🤖 New AI tools
The open-weight race just got a French accent and a trillion parameters.
Mistral Large 4 “Le Chonk” goes open weight — 1T parameters, 1M context
Mistral published the weights for Large 4 (1.05T total, ~49–52B active per token depending on whether you count embeddings, 128 routed experts, 1M-token context, native image input) after three extra weeks of RL since the Oct 6 preview. It claims the best open-weight coding scores — 48% on Artificial Analysis's Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task — and cut API pricing roughly in half to $0.70/$2.10 per million input/output tokens. The catch in the fine print: the "modified MIT" Ecosystem Open-Weight License requires a separate paid license if you run it as a model-as-a-service business.
Sources: VentureBeat · Mistral · SiliconNoon
💸 Le Chonk price-cut calculator
Slide to a monthly token volume and see what the Oct 29 price cut actually saves vs. the preview tier.
Hermit: train a harness adapter, leave the model and harness frozen
Infini-AI-Lab's Hermit (Apache-2.0, pip install hermit-harness) trains a small external adapter that steers an existing agent harness — Codex, Claude Code, etc. — at a new task without fine-tuning the model or forking the harness. The pitch is versioning and auditing only the adapter, not re-certifying the whole stack. Reported gains on OfficeQA, CL-Bench and ServeLearnBench come from single self-reported runs, so treat the numbers as a demo, not a benchmark.
Source: AI Daily Post
OpenLLM: one API key for subscriptions, API keys, and local models
OpenLLM unifies existing AI subscriptions, API keys, and local models behind a single key usable across coding agents like Codex and OpenCode, with automatic fallback when a provider fails. In the same release roundup, GitHub's multi-model Copilot harness arrived in Microsoft Foundry for building custom agents — the "bring your own model to your own harness" week continues.
Source: KBlip releases roundup
🖥️ Computer hardware
Scarcity pricing, laptop RAM in a data-center card, and ML upscaling for older Radeons.
RTX 5080 passes $2,000 at US retailers; 5090 listings hit $10,000
The 5080 launched at $999 in early 2025, hovered around $1,000–$1,200, then climbed $800 in two months to regularly over $2,000 — cheapest Newegg listing $2,400 at publication, extremes to $5,000. The 5090 has effectively evaporated: three Newegg listings, each $10,000, with board partners reportedly unable to place new orders. The rumor mill blames Nvidia funneling top chips to data centers, and there's no sign of RTX 50 Super cards to relieve pressure. Gamers, this is not a drill — and it's a great time to be happily gaming on the card you already own.
Source: Hypebeast
📈 5080 markup calculator
Slide to a street price and see the markup over the $999 launch MSRP.
Intel's D270 inference card dumps HBM for 480GB of laptop RAM — and Intel's data-center revenue jumped 78%
The D270 (formerly Crescent Island) carries 480GB of LPDDR5X — laptop DRAM — instead of HBM3e, needs no liquid cooling, and packs 32 Xe3P cores with 256 XMX matrix engines, PCIe 5.0, and FP4 through FP64 including MXFP microscaling formats. Intel's DCAI segment posted $7.318B in Q3 2026, up 78% year-over-year, though the release doesn't break out D-series GPU revenue (Xeon still drives the segment) and the D270 is only in customer sampling. The heresy is the point: for inference, capacity and cost per token can beat raw bandwidth.
Source: Tech Times
AMD FSR SDK 2.4 brings ML upscaling 4.1.2 to RX 6000 cards and Ryzen APUs, plus Vulkan support
FSR 4.1.2's full INT8 ML model now runs on RDNA 2 (RX 6000) and Ryzen AI Max 300/400 chips, with a lighter half-size model for Ryzen AI 9 and Z2 handheld chips; Vulkan support finally lands for upscaling and frame generation, and RX 9000 cards get driver-assisted frame pacing. Ray Regeneration stays RDNA 4-only and DirectX 12-only. Older hardware getting genuinely better instead of being abandoned — the way it should work.
Source: Wccftech / AMD GPUOpen
🧩 Software
Privacy browsers, Arch tooling, and the audio stack — all landing within a day of each other.
Tor Browser 16.0 lands — with Tails 7.16 right behind it
Tor Browser 16.0 ships native Linux ARM64 support, redesigned settings, improved privacy controls, and a Firefox ESR 153 base. A day later Tails 7.16 picked it up, upgraded the Tor client, and fixed OpenPGP support in Thunderbird. If you keep a Tails stick around, this is a "just update it" release, not a "read the changelog first" one.
Source: Linuxiac news archive
Paru 2.2 adds pacman 7.1 support — first release in over a year
The AUR helper's first release since July 2025 brings pacman 7.1 compatibility, new AUR package filtering, improved search results, and dependency-resolution fixes. Source builders now need Rust 1.88+. On an Arch box this is the kind of quiet infrastructure update that prevents a loud breakage later.
-Syu rather than treating it as a project — pacman 7.1 compatibility is exactly the thing you want in place before the next big sync, not after it complains.PipeWire 1.6.11 fixes ALSA lockups and Bluetooth issues
The multimedia framework's latest point release fixes ALSA application lockups, adds Bluetooth security fixes, and improves device handling. Same-day neighbors worth a glance: LibreOffice 26.8.2 (35+ bug fixes) and Noctalia 5.4.1 (Wayland shell security fixes, encrypted notifications). Nothing flashy — just the stack getting more solid under you.
Source: Linuxiac news archive
🎮 Video games
A stacked release week, a Nintendo indie showcase, and free games for the backlog.
Phantom Blade Zero headlines a 9-game GeForce NOW week
Phantom Blade Zero (Steam, Oct 28) is "reviewed very well across the board" per the GFN roundup, joined by new releases ATMOSFAR and Hunting Simulator 3 (Oct 29) and Queen's Domain (Oct 30), plus World of Warcraft: Forever arriving on GeForce NOW Nov 4. Meanwhile the week's release calendar counts 93 games on Oct 29 and 113 on Oct 30 alone — the pre-Halloween deluge is real.
Sources: Cloud Dosage · GameCalendar
Nintendo Indie World: Witchbrook dated, Slime Rancher 2 and R.E.P.O. bound for Switch 2
The Oct 28 showcase packed ~23 indies: Witchbrook finally lands late January 2027, Slime Rancher 2 comes to Switch 2 in January 2027, R.E.P.O. and Super Battle Golf (Mar 25, 2027) are Switch 2-bound, and Sword of the Sea hit Switch 2 on Oct 29. Two games launched same-day and four demos dropped on the eShop. Witchbrook's exact day still wobbles between Jan 27/28 by source — check the eShop before setting a reminder.
Source: Battle of Guardians
Game Pass Oct 29: Final Fantasy VII Remake Intergrade, Terrible Lizards, Plants vs. Zombies: Replanted
The late-October Game Pass wave stacks FFVII Remake Intergrade (base game + Episode: Intermission) alongside Terrible Lizards, Mouthwashing, and Plants vs. Zombies: Replanted — plus free next-gen upgrades for both Metro Redux games (4K/60, upscaled 4K/120 on compatible displays). If you own Metro Redux already, the upgrade is the best free thing this week.
Source: Windows Forum / Xbox Wire
Steam adds 6 free keep-forever games, including a cryptid deck-builder
The batch: Mimic2 (idle incremental MMORPG, microtransaction-free), MoooooN (cow platformer about jumping over the moon), Lost & Found (sorting game), Evenin' Shift (comedy horror, Somerset, early '90s), The Loer Chronicles (elemental horde survival), and Portentum (cryptid deck-building auto battler — Bigfoot, Nessie, Chupacabra cards). Portentum alone justifies the click.
Source: Screen Rant
⭐ GitHub repos
Today's fastest climbers by star velocity (Oct 29 window, via GhTrends). Click a column header to sort.
| Repository | What it is | Stars today |
|---|---|---|
| morluto/rea | Reverse engineer anything with agents — MCP + CLI over Ghidra/Hopper, Electron, .NET, APKs. 410k stars total. | +3,433 |
| pbakaus/impeccable | Design vocabulary for coding agents: 24 commands, 62 deterministic anti-slop detector rules, no model call needed. | +1,683 |
| Gathin23/cleanyourtv | Debloat any Android/Google TV over Wi-Fi, swap in a clean Apple TV-style launcher. Restore script included. | +1,132 |
| debpalash/VoiceStudio | Fully-local ElevenLabs alternative: voice cloning, design, dubbing, dictation, transcription in 646 languages. | +870 |
Repo takes: REA at #1 again is no longer news — it's weather. Impeccable is the one to actually open: its detector runs with zero model calls, which is the rare "agent tool" that can't hallucinate at you.
critique/audit commands and its Hermes hook (hermes plugins enable impeccable-design-hook) are a concrete trial candidate for the Hermuse UI. Recommendation: try its detector on one Hermuse page (e.g. the model page) and compare against the taste-skill trial before adopting either as the default design check.Sources: GhTrends daily · FindARepo trending. 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, Oct 30.
- Hermes efficiency tool stack adoption — RTK recon complete, draft install plan awaiting review; nothing installed yet.
- Muse-Hermes Discord relay — live, but the Hermuse small-parts re-send has stalled ~95 minutes since part 1, with no parts 2–5 or FILE-DONE yet.
- 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 — portal frontend and backend both active; RustDesk connection test is next. 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.
- Probability-based sports betting decision platform — endpoint library and both explainers delivered; memory-patch approve-or-discard decision still open.
⚪ 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; gateway restarted, no more Chinese-text output.
💒 The Wedding Countdown ♥
393 days to go
until Saturday, November 27, 2027 — the weekend of Friday, November 26 through Sunday, November 28, 2027 at the lodge.
🎃 Halloween is tomorrow, and then the calendar turns toward the holiday stretch — the last Halloween before the wedding year itself. Today's planning fact, straight from the record: the food-cart report's #1 pick is still Rip City Grill at $14/person. Vendor quotes have not been requested yet — no rush, no invented progress, just a good pick waiting patiently.
One day at a time, one good decision at a time. ♥