HOME GITR3KT Morning Tech Brief — No. 62 · Saturday, October 31, 2026 Saturday · #ff6b4a ← All briefs
🎧 THE HOME LAB BRIEF — TODAY'S EPISODE
Morning Tech Brief — Oct 31, 2026 · No. 62

Delivered with the brief, and the voice is live: the player vocodes the clean voice master in your browser, in real time. Pop it out, keep reading or wander the site — then drag the sliders in the Robot Voice Tuner and hear it change while it plays.

Broadcast MP3 Clean voice MP3
Morning Tech Brief · Saturday, October 31, 2026 · No. 62

Your TV is watching back, and GPUs have left the building

A researcher took apart his own smart TV and found the surveillance state in the firmware, the frontier models went full cyber-defence, and the graphics-card market achieved peak absurdity. Happy Halloween.

Today's edition in one breath: Google's Gemini 4 Carbon tops the Artificial Analysis index while locked to cyber defenders, OpenAI's GPT-6.1 Terra undercuts its own predecessor on price, Liquid's 2B models now run offline on a phone, and Nvidia's top two consumer cards have quietly vanished from its own US store. Plus: the Low Level TV reverse-engineering story Matt flagged, with the receipts, the caveats, and what it means for the lab's REA work.

🤖 New AI tools

Cyber defence is the week's theme, and the small models are the ones you can actually touch.

Model release · Oct 30

Gemini 4 Carbon tops the charts — and almost nobody can use it yet

Google DeepMind's Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index v5 — a point ahead of Claude Opus 5.5 — with headline cyber numbers of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, plus a 1M-token output window. The catch is access: for now it's limited to a subset of cyber defenders through Google's Fairwind Program at $2/$10 per million input/output tokens, with broader access promised later. A frontier model whose launch benchmark suite is vulnerability discovery, exploitation and patching is a statement of intent, not a coincidence.

Sources: AI Tools Recap · The Century Report, Oct 30

Model release · week of Oct 26

GPT-6.1 Terra: cheaper agentic coding, as OpenAI prunes the family tree

OpenAI's GPT-6.1 Terra targets cheaper agentic coding, scoring a reported 15 percentage points higher than GPT-5.6 Terra at about one-third the cost, as OpenAI reportedly discontinues the old Terra line and consolidates around a three-tier Luna / Sol / Astra family. If the numbers hold, this is the rare release that's good news twice: better reasoning effort for less money, and fewer model names to keep straight. Treat the family-tree reshuffle as reported, not yet an official OpenAI blog post.

Source: This week in tech — 31st Oct 2026 (Medium)

Local models · Oct 30

LFM3 now runs fully offline on phones — 40 tokens/sec on an iPhone 16 Pro

Liquid AI's LFM3 family (six 2B-class models on one backbone, 256K context on the 1.7B text model, ~1GB at 4-bit) now runs entirely on-device through the open-source Private Mind app: a Reddit report puts LFM3-1.7B at ~40 tokens/sec with sub-half-second first token on an iPhone 16 Pro, and the vision model at ~20 tokens/sec. Those phone numbers are single-source and unverified, but the direction is what matters — a 2B model that fits in a gigabyte is a model that fits everywhere, including hardware the lab already owns.

🔗 Lab callout — Local LLM stacks paired with Jev / Tiny Hermes: this is the cheapest experiment on the board. Recommendation: try LFM3-1.7B on the lab rig this week — at ~1GB it costs nothing to serve next to the Qwen pair, and it doubles as the obvious candidate for the Pi 5 side of the house, where Liquid explicitly tuned for Raspberry Pi 5-class hardware. Compare it on inbox-triage-style tasks against the current small-model setup before judging it on coding, where tiny models are weakest.

Sources: AlexTech · Liquid AI

🧮 LFM3 memory math

Weights are ~1GB at 4-bit; a full 128K-token KV cache needs ~1.6GB. Slide the context and see the total footprint.

On a 16GB card this is a rounding error — which is exactly the point. The lab's VRAM fights are about 30B+ models; 2B models are a solved problem.

Agent runtime · Oct 30

AWS + UC Berkeley's Infinity: 75,000 concurrent agents, event-driven

AWS and UC Berkeley released Infinity, an agent runtime built around event-driven execution and rated for 75,000 concurrent agents. The number is a benchmark claim, not a lab result, but the architecture choice is the interesting part: event-driven execution is how you stop paying for thousands of agents sitting idle polling. For anyone running more than a handful of agents, the runtime — not the model — is increasingly where the bill lives.

🔗 Lab callout — Hermes efficiency tool stack adoption: Infinity sits at the harness/runtime layer the lab already invests in, but it's a research release, not a drop-in Hermes replacement. Recommendation: watch it — revisit if it ships a Hermes-compatible adapter or a reproducible 75k-agent demo on ordinary hardware; do not re-architect the harness around a press number.

Source: This week in tech — 31st Oct 2026 (Medium)

🖥️ Computer hardware

The GPU market didn't cool off. It went full haunted house.

GPUs · Oct 30

AMD GPUs get a second price hike in two months — RX 9070 XT nears $1,000

ChannelGate reports AMD will raise GPU + GDDR6 bundle prices for board partners in November, after an October increase, driven by an 8–10% rise in GDDR6 prices in Q3. The RX 9070 XT launched at $599 in March 2025; most listings now exceed $900, with premium cards past $1,000. Nvidia GPU prices rose 40%+ in the prior three months per the same coverage, AMD 28.3% — so this is AMD catching up, not AMD leading. Either way, the $599 card is a museum piece now.

🔗 Lab callout — Dynamic VRAM spillover / Steam default GPU — slot-1 A770: the lab's two 16GB A770s, bought before the madness, keep looking smarter. Recommendation: buy nothing, hold the slot-1 setup — no GPU purchase in this market improves the lab's inference-per-dollar, and the spillover work already squeezes more out of the cards on hand than a new card at 1.7× MSRP would add.

Sources: PC Guide · Wccftech

GPUs · Oct 30

RTX 5080 vanishes from Nvidia's US store — and the 5090 went with it

Nvidia's own US Marketplace no longer lists the RTX 5080 or 5090 — the 5090 disappeared around Oct 16 (briefly returning at $5,599 vs its $1,999 MSRP), and the 5080 followed on Oct 30, leaving the 5070 Ti as the top listed card. The UK store still shows both. Nvidia says the 5090 isn't discontinued; it hasn't explained the 5080. Meanwhile street checks found a PNY 5080 at $1,999.99 at Best Buy and a third-party 5090 at $9,399.99 on Newegg. A marketplace delisting is not a discontinuation — but when your own store won't sell your flagship, the scarcity story writes itself.

Sources: TechTroduce · PCGeek price check, Oct 30

📈 GPU absurdity calculator

Slide to a street price and see the markup over launch MSRP ($999 for the 5080, $1,999 for the 5090).

Cloud silicon · Oct 27–30

Oracle puts Vera Rubin NVL72 on general availability — and Google TPUs come to OCI

Oracle's expanded Nvidia partnership makes Vera Rubin NVL72 systems and Vera CPUs generally available on OCI, with an 8-GPU HGX Rubin NVL8 config planned, and lets customers scale CPU and GPU independently; separately, Google TPUs are coming to Oracle Cloud users. Nvidia shares fell 4.7% to $203.71 on Oct 30 anyway — the market read "more ways to buy accelerators" as "more competition for the accelerator dollar." No order value or revenue is attached to the announcement, so treat it as availability, not demand.

Sources: TradingView / GuruFocus · This week in tech (Medium)

🧩 Software

One-edition special first: the video Matt sent, researched properly.

★ One-edition addendum · Low Level TV · Oct 30

Reverse engineering proves what we suspected: a Samsung TV's firmware, taken apart

The video covers researcher @log4j's reverse engineering of his own 2025 Samsung smart TV's Tizen firmware. What was actually found, per the researcher's posts and corroborating coverage: a security-policy.xml listing functions including an always-on microphone path, automatic content recognition (ACR), geo-positioning, network mapping, a targeted-ads service and user-profile building; on-device computer-vision models monitoring activity inside apps (videos watched, ads skipped, playback progress); and, on TVs without a built-in mic, code that looks for microphones on connected soundbars or Bluetooth devices and routes audio to Bixby. The evidence is the researcher's firmware analysis plus a Flipper Zero test showing a mic opening for ~5 seconds on a rooted set — not a proven recording exfiltrated off an unmodified TV, and the researcher's full write-up is still in preparation. Samsung has not publicly addressed the specific claims; a community note points out Samsung openly advertises sound-detection features (Pet Care / Family Care bark-and-cry alerts via SmartThings), which covers part — but only part — of what's listed. The twist that made the rounds: OpenAI suspended @log4j's account for alleged “Cyber Abuse” after he used its tools for the local reverse engineering, and Claude has reportedly been truncating similar REA work mid-generation. Reverse-engineering your own device is not new; the cloud labs treating it as abuse is.

🔗 Lab callout — REA reverse-engineering toolkit: this is the lab's own thesis, playing out in public. Recommendation: study the write-up when it lands and diff its method against the lab's REA toolkit — firmware extraction, the security-policy manifest, and the ACR/audio code paths are exactly the surfaces REA is built to map. Do NOT route this kind of work through cloud models: the bans are the proof that the cloud path is unreliable for REA. The lab's local-model lane (Qwen pair, and now LFM3-class small models for triage) is the durable route — keep REA local, keep the evidence raw, and treat the Samsung claims as allegations until the write-up's artifacts can be checked.

Sources: Low Level TV (video) · Decrypt · TokenPost

Linux · Oct 30–31

Linux roundup: kernel 7.3.1, Wine 11.21, XanMod kernels, Arch refreshes

Kernel 7.3.1 lands with KASLR-leak, coredump-race and io_uring fixes, alongside XanMod EDGE/MAIN/LTS builds for 7.3.1, 7.2.14 and 6.18.60; Wine 11.21 brings faster 32-bit OpenGL via WoW64, voice-capture DSP and ARM64 work; Arch refreshed linux-rt-lts 6.18.60, Node.js 26.13 and the mise dev-env tool on Oct 30–31. None of these is a headline alone — together they're the rolling-release drumbeat doing its job. On an Arch box, this is a pacman -Syu morning, not a research project.

Sources: Linux Compatible archive · Arch: linux-rt-lts · Arch: Node.js 26.13

Distros · Oct 30

Ubuntu 27.04 “Tenacious T-Rex” opens for development

The Tenacious T-Rex archive opened Oct 30, with auto-sync from Debian pulling in post-26.10 packages ahead of a scheduled Apr 22, 2027 release (beta Apr 1). Expect GNOME 52 and possibly the Linux 7.6 kernel. Matt's box is Arch, so this is spectator sport — but Ubuntu's six-month cadence is still the metronome the wider Linux desktop marches to, and T-Rex is a gloriously unserious codename for a serious release.

Source: OMG! Ubuntu

Desktops · Oct 30

Manjaro testing previews GNOME 51, Plasma 6.8 and COSMIC 1.10

Manjaro's Oct 30 testing update — a preview of Manjaro 27.0 Caldonia — bundles GNOME 51.1, KDE Plasma 6.8.2 and COSMIC Epoch 1.10, plus Firefox 159, and retires X11 session support in Plasma 6.8. It's a testing branch, so the usual caveats apply, but three desktop environments landing a major version in one update is a good weather report for the Linux desktop: Wayland is no longer the future, it's the default with the old session being shown the door.

Source: Manjaro forum

🎮 Video games

Halloween patches, a one-life-mode fix, and a desert DLC.

Patch · Oct 30

Deadlock's Halloween update: candy cauldrons, a beefier Mid-Boss, and the meta's worst offenders nerfed

Valve's Oct 30 Deadlock update touches 34 heroes, 38 items and 33 general systems: the Mid-Boss gets 20% more health, resistance reductions are halved against it, and its Rejuvenator now splits into three breakable crystals; Rat King, Nurse Harrow and friends take the hardest nerfs while Abrams and Drifter get straight buffs. Crates are pumpkins, a cauldron hands out consumable candy, and a hideout upstairs teases a paid Wardrobe skins feature. A balance patch wearing a costume — the good kind.

Sources: TheGamer · SteamDB patch notes

Patch · Oct 30

Modern Warfare 4's Oct 30 patch fixes the premature Search and Destroy kick

Infinity Ward's Oct 30 notes fix players being kicked for inactivity prematurely in Search and Destroy, plus a private-match black screen on team-switch. The open question, honestly flagged by the coverage: the Oct 25 notes already claimed the spectate AFK kick was addressed, and the new note doesn't say whether this is that bug returning or a second timer. In a one-life mode, an idle timer that reads a dead player as an absent one isn't a minor bug — it's a bug aimed at the mode's whole point. Fixed per the notes; cause still unsettled.

Source: Hardcore SND

DLC + patch · Oct 29

Crimson Desert Enhanced gets its first DLC: Charting the Unknown

Pearl Abyss shipped the Charting the Unknown DLC for Crimson Desert Enhanced on Oct 29 alongside a major patch — improved NPC interaction UI (chat and gift options), plus a host of fixes. The devs say no weekly updates, just major patches at a regular clip. First-DLC energy with a real patch behind it, not just a content drop and a prayer.

Source: EGW News

Upcoming · Nov 9–13

Diablo II: Resurrected Ladder Season 16 starts Nov 13 with Patch 3.4

Patch 3.4 rolls out Nov 9 (Season 15 ends 11 a.m. PST that day), and Ladder Season 16 starts 5 p.m. PST Nov 13 with item changes in tow. A quarter-century-old game still running a ladder calendar this tight is either admirable or a cry for help — either way, the ladder crowd knows what to do with the date.

Source: PatchBeat patch notes

⭐ GitHub repos

Click a column header to sort. Star counts are today's snapshot, not gospel.

RepoWhat it isStarsLanguage
Orkas-AI/OrkasOpen-source desktop AI team app: goals become to-dos, specialist agents coordinate to finish them. +1,004 stars today.5.1KJavaScript
lightpanda-io/browserHeadless browser written from scratch (Zig) for agents and automation — not a Chromium fork. v1.0 shipped Oct 2.38.5KZig
jakubkrehel/skillsSkill pack with an acceptance bar — the agent-skill format maturing from vibes to a bar you can check.9.4KMarkdown
Gathin23/cleanyourtvDebloat any Android/Google TV over Wi-Fi for an Apple-TV-style home. 107 TV models listed — and very on-theme today.47KPython
thedotmack/claude-memPersistent context across agent sessions: capture and compress what matters, recall it next time.trending—
🔗 Lab callouts — repos: lightpanda browser → Laya + Nous browser automation trial: a from-scratch headless browser built for agents is exactly the kind of tool that could replace fragile DOM-scraping lanes — try it as the browser backend for one automation trial before buying deeper into it. claude-mem → Codebase memory MCP competitor research: persistent agent memory is the same problem Graphify/codebase-memory-mcp attack from the code side — compare it against the current memory approach on one real multi-session task, don't adopt on star count alone. cleanyourtv → REA / home lab: watch — it's the practical, Android-TV-only cousin of today's Samsung story, and a reminder that debloat tooling only works where the platform lets you in.

Sources: GitDiscover daily picks, Oct 31 · Open Source Radar, Oct 31 · Trendar

📋 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. Roster checked Oct 31, 2026.

🟢 In progress

  • Morning tech news briefing — this edition, No. 62, Oct 31.
  • Coupon-stacking AI tool — Coupon Stack v3 live on seed data; live ingestion waits on the Reddit filing due today and proven store endpoints.
  • Probability-based sports betting decision platform — Probability Desk live as Tool T5: manual-entry devig, EV, fractional-Kelly, NO BET as a first-class verdict.
  • gitr3kt.com public website — Tools T4–T6 live, snarky house voice on the main site; board ends at Live 9, Shelf 11, Progress 0.
  • Weekly tech podcast — Brief No. 61 episode live with pop-out player and live in-browser vocoding; automatic from No. 62; ear-test pending.
  • Hermuse command center — portal frontend and backend both active; RustDesk connection test is next.
  • Google The Dalles applications — 2× wage target — resume approved; FM application still unsubmitted at Step 1, pending Matt swapping in the final clean-named resume and checking the legal-name mismatch.
  • Hermes efficiency tool stack adoption — RTK recon complete, draft install plan awaiting review; nothing installed yet.
  • Muse-Hermes Discord relay — stall cleared Oct 30 17:03 PDT; coupon build blocked pending re-send of chunks 8–13, VRAM search returned no counters.
  • Local LLM stacks paired with Jev — A3B testing scrapped per Matt; unit stays the 262k flip-flop specialist; fastest-10 32GB-VRAM research delivered, Qwen3.5-9B flagged as next serve test.
  • 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, codebase-memory-mcp parked as phase-2 complement.
  • 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.
  • 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; gateway restart it was parked behind now complete.
  • 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.
  • Cartridge dumping station — all-consoles guide delivered; no purchases until Matt picks.

🟡 Blocked — with the reason

  • Dynamic VRAM spillover — spillover 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 — Bullflow APK pull sidebarred until the remote portal is proven; emulator Stage 1 held ready.
  • 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”; guide updated with RuView home-base track.
  • Modular Pi 5 plug-in device — plan posted Oct 17; SDK token is the next input needed.
  • 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.
  • Shepherd Outsourcing debt-validation letter follow-up — both letters still unsent; LVNV window closes in ~12 days (Nov 12), Capital One's in ~29 days (Nov 29). Reminder scheduled Nov 5.
  • 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.

✅ 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; overnight window ended 8 a.m. PDT.
  • GSHCS appointment, Hermiston — Oct 27 3:20 p.m. appointment passed, tracking closed.
  • Hermes Chinese-text output fix — complete; gateway restarted, no more Chinese-text output.

Roster note: three items from yesterday's snapshot — Pi Muse gadget, Riley Brown “7 Insane Use Cases” picks, and reDroid — no longer appear in the current tracking roster. State unclear, not presumed closed; they are flagged here rather than silently dropped.

💒 The Wedding Countdown ♥

392 days to go

until Saturday, November 27, 2027 — the weekend of Friday, November 26 through Sunday, November 28, 2027 at Scholls Valley Lodge.

🎃 Happy Halloween — the last one 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, and vendor quotes have not been requested yet. No rush, no invented progress — just a good pick, waiting patiently, while the candy bowl gets emptied tonight.

One day at a time, one good decision at a time. ♥