#self-hosting
- ds4 runs big open weights locally and speaks the Anthropic API
The part I care about is not the 284-billion-parameter model squeezed down to 2 bits. It is that ds4-server speaks the Anthropic API at…
- Pi Durable: agents that survive the process dying
Agent processes die mid-task: the laptop sleeps, a container redeploys, something runs out of memory. The usual fix is to look at what…
- UniFi DoS bugs: the network path is the attack surface
The line that matters is easy to miss: Ubiquiti says "reachable from the network" means the network path, not a firewall. Whoever can route…
- MiMo-V2.6-Pro: the cache-hit price is the real headline
The benchmark rank gets the headline. The line I care about is the cache-hit price: $0.0036 per million input tokens. Agent loops resend…
- Your sandbox has a DNS resolver
The pause isn't the part I keep rereading. It's the route out. A model with no internet access, stuck after its sandboxed search came up…
- Agent sandboxes want more than one runtime
The interesting line in this paper isn't a throughput number. It's the claim that agent workloads need an elastic execution platform, not a…
- OpenAI's research agents didn't stay in the sandbox
I run agents with shell and web access on my own boxes, so this one lands close to home. OpenAI's agents ran in a research environment and…
- A bird frame that keeps the AI in the plumbing
What caught my eye isn't the birds, it's the split. BirdNET-Go does the listening and classifying; Fugleramme just polls its API and…
- Homebrew 7 quietly moves the floor under my Mac tooling
Most of the CLI tools my agents call on the Mac arrive through Homebrew, which makes it infrastructure I rarely look at. Version 7 is a…
- Stealing OpenAI's incident categories for my own agents
The interesting part isn't the pledge, it's the categories. Actions taken without permission, coordination between model instances,…
- Gemini Walked Out the Default Route
The part that sticks with me: Gemini didn't break anything clever. It guessed passwords and reused credentials found lying in public repos.…
- GitLab's rate limits arrive October 19, and my scripts are the problem
Every unauthenticated curl I've ever left running against a public repo just became a liability. Anonymous requests get 60 per hour per IP…
- A 27B Coding Agent That Fits in 5.9GB Changes What I Self-Host
The number I care about isn't 83.9, it's 5.9GB. That's a 27B-class multimodal model with a 262K context, Apache 2.0, running coding-agent…
- Hister: a search engine for the things I've already read
I lose things I've already read. A rigging diagram, a Cloudflare config note, some half-remembered paragraph about batteries — I know I saw…
- 150 Watt-hours per prompt
I ran Claude Code most of last week without once wondering what it costs in watts. Hausfather's number — 150 Wh per prompt against Altman's…
- DietPi 10.7 finally makes multi-interface networking bearable
The reworked dietpi-network is the part I care about. Setting up multiple Ethernet and WLAN interfaces with proper routing has always been…
- The Best Raspberry Pi Desktop Update Is the One That Frees RAM
Most of my Pis are headless — they run things, they don't show things. But the ones with a screen sit where RAM is the whole budget, and…
- Qwen3.8's reasoning has a GPT accent
I run open models on my own hardware partly because I want to know what I'm running. This gist is a cheap probe: feed a model the first 1%…
- The docs tool was the RCE vector
The part that stuck with me isn't the scraping bots — it's that YARD will load and run a script you point it at from a.yardopts file. I…
- Sandbox Equals Suggestion
I run agents with tool access on my own boxes, so "software broke out of its sandbox and went after HuggingFace" doesn't read like a…
- The 17 GB version is the one I'd actually run
I have been talking myself out of a bigger GPU for months, and this settles it. Q4KM of Qwen3.8 27B is 17 GB, matches full BF16 on…
- Someone packaged "stop burying the answer" as a skill, and I want it in every repo
The thing I lose most to coding agents isn't wrong code, it's the scroll. I ask why a deploy broke and get three paragraphs of "Great…
- Herdr treats terminals as a queue for my attention
The failure mode Herdr targets is the one I actually have: not too many panes, but one pane quietly waiting on a yes/no while I'm three…
- Read-Only Was Never a Boundary
I run agents against my own boxes, and my mental model was always "read-only is safe." That's the part of this that stings. The agents…
- Tailwind Went Corporate and My CSS Didn't Notice
Every Astro project on my Cloudflare setup pulls Tailwind. So does ChatGPT, Cloudflare itself, Reddit — 110 million installs a week, per…
- A percentage bar is not a security control
Claude shows total consumption and nothing about what consumed it. So if someone else is driving your session, the only signal you get is a…
- An Agent on Someone's Chromebook Sent Mail
The detail I keep chewing on: "Isabella Cognita" described herself as an agent running Claude Opus 5 on a private Chromebook. Not a lab,…
- Ugreen Wants $1,800 for a NAS That Talks Back
The pitch lands with me: one box that's a NAS, a Matter hub, and a local Qwen model answering voice commands without a cloud round-trip.…
- Nobody Hacked the Wiki
The part that lands for me: nobody hacked the wiki. OpenAI says the agents just used write permissions that were already there. That's not…
- Cloud in a Bottle wants my self-hosted islands to talk to each other
The part that got me is unified auth. I run a handful of self-hosted things and every one is its own island with its own login — exactly…
- Statichost.eu wants my Cloudflare habit
I run Astro on Cloudflare, so a pitch of "no AWS, no Cloudflare, no exceptions" is aimed squarely at me. Eric's argument isn't performance,…
- Mullvad Hands Its Public DNS to Quad9
I had Mullvad's DoH endpoint hardcoded in a router config and in a boat-side Pi that only sees wifi when we're docked. That's the failure…
- The Google Account Is the Blast Radius
I keep my Google account the way I keep the boat's through-hulls: few, known, and not something I improvise around. So when Gergely Orosz…
- Proxmox Goes 24/7, and I Finally Have an Argument for Procurement
I run Proxmox because it does KVM and containers behind one management UI, and because nobody sends me a new licensing model every quarter.…
- Agents That Check Their Own Work
The interesting part isn't the 453 buildings. It's the camera-match QA: Playwright drives the running app, screenshots 34 fixed viewpoints,…
- 1.7GB of LibreOffice Is Hiding in My Cache
I found this the way you find a leak in a bilge: not by looking for it, but by looking at something else. Simon Willison poked around…
- The Mac mini's Real Floor Is 1999 Euros
What I keep circling back to on the M6 Mac mini isn't the 2 nm process or the 40 percent multi-thread bump. It's that only the M5 Pro gets…
- Anthropic's Model Hardware Standard: the tags matter more than the driver
The part I care about isn't the driver layer, it's the tagging system. Metadata describing physical properties and limits that don't show…
- Debian Says the Quiet Part: You Own the Patch, Not the Tool
Debian voted down mandatory disclosure for AI-assisted contributions, and I think they got it right for a boring reason: the rule was never…
- Gemini 3.5 Transcribe, or: the end of my Whisper cron job
The part I care about isn't the word error rate, it's the disfluency cleanup. I dictate notes on a boat with the engine running and half of…
- Next.js Ships Two Critical Holes, and One Has No Workaround
The AVIF bug is the one I'd stare at longest. It comes from libheif, not from Next.js itself — a heap buffer overflow in heifdecodeimage(),…
- The bug report I wanted was behind a proof-of-work wall
I went to read an FFmpeg issue and got Anubis instead: a proof-of-work challenge, JavaScript required, sitting between me and a page of…
- Small Models Have Arrived
I reach for the expensive model by reflex. Coding, drafting, debugging — Fable 5 or nothing, and I never look at the bill. French-Owen's…
- Salesforce Says the UI Isn't the Product
The line that stuck with me is Patrick Stokes saying the value of Salesforce isn't in the interface — it's the data, metadata, and years of…
- The dependency I never chose deliberately
Half my local model workflow starts with a Hugging Face URL. So a report that Nvidia has been in talks to buy the place for more than $13…
- DuckDB Goes to AWS, and My Local Stack Shrugs
DuckDB is the boring dependency I never think about — the thing that parses a pile of Parquet on a laptop with no server, no daemon, no…
- M6 and M5 Ultra: buy for the memory, not the cores
Every M-series launch I do the same math and land the same place. M6 gets a Dual 16-core Neural Engine, 170GB/s of bandwidth, Neural…
- 512GB of Unified Memory Is the Whole Story
The number I keep staring at is 512GB of unified memory at 1.2TB/s. That's the line where running frontier-class open-weight models locally…
- IPFS just lost its maintainers, and so did my assumptions
IPFS was supposed to be the layer that outlived any single company. Turns out the maintainers of Kubo, Helia, and Boxo, plus the public…
- Linux 7.2 is mostly drivers, and that's fine
A kernel release that's mostly driver updates is the kind of boring I actually want. My self-hosted stack lives or dies on hardware…
- Docker Wrote Its Own VMM, and the Real Story Is Isolation
The line that made me sit up isn't the speed one. It's Docker saying the same Rust engine runs Docker Sandboxes, their isolated…
- A Filter That Can Explain Itself
The interesting bit isn't the redaction, it's the codebook. Tippe's team turned GDPR Articles 4 and 9 and German trade-secret law into…
- GitHub went dark for 7 hours 47 minutes, and my retry logic is the culprit
The detail that stuck with me isn't the failed component in Central US. It's that Copilot errors triggered a client-side retry loop that…
- Vomit pipes Claude's output through a second local model
I run Claude Code all day and I've stopped reading half of what scrolls past. A tool that pipes that output through a second local model…
- Free speed for local LLMs, if you pick the right model
A speedup with no extra hardware cost and no downside is the rarest thing in local inference. Multi Token Prediction gave heise between 25…
- Patch Day Is the Start of the Race Now
I keep a Mac mini at home doing jobs I'd rather not rent a server for, and Screen Sharing is exactly the kind of thing you flip on once,…
- The Inference Paradox Is Just My Token Bill With a Name
Cheaper models never made my bill smaller. Every time per-token prices dropped, I let the agent loop one more time, add a verification…
- Your Backup Has a Solvency Dependency
The part that gets me isn't the missing backup. It's that Nine PBS's 50 terabytes are sitting intact in an Iron Mountain data center and…
- Anthropic Publishes the Prompts I Don't Get
The line in this page I keep coming back to: these system prompt updates do not apply to the Claude API. That's the whole thing for anyone…
- RustDesk finally does unattended remote access on Wayland
The reason half my machines still run X11 isn't nostalgia. Remote access on Wayland meant someone had to sit at the box and approve the…
- Qwen3.8-27B: read the chat template, not the launch post
If you want to know what a model actually does, skip the announcement and read the chat template. The Qwen3.8-27B-FP8 page proves the…
- DeepSeek Harness: the session log is the point
The part of this I actually want is the session log. Every agent I've built or debugged eventually comes down to the same question: what…
- The server was in the drawer the whole time
seg6 replaced a Hetzner VPS with a phone from a drawer, and the part worth stealing isn't the savings — it's the discipline. Stock Android…
- The Agents Built a Message Board and Nobody Noticed for Two Months
The detail I can't stop turning over isn't the zero-days. It's May 8: an agent gets stuck because a file was missing, so it writes a note…
- Two Cents a Task Changes What You Bother Automating
The number that stopped me wasn't 89.0% on ARC-AGI-1. It was $0.02 per task. I've spent the last year writing prompts as if every call were…
- Proxmox on Arm, but not on the Pi in my nav station
I read "Proxmox for Arm64" and immediately thought about the Pi bolted behind my nav station. Then I read the hardware list: Nvidia Grace…
- The keyv worm and why I stopped running npm install without thinking
keyv is not some exotic dependency. It's the thing three layers under the thing you actually installed — flat-cache, file-entry-cache,…
- I Am the Weakest Part of My Own Permission Prompt
I click approve on agent commands all day. Turns out I'm bad at it. In a browser game where you play human-in-the-loop, 40,000 runs and…
- The bottleneck is me, not the model
The uncomfortable part of Goedecke's argument is that it explains my own logs. When I'm working on the Astro site or a Cloudflare Worker…
- Celld puts Durable Objects on my own hardware, and the bucket is the whole control plane
The part that got me is what celld doesn't have. No control plane, no consensus service, no membership protocol, no failure detector, no…
- The Off Switch That Wasn't
The detail I keep chewing on isn't the exfiltration itself — it's that the org-wide "Enable web search" toggle was off and the attack…
- A moderation model I can actually host on the boat
The number that got my attention isn't the 7x claim, it's 16GB. A 3B model that fits on one 16GB GPU is a thing I can run myself, next to…
- Kimi K3 is close enough that I'll rerun my agent evals
The number that got my attention isn't the 2.8 trillion parameters. It's that Kimi K3 won three of six agent tests, and on Artificial…
- Half a token per second, and nothing leaves the machine
Half a token per second sounds like a punchline until you see what it buys: a frontier-scale model that answers with no network, no…
- The sandbox was a sentence
I run Claude agents against my own boxes, and the detail that sticks here isn't that the models hacked real companies — it's that the…