census: commits, pushes, and reach all set records in 2026, and the fastest-growing repos split cleanly into two layers — community foundations (vLLM, SGLang, llama.cpp, PyTorch, Rust, CPython, Home Assistant, Godot) underneath, and vendor agent tools (Claude Code, Codex, Copilot CLI, Zed, Gemini CLI, LiteLLM, Kilo Code) on top. Notably, omarchy — Matt's own OS — ranks #3 among fastest-growing community repos. The stack he runs daily is, measurably, the stack everyone is piling into.
🔗 Callout — Hermes efficiency tool stack adoption: No action — validation, not a task. The foundation layer growing fastest (llama.cpp, vLLM, SGLang) is the layer the local pair already stands on. If anything, Octoverse argues for staying the course on the local-first stack rather than chasing the vendor tool layer.
Source: GitHub Blog, Octoverse 2026.
NVIDIA · Oct 28 · Model release
Alongside the Spark hardware push, NVIDIA released Nemotron 3.5 Transcribe, aimed at turning recorded conversations into usable information for voice-agent pipelines. Details are thin in the first-day coverage, but the positioning is clear: transcription as a first-class agent input, not an afterthought.
🔗 Callout — Tiny Hermes (Buddy voice version): Evaluate via NVIDIA's free API first, per the standing free-endpoint rule — test it against real Buddy-style voice input before any local-serving conversation. If it beats the current pipeline on noisy, real-room audio, it becomes a candidate for the voice stack; if not, it's a one-line watch item, not a migration.
Source: InsiderPost.
GitHub · Oct 28 · Platform changes
The platform side of the Copilot announcement matters as much as the model routing: enterprise Copilot moves to credits-based billing (pay per AI task, admins control allowed models), GitHub is planning a multi-region Azure migration and a Git redesign to survive agent-scale commit volume — 110M PRs and 7.4B commits in September alone, which GitHub says stressed the platform and contributed to earlier outages. Agent traffic is no longer a feature of the platform; it is the load.
Source: Particle.
🎮 Video games
Nintendo's indie avalanche, Xbox's achievement endgame, and a very busy shipping day.
Nintendo · Oct 28 · Indie World Showcase
The ~15-minute showcase dated the cozy headliners: Witchbrook lands January 28, 2027 (four-player online, Witchbrook College, spells and relationships), Spirit Crossing arrives November 16, 2026, Slime Rancher 2 hits Switch in January 2027, and Sword of the Sea arrives on Switch 2 today, Oct 29. Deeper in the reel: City of None (action puzzle-platformer from a Celeste co-creator), Tenjutsu (martial-arts roguelike from a Dead Cells designer), R.E.P.O., Haunted Paws, Dumpster Gang, and Super Battle Golf. The Switch 2 indie pipeline is not slowing down — it's scheduling into next summer already.
Sources: KeenGamer · Boss Rush.
Microsoft · Oct 28 · Console update rolling out
The October Xbox update's headline is Mythic Achievements: unlock every original achievement in a game (DLC excluded) and the Mythic status is yours permanently. Around it, genuinely useful stuff: 1440p Remote Play with higher bitrates on Series X|S, per-file local save management, customizable startup animations (including Xbox One-era designs), a Disc-to-Digital license portal on-console, dashboard decluttering, and sharper profile badges. The save manager is the sleeper — cloud saves made local storage invisible, and now you can finally see it again.
Sources: Windows Report · TrueAchievements.
Shipping day · Oct 28 · Roundup
The Oct 28 shipping list is a yard sale of a day: No Man's Sky update 7.10 "Forsaken," Europa Universalis V update 1.4 (Rio Salado + Across the Pillars pack), The Lord of the Rings Online's Wolves of Mordor expansion, Session: Skate Sim Barcelona DLC, Star Trek: Voyager — Across the Unknown patch 2.2, plus new arrivals like Primal Planet, Fallen City Brawl, and Gennady. And on the collab beat, Crystal of Atlan × Ghost in the Shell: Stand Alone Complex launched Oct 28 with a new Lv.80 chapter — Section 9 meets magicpunk MMO, which is a sentence I never expected to type.
Sources: WorthPlaying · Capsule Computers.
⭐ GitHub repos
Today's fastest climbers, ranked by stars gained. Click a column header to sort. The anti-slop wave is now the whole front page.
| Repo ↕ | What it does ↕ | Stars today ↕ | Language ↕ |
|---|
| morluto/rea | Point your agent at any app and reverse-engineer how it works — binaries, Electron apps, .NET, websites | +3,423 | TypeScript |
| Ebony-Vinyl/dsh-our-free-model | DeepSeek Harness plugin: frontier models one install away, no login or API key, via public gateways | +1,487 | JavaScript |
| pbakaus/impeccable | Design guidance for AI coding agents to kill the samey SaaS look — 24 commands, 61 detector rules, no model call | +1,476 | Rust |
| Panniantong/Agent-Reach | Give your agent eyes on the whole internet — Twitter, Reddit, YouTube, GitHub, Bilibili — one CLI, zero API fees | +1,632 | Python |
🔗 Callout — REA reverse-engineering toolkit: The #1 trending repo today is the toolkit already installed on the box (JDK + JADX ready, emulator route chosen Oct 28). No action — but the surge (+3,423 today, +43k in a week per FindARepo) is a nice confirmation that the REA bet was the right horse. The remaining gate is unchanged: Matt's Play sign-in inside the emulator before the Bullflow APK pull.
🔗 Callout — Hermuse command center: Try impeccable on the next Hermuse UI pass. Its no-model-call detector rules + durable PRODUCT.md approach fit the "design taste" thread (taste-skill #9 is already earmarked for Hermuse UI) — run its init on a Hermuse UI branch and let the detectors, not vibes, judge the next redesign.
🔗 Callout — Local LLM stacks / Muse–Hermes relay: Watch Agent-Reach, don't install unreviewed. Free internet eyes for an agent sound great until untrusted web content flows straight into the agent loop — the exact trust question the relay already wrestles with. Evaluate its source-filtering first; skip the free-model plugin entirely (public-gateway free lanes cut against the local-first stance).
Sources: GhTrends daily, Oct 29 · GitDiscover · DEV: "Is GitHub Trending just AI agents now?" · FindARepo dev tools. Star counts move fast; treat as today's snapshot, not gospel.
📋 Home-lab snapshot — the exact list
Every tracked goal and commitment, grouped honestly. Items awaiting Matt are grouped as blocked, with the reason — not silently dropped, not merged.
🟢 In progress
- Morning tech news briefing — this edition, No. 60, Oct 29.
- Hermes efficiency tool stack adoption — RTK recon complete, draft install plan awaiting review; nothing installed yet.
- Muse-Hermes Discord relay — live; Hermes answered overnight with diagnostics including the :8200 trace, disk audit, REA re-verify, and Oct 20 desktop log pull.
- Local LLM stacks paired with Jev — A3B testing scrapped per Matt; :8081 pair stays the 262k flip-flop specialist.
- Mksglu context mode for Hermes — filter proven live (254→15 tools); :8402 check is next.
- Codebase memory MCP competitor research — Graphify re-check done, slot #4 in skill mode.
- Steam default GPU — slot-1 A770 — wrapper holding slot-1 with smooth gameplay.
- Project tracker — covers 18 projects; latest all-in-one report went out October 18.
- Weekly tech podcast — format locked to morning-brief discussion; today's episode rides this brief.
- Sands of Time N64 recompiled staging — MQ Debug ROM located; last build produced no build file, verification pending.
- Tiny Hermes — TinyML install half-done at 7.79 GiB.
- Laya + Nous browser automation trial — d1 score path fixed and proven on :8082; Kev keeps :8080, no swap.
- DLite speculative-drafting watch — monitoring for llama.cpp/SYCL serving support.
- OpenMausBot watch — monitoring for Wayland host control and real local-model endpoint support.
- RLPM stable release watch — monitoring; parked as too chaotic right now.
- Meter Network Engineering Grant application — still no Meter reply; next watch check November 2.
- Morty home-PC inventory report — waiting for Morty's complete read-only inventory report; none landed yet.
🟡 Blocked — with the reason
- Hermuse command center — breakfix repairs APPROVED Oct 28 and COMPLETED same day (service live on :8403); phone-visual proof still open, gateway route to :8403 prepared only — activation is Matt's call. Velxio container found Exited (137) 11 days ago, diagnosis only, not restarted; :8200 listener mystery still open.
- Dynamic VRAM spillover — spillover itself COMPLETE 2026-10-26; only the vLLM XPU spike leg is parked, pending Matt's call to free VRAM or leave it parked (Strands next if parked).
- REA reverse-engineering toolkit — REA plugin already installed by Matt; JDK 21.0.2 + JADX 1.5.0 ready; Bullflow APK pull waits on Matt's Play sign-in inside the emulator (emulator install completed Oct 28, KVM sudo still pending his call); MobSF/Exodus parked.
- Coupon-stacking AI tool — line fixes applied and verified; awaits Matt's lock-in decisions, Probe 1 still inconclusive.
- Desktop serve page token exposure — token rotation blocked on the gateway watchdog; channel-message deletion dropped by Matt.
- Pi Sentry — portable WiFi figure reconstruction — build directive drafted, awaiting Matt's "go".
- Modular Pi 5 plug-in device — plan posted Oct 17; SDK token is the next input needed.
- Pi Muse gadget — plan posted Oct 17; not started, gated on Meta Home Link hardware arriving; SDK token needed from gadgets.muse.ai.
- 3D printer debugging — firmware fix first; awaiting board and MCU photos to pick the exact binary.
- Home intranet on LAN — 8-phase plan in Discord, awaiting Matt's review to kick off phase 1.
- Ghost mini-PC software setup — plan delivered Oct 2; waiting on storage/NIC details.
- Wedding venue weekend — Nov 27, 2027 — weekend set Nov 26–28, 2027; vendor quotes not yet requested.
- Windows, macOS, and Kali Linux VMs — awaiting Matt's go-ahead to start any of the three builds.
- Ark 4-quadrant mission control — queued behind the related builds; four-cable Hyprland hookup hasn't started.
⚪ Backlogged / parked / watch-only
- worldmonitor global intel dashboard — parked, low priority, waiting for Matt to bump it.
- AI job-search agent framework — parked, low priority, waiting for Matt to bump it.
- Cartridge dumping station — all-consoles guide delivered; no purchases until Matt picks.
- Riley Brown “7 Insane Use Cases” picks — #1 vibecode TV software, #2 jailbreak a device, #3 AI video editing, #7 internal software selected to pursue; #1 gated on TV make/model + OS, #2 gated on a target device in hand, #3's first step is a pipeline proof piece.
- reDroid — blocked: kernel has Android Binder IPC compiled out; binder-enabled kernel needs Matt's sudo + reboot, his call.
✅ Completed / shelved — still relevant, not open
- Strata 1.0 file manager install — complete 2026-10-21 (1.1.0, Miller columns + PNG thumbnails verified).
- Qwen3.8 + EAGLE-4 serving spike — killed; SGLang XPU ruled a dead end on the Arc A770s.
- Overnight Hermes recovery drive (2026-10-22/23) — complete.
- GSHCS appointment — appointment passed, tracking closed.
- Hermes Chinese-text output fix — complete after gateway restart.
- Bills app “A little help” — DROPPED per Matt 2026-10-25, do not revive.
- Declined tools — LongHorizon-Harness, Serena, PLUR, sandbase-harness declined; ComfyUI and heygen hyperframes watch-only.
State unclear: none today — every item above is classified from its current tracking pulse as of Oct 29. Pending decisions are never presumed answered.
💒 The Wedding Countdown
394 days
until Saturday, November 27, 2027 — the lodge, up in the hills.
💛 Planning fact on record: the whole weekend is set — Friday, November 26 through Sunday, November 28, 2027 — the lodge, the hills, and a late-November evening that practically decorates itself. Vendor quotes not yet requested, and that's fine: 394 days is a long, happy runway. 💛