A nightly script collects AI leads from free sources, a local model scores them, Claude writes five deep-dive notes in one half-price Message Batch, and the agent's morning brief carries a verdict and a link for each. Sources, setup choices, code and the first test run.

Self-hosted Honcho gives my home AI agent long-term memory: llama.cpp on a GPU box does the reasoning, Ollama on a Mac mini does the embeddings, and the stack runs on my own machines. Compose changes, per-feature model config, the embedding bootstrap order and a day-one test.

Five provider keys moved from a .env file into a dedicated Proton Pass vault. The agent reads them at startup with a viewer-only token that expires in 90 days and logs each read with a reason. Vaults, tokens, plugin config, migration, testing, and a wrapper that hands one secret to one tool.

One Audiobookshelf container, read-only NFS from the NAS, five libraries, per-user accounts, and iPhone access on the LAN or over a VPN. Setup steps and the settings that matter.

Text an agent a product link and a target price; a phone alert arrives when it hits, including the lowest used offer on Amazon. PriceBuddy self-hosted and hardened, a small helper the agent calls, and an alert job that uses no model at all.

A used desktop with an 8 GB RX 7600 and a gaming PC with an RTX 5090 run the same ComfyUI graphs on Z-Image Turbo and Krea 2 Turbo. Both make clean photoreal images; the 5090 is 20x faster and uses 5 to 6x less energy per image.

A used 14900K desktop with 128 GB of DDR5 and an RX 7600 runs Qwen3.8-Flash-Next with vision and a 262K window at about 12 tokens per second, with benchmarks for five models and the settings behind them.

Paperless-ngx 3.3 on a small homelab box, Tika and Gotenberg on the NAS, a local 27B model for labeling and chat, and a WireGuard hub on a VPS so the archive opens on my phone from anywhere even though home internet is 5G CGNAT. Hardware split, versions, settings, and every gotcha with its fix.

oMLX with the MTP head and the n-gram table on SSD holds about 60 tok/s from empty to 245K; llama.cpp with MTP runs 59 fresh and 46 at 245K. Full window, vision on, identical output.

Dell Pro Max with GB10 at about $5,523 versus the $6,950 DGX Spark, Strix Halo, Mac Studio M5, and used 3090s. Plus what a two-box pair really does.

Every major result on experience in language models, 2021 to October 2026, with what each one measured and what it did not.

A friend ran two copies of a local model as Freud and a depressed android for twenty turns and sent me the transcript. I turned it into 50 consistent photographic frames, and it gave me a project to do with my daughter.

A LoRaWAN pressure probe on a gas wellhead reports every 20 minutes to a self-hosted Prometheus and Grafana stack, showing exactly how the well recovers between draws. Here is how to build it.

Water on the sensor's rope pages a phone in about 25 seconds and gets announced out loud in the house. Here is how to build it on a self-hosted LoRaWAN and Prometheus stack.

New research traces prompt injection to a single mechanism: LLMs decide who is speaking from writing style, not from role tags. Forged reasoning takes attacks from near-zero to 60% success. Remove the style, and it collapses to 10%.

One anonymous source says $12.9 billion. Both companies are silent. Here is what is actually known, what the communities fear and hope, and the questions worth asking before anyone predicts anything.

Benchmarks for Qwen's new hybrid-attention MoE at its native 262K context on a 128GB M5 Max: the architecture that makes it fit, the day-0 recipe, and a depth sweep from 0 to 262K tokens.

A late-1990s DSC PowerSeries panel already knows every door and window in the building. Two resistor dividers and a XIAO ESP32-S3 read all of it into Home Assistant and Grafana, read only.

A newer, well-recommended coding model fits my 32GB GPU at max quality and runs faster than the one I keep. I kept the old one anyway, and the numbers say I was right to.

Qwen3.6-27B scored higher on BFCL. Both models scored 58/59 on the real workload. One answers in 0.87s, the other in 4.27s.

A ~31B Mixture-of-Experts model at 4-bit, on one 32GB card, that beat a dense 32B on the real BFCL function-calling benchmark. What it is, what it is good for, and the numbers.

Every core LLM idea, defined in order and built on the one before it: weights, tensors, attention, quantization, LoRA.

signal-cli from scratch on an ARM64 Pi, an agent runtime built from source for the Signal channel, and a read-only security model enforced in layers. Everything needed to rebuild it, including every trap.

A cheap cloud-locked AI gadget, reflashed to answer from your own LLM. No cloud, no subscription.

One small computer, a few watts, and a 290-source offline library that keeps working when the internet is off, and comes with you when you travel.