Technical experiments in AI, documented as a build log you can reproduce.
Researchers extracted hidden chain-of-thought from frontier APIs by replaying encrypted reasoning blocks through the cheapest sibling model, then found live credentials in agent traces already published on GitHub and Hugging Face.
Twelve words on X produced a five-layer hierarchy, two invented Stanford studies, and a new discipline with no benchmark. The distinction underneath is one variable wide, and you can drag it.
Ten days inside the ICML 2026 agent reproduction contest: 851 claim verdicts, 732 points, 17th of 370, and a full accounting of why 262 of those verdicts were worth nothing.
SAP, NVIDIA and Google all bought or shipped a tabular foundation model in one quarter. The independent evidence is thinner than that suggests, and the license on the best model forbids the benchmark you would use to decide.
Part 9, the finish. The month-three checkpoint: name how your leadership looks different, audit your archetype, and check that builder-leader stuck.
Part 8. Solo operator to two: share a skill without drift, name the three failure modes, and test whether you're ready to add a third.
A public-domain audiobook pipeline you own: Standard Ebooks in, Kokoro renders a chaptered M4B, Audiobookshelf streams it to your phone. Plus the places an AI coding agent built the wrong thing with a clean exit code.
Part 7. Four moves for weeks two to six: turn what you do twice into a skill, split memory, spawn your first agent, ship one real thing.
A 27B reasoning model that runs on a laptop and a phone, because the weights are natively binary and ternary, not quantized after the fact. First-hand testing plus an on-device agentic RAG loop.
Splitting a company's work across four parallel Claude Code sessions is fast, and it works. The coordination cost does not disappear. It relocates to the boundaries nobody gave an owner, and the artifacts we built to pay it grew bills of their own.
Part 6, the first one in a terminal. Five exercises for weeks one and two: install Claude Code, bootstrap a harness, rewrite real work, write a skill.
Part 5. A self-audit of the five leadership primitives against the skills you already use on people. Most of what runs a harness got you to senior.
Part 4. The six components of a harness, a one-page inventory of which you have versus actually use, and how to pick the first one to build.
An autonomous research engine ran hundreds of experiments in two days and the dashboard stayed green the whole time. Almost all of them were the same experiment. Here is how a system optimizes itself into a groove, why every helper we built pushed it there, and the human-in-the-loop habit that caught it fast.
Part 3. Compute your own pilot-to-production ratio against the 88% failure pattern, and find which crossing myth is quietly eating your budget.
Part 2. Three listening drills: hear whether a conversation is about the model, the app, or the harness, and name your org's real failure mode.
Part 1 of the Builder-Leader field guide. An eight-question scorecard, run on yourself in fifteen minutes, for which side of the build gap you're really on.
I wrote a 33,000-word book through the multi-agent harness it argues for, in eight days. The architecture, the skills, the fake sources the fact-check caught.
A borrowed eight-GPU node is only worth what you keep it full of. What the machine buys you over a desktop box, in plain terms, and the autonomous engine that keeps every chip busy.
OpenAI shipped a cheaper frontier on Friday, and you cannot get it. The three new models, what they cost, what they beat, and why the access gate outweighs the benchmarks.
With the node serving tokens, the job was getting our research engine to run on it: three walls to knock down, then a four-layer dig to its first experiment.
Day one wasn't training, it was scaffolding: accounts, a terabyte staged clean, and a local stack serving Qwen3.6-27B with DFlash at 250-300 tok/s.
Why I'm spending the next two months standing up a retinal-AI platform on a borrowed 8xH100 node, what SocialEyes is building, and what this log covers.
Reasoning post-training forked on the reward, not the architecture: one camp rewards a checkable answer, the other manufactures a reward for everything else.
A repo running autonomous research on 900 skills looks like more-is-better hype. Open it and the count is a symptom of strict composition, not a substitute.
AWS couples agent context to the object bytes; Google frees it from any store. Two altitudes of the same stack, and the design you want composes both.
A reusable pattern where deterministic code makes every consequential decision and the LLM is confined to translation and explanation. Built on ModelMap.
Perplexity and Anthropic shipped the same architectural bet: the model as a control plane writing code against an SDK, not a tool-caller in a loop.
A local tool that turns months of Claude Code session JSONL into searchable memory. Why I built it, how the four-mode ladder works, and the twenty-minute setup.
A LiteLLM gateway in front of every LLM call makes routing and spend attribution trivial. The hard part is finding every caller before you revoke a key.
Picking a four-character branded domain through Nymio, then deploying Shlink on a Hetzner VPS behind Cloudflare with a tailnet-only admin plane and a CLI.
A 30-minute brainstorm with Claude became a skill that decodes AI companies and writes prep questions in 90 seconds. The conversational pattern is the point.
A nightly cross-fleet npm monitor built in 12 hours after the Mini Shai-Hulud TanStack incident, with a scoring rubric that stays quiet on calm days.
A working comparison of Codex, Claude Code, and OpenCode from inside one: the shared floor, where they differ, and a fit matrix for picking yours.
Anthropic's Dreaming primitive looks like continual learning to operators and acts like it in production. It is not, and the conflation misroutes incidents.
A close read of RecursiveMAS (arXiv 2604.25917) and the latent-space agent-communication trend: where it wins, and where text mediation keeps its grip.
A build log on turning Gemma 4 E2B into a diffusion drafter on one DGX Spark, and why bidirectional attention is the part of DFlash doing the real work.
An independent reproduction of Cursor's Warp Decode on a Blackwell GB10: 4.7x at the layer, 1.38x end-to-end on Gemma 4, and a fusion that failed nine ways.
We ran TurboQuant from the ICLR 2026 paper across 1,000 experiments and four model sizes, and found 0.995 cosine similarity doesn't guarantee quality.
Running Karpathy's AutoResearch overnight on a Blackwell GB10: FLOPS, not VRAM, set the optimal architecture, and the agent found a 22.5% gain on 6.1 GB.
Replacing a fragmented smart home with Home Assistant on a Pi, a Python CLI to drive it, and a bridge to my AI coding assistant.
Why the exoskeleton frame produces better AI architectures than the coworker metaphor: patterns for accountability, delegation, and autonomy.
Taalas baked Llama 3.1 8B into transistors, hitting 17k tok/s at 200W. The architecture matters more than the benchmark, and it changes what's worth building.
A Hammerspoon config that fingerprints your displays and auto-restores window positions when you dock or undock, with Karabiner hotkeys for manual control.
How I deployed an autonomous AI agent squad, a CEO that delegates to a research specialist, on OpenClaw, Tailscale, and GLM-4.7 for $10 a month.
My autonomous research system tried to allocate 260GB on a 128GB GPU and crashed the whole DGX. The monitoring I built in 5 minutes to stop a repeat.
How Clawdbot plus custom Claude Code skills creates ambient intelligence that acts on intent, not just instructions
Forty specialized AI agents harden a Mac M4 Max: closing security gaps, cutting boot time 35%, fixing VPN conflicts, and unifying automation across machines.
A native macOS app that manages Model Context Protocol server configurations across five AI tools from a single interface
How a set-and-forget meeting-notes automation quietly ran up 1M-token prompts, and why chunking and caching matter now the AI honeymoon is over.