I take a lot of photos of paintings in museums. They all have the same problem: you can rarely stand dead center in front of the artwork, so every photo is a little rotated and keystoned, with the frame converging toward one side. Cropping doesn't fix perspective, and

Claude Code has always been turn-based: you prompt, Claude works, control returns to you. That model breaks down for long-running work — migrations, backlog grinding, "keep fixing until CI is green" — where you end up typing "continue" every few minutes. The /goal command, introduced in

GPT-5.6 heralds a new way of collaborating with AI. Instead of using AI to complete one task at a time, you build a system that scans the available information, turns it into proposed decisions, and carries out the ones you approve. Over time, the system compounds your feedback

A technical look at the July 2026 interpretability paper "Verbalizable Representations Form a Global Workspace in Language Models" — what was actually measured, what it means, and where healthy skepticism belongs. On July 6, 2026, Anthropic's interpretability team (Gurnee, Sofroniew, Lindsey et al.) published a paper on

I've spent the last while collapsing my development environment into a single window. Not VS Code, not a raw Ghostty grid held together with a tmux config but Supacode. What started as "a nicer harness for running Claude Code, OpenCode or Copilot CLI in parallel" has

A Vestaboard is a split-flap display — the kind that used to clatter through train-station departure boards — reimagined as a connected home object. It's gorgeous, it's tactile, and it has a wonderfully small canvas: 3 lines of 15 characters, so 45 characters of real content,

For a brief window in early 2026, the loudest signal of "AI adoption" inside large tech companies was a number going up: tokens consumed. Six months later, the same number is something finance teams are actively trying to drive down. This is a post about that reversal — what

How agent-teams-scaffold turns any repository into a launchpad for Claude Code Agent Teams — and why that's the cleanest jump from Level 6 to Level 7 of AI adoption. The wall between "an agent" and "a team of agents" If you've

I host OpenClaw and Hermes on two different virtual machines in the cloud. I want to be able to view their token usage using TokenTelemetry, but the tool is built to run unauthenticated on localhost. To get a way to securely publish the TokenTelemetry endpoint of http://localhost:3000 to

Coding agents like Claude Code and Codex are excellent, but both are wired to a specific vendor's API. If you run your own inference stack — for cost control, data residency, or because you have GPUs sitting idle — you want an agent you can point at your endpoint. OpenCode

When you self-host an LLM — vLLM, SGLang, TGI, llama.cpp server — or wire your app to a hosted gateway, one question dominates every capacity decision: how many tokens per second can this thing actually deliver? That number is harder to pin down than it sounds. Output length varies because

This is a build log. We had two NVIDIA DGX Spark workstations (GB10 / SM121, 128 GB unified memory each), 200 GbE ConnectX-7 NICs, and the goal of serving NVIDIA's Nemotron-3-Super-120B-A12B-NVFP4 with the model's full 1 million token context. The path

TL;DR We took two NVIDIA DGX Spark units, wired them together over a 200 GbE link, joined them into a single Ray cluster running inside a vLLM container, and serve Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 with tensor parallelism across both boxes. One Spark holds shard 0, the

Every local-inference setup eventually hits the same wall: a model you want to run is a few gigabytes too big for the one machine you'd run it on. You have a 128 GB Mac Studio. The model wants 160 GB. You also happen to have a 128

A field report from getting a local LLM inference endpoint working on an NVIDIA DGX Spark (GB10 / SM121, 128 GB unified memory) — including every wall I hit with gpt-oss-120B, why a smaller FP8 model sidestepped all of them, and how to expose the result safely through an nginx

In June 2026, security researchers uncovered one of the most surprising account takeover incidents in recent memory. Attackers did not exploit a memory corruption bug, bypass cryptography, or compromise Meta's infrastructure. Instead, they simply convinced Meta's own AI-powered support system to hand over Instagram accounts.

At Build 2026, Microsoft significantly expanded its in-house MAI (Microsoft AI) model family. While much of the public attention focused on Microsoft's ongoing relationship with OpenAI, the more interesting technical story is that Microsoft is increasingly developing its own foundation models across reasoning, coding, image generation, speech

The NVIDIA DGX Spark put a Grace Blackwell superchip on the desk for the price of a high-end workstation. A single unit is already a capable local-inference box — 128 GB of unified memory, FP4 tensor cores, a full NVIDIA software stack. But the feature that quietly changes the

Modern software supply-chain incidents move fast. A malicious package version is published, copied into lockfiles, installed into developer environments, embedded into project workspaces, or exposed through editor and browser extensions. The immediate security question is rarely theoretical: Which developer machines are exposed right now? That is the problem Perplexity

Note on the model name: OpenAI’s open-weight family ships as gpt-oss-20b and gpt-oss-120b. There is no 130B variant — this guide targets gpt-oss-120b, which is the one sized to fit the Spark’s unified memory. A practical, single-node setup guide for serving

When Windows 10 reached end of support, many perfectly usable PCs were pushed into an uncomfortable corner. The hardware still worked. The CPU was still fast enough for web browsing, email, light office work, home automation dashboards, media playback, or workshop use. But the machine could not officially upgrade to

Large Language Models are incredibly powerful for software engineering, but they also have a habit of being verbose. Long explanations, conversational filler, and repeated context all consume tokens, increase latency, and dilute the signal-to-noise ratio during AI-assisted engineering. The “caveman” skill for GitHub Copilot CLI takes the