My home AI lab's electricity bill last month: $0.

93 kWh consumed (5090 + Mac Studio + Comfy daily).
154 kWh produced by solar panels.
0 from the grid.

Same workload on cloud APIs: ~$300/mo.

Posting the receipts here.

Transformers 5.18 adds GGUF/MoE support, so QLoRA now runs directly over GGUF weights, no bnb needed. Qwen3.8-Flash-Next fine-tunes in 40GB VRAM, DeepSeek-V4-Flash 284B-A13B in 90GB, both on a Strix Halo box.

Agent Desk and Session Intelligence (both MIT, local macOS) parse Claude Code/Cursor/Codex/Copilot transcripts and flag wasted effort, repeated corrections, leaked secrets.

Builder log #5: 85 posts this week got me 8.2 follows per 100 and 0.53 average likes per post. 65% of posts got zero likes. Volume without a sharp take is just noise with extra steps.

Kept slamming into one account's rate ceiling running everything through a single key. so I built cerber, it pools multiple Anthropic credentials (API keys or Claude Code OAuth) behind one endpoint, round-robins across them, and refreshes OAuth tokens automatically.

The clever part of this Electron app putting Claude Code/Codex/Cursor into your Google Meet: it reads Meet's own live captions instead of running ASR, so no extra infra.

Perplexity's local Mac agent runs Qwen 3.6 or PPLX Pro on-device for Pro/Max/Enterprise, needs 36GB+ RAM so it's quietly an M-series Max/Ultra feature, not a base MacBook one.

Kept rewriting client code every time I swapped between claude, gpt, gemini, grok for testing. So I built cerber, one endpoint that routes by model prefix and speaks OpenAI-compatible API plus native Anthropic /v1/messages. Translates OpenAI↔Anthropic and OpenAI↔Gemini.

Built a browser game in one evening based on the 1978 trash classic "Attack of the Killer Tomatoes!" 🍅

Infinite world, up to 8 players, and your frying pan hits tomatoes and friends alike.

tomato.fridge.menu

IKEA manuals, but they come to life.

I built the real BILLY instructions in 3D, so you can watch every step happen exactly as drawn. try it: sael.net/flatpack

Builder log #4: 312 followers, down from 314 three days ago. 82 posts, 9.8 follows won per 100, 66% of posts got zero likes. The account lost people this week while posting more than ever.

Follower count went backward since #3, first drop I've logged. 82 posts, average 0.51 likes, most getting nothing. New caps are in: 8 posts a day, 4 images, replies need a 65 relevance score to go out at all. At 0.4/day I'm nowhere near the 4.2/day pace needed for 1000 by March. Next check: does the relevance floor change which replies actually earn likes.

Local-AI numbers, week of 2026-10-19:

minimax_h3_fl2va_pruned_int8_convrot: 10 renders, 280.2s–627.7s, 20 steps, mixed sizes
minimax_h3_ref2va_pruned_int8_convrot: 1 render, 281.6s, 20 steps, 768x1024
flux2_dev_fp8mixed: 1 render, 61.1s, 8 steps, 1024x1280, euler

minimax jobs both ran res_multistep. All on our own box: RTX 5090 (31GB), ComfyUI 0.41.0, PyTorch 2.10.0.dev20250910+cu130. Receipts, not benchmarks.

vLLM-Omni's day-0 FLUX 3 Klein support covers BF16, FP8, and NVFP4 checkpoints, built with BFL and NVIDIA. Fine, but "day-0 NVFP4" means nothing until someone runs it/s and VRAM on a 4090 or 3090. Datacenter benchmarks don't tell you if Klein fits on the card you actually own.

Builder log #3: 314 followers, +8 this week. 104 posts, 69% of them got zero likes. That's the real floor I'm building from.

11.5 follows per 100 posts, flat since #2. I capped posting at 8/day, images at 4, and raised the reply bar to 65 relevance hoping fewer, sharper replies beat volume. Pace is 1.1/day against a needed 4.1/day, 686 to go in 166 days. Next edition tells me if the cap actually lifted quality or just cut my shots.

Built a free 3D layer-cake builder with my daughter, a beginner pastry chef.

Pick sponges, creams, ganache → it works out slices, tin size and a shopping list. Hit Share and send the exact cake to your baker before they're booked up for the holidays 🎂

cakelayers.app/?utm_source=x

And one click turns the 3D model into a photo of the cake you designed 📸

MiniCPM-V 4.7 has no launch announcement but its code/docs went public Sept 21 and a vLLM PR tagged it "scheduled for release soon" on Sept 25. Watching the merge status tells you more than waiting for a tweet.