Recent content on Notes from the Rabbit Hole
DeepSeek Harness is a coding-agent harness from DeepSeek. It’s labeled as a developer preview, and it puts DeepSeek’s models inside a serious, modular environment for working on code. That makes it interesting. It also puts the software awfully close to a person’s source code, credentials, shell, files, and network. I haven’t adopted it. I decided to inspect it first.…
What I’m focused on right now
I have a fascination, and sometimes an irreverence, for the way buzzwords can spread like wildfire in business. This is not a new habit. I have been making a small, happy nuisance of myself about it for years. In 2013 I started writing for the Red Hat Developer Blog about a DevOps transformation we were running inside Red Hat IT, and somewhere in that series I went on a tear about managerspeak .…
This article started with me talking to my agent over Wispr Flow . I was dictating my thoughts while it transcribed them and kept the whole conversation available as context. That’s very much what my creative flow looks like. But while I was speaking, it occurred to me that something important was happening: I wasn’t editing myself. When you type, you filter. You form the sentence,…
On April 24, 2026, DeepSeek launched the V4 preview : the Pro model a 1.6-trillion-parameter mixture-of-experts with 49 billion active per pass, one million tokens of context. The world’s most capable open model family went live on Huawei’s Ascend chips and CANN software , not CUDA. Alibaba Cloud and Tencent Cloud served it within hours . Huawei’s own chips had trained part of…
On Monday you’re a product strategist. On Tuesday you’re an architect. Wednesday is sysadmin. Thursday is marketing. Friday you run the numbers and realize you’re an accountant. Saturday you look at the week and wonder who checked any of it. The question is always the same: which of these ten jobs do I do today? The cruelty of the solo-founder job is not the difficulty. Each job…
The cheapest AI model isn’t the one with the lowest token price. It’s the one that gets acceptable work across the finish line with the fewest retries, repairs, and wasted minutes. That sounds obvious, but it is not how most of us pick models. When a model family has a cheap tier, an expensive tier, and a middle tier, the middle feels responsible. It sounds like the sensible default.…
Personal agents get useful by learning one person. A company agent has a harder job: it has to remember enough for people to work together without treating the whole company as one person with one permission set. That’s the problem QM is trying to solve. Y Combinator has open-sourced QM after learning what it takes to manage a fleet of agents at work . YC says it provisioned more than 50…
In September 2025, Anthropic had a documented Claude share-page search-indexing failure on its hands. Google had indexed just under 600 shared Claude conversations, according to Forbes , which reported the incident on September 8. These weren’t ordinary unshared chats, but public snapshots created by Claude’s Share feature. The first warning was documented in public. Then, 322 days…
I think the United States should consider building an American Stewardship Fund: a permanent public inheritance that turns some portion of a major economic windfall into capacity for people who aren’t lucky enough to own the scarce assets behind it. That’s the argument. The mechanism is not settled, and I’m not pretending I’ve got a complete federal policy blueprint tucked…
A city can put a remarkable amount of information online and still make it hard to use. You can find Raleigh police records, fire statistics, old incident summaries, inspection histories, and a public active-incident map. But each service has its own vocabulary, its own search form, and its own idea of what the result means. A resident arrives with a normal question. The city replies with a…
Walk into Breakwater Nine with me. The old ferry lanes are still painted on the floor, though the terminal hasn’t carried commuters in years. Tidal shuttles now bring medical supplies to clinics on the upper deck. Salt has worked its way into the freight lift’s access reader. A courier is standing there with a case of insulin and eleven minutes of cooling left. The credential in her…
You have a decision you can’t stop turning over. You’re not in crisis. You don’t need a therapist. But you’ve talked yourself in circles enough times that you can predict every branch of your own internal argument. What you need is someone to help you think: without steering, without diagnosing, and without pretending the decision belongs to anyone but you. That’s…
You’re standing near Fayetteville and Hargett streets downtown, and you need a bus. The useful answer isn’t a lesson about how transit information gets published. It’s this: the closest stops in this lookup were on East Hargett at Fayetteville, Wilmington at Morgan, and East Morgan at South Wilmington. You shouldn’t need a scavenger hunt through city websites to get an…
Stop telling me it’s done. Tell me what you checked, what you assumed, and what you couldn’t verify. /ponytail was always loaded. That was the problem. Not that it was bad at its job. Not that the YAGNI instinct it encoded was wrong. The problem was that it had no idea when to shut up. I’d be writing a personal email, organizing a research note, configuring a smart home device,…
Late in a long interview process, a take-home assignment landed in my inbox. I opened it, read through the prompt, and laughed. It asked me to present an initial executive point of view on helping an organization move from AI-assisted software development toward something more autonomous: how to diagnose where the organization sits today, how to sequence the next moves, what to protect, and how to…
I keep coming back to what happens after something breaks. A skills repository matters when the result of one skill changes what the next one can do. Sometimes that’s modest: a conversation about somebody’s actual routines, constraints, and voice can give a brand-design skill something better to work with than a pile of aspirational adjectives. Your city’s public data can move…
What if the most important AI development isn’t a model that knows more, but one that knows less? That is the question at the heart of Andrej Karpathy’s “cognitive core” thesis. In a tweet and a podcast interview , he laid out a vision that sounds almost heretical in an industry built on scaling: build a small model (a few billion parameters) that intentionally sacrifices…
Almost every American knows where to find an abandoned railroad track near them. Go look. There is probably one within driving distance of where you live right now. Weeds growing between the rusted rails. Ties rotting into the ground. Maybe the right of way has been paved into a walking trail. Maybe it just sits there, too much trouble to remove. Those tracks used to keep towns alive. A railroad…
Right now, somewhere a CEO is yeeting a six figure AI platform at a team and praying something sticks. That was the opening image Shannon Ryan left us with at this month’s AgileRTP meetup, and it only got sharper from there. Shannon is VP of Marketing and a technology strategist at Veritas Automata , a firm that works with teams adopting AI in highly regulated industries: life sciences,…
I wrote about the Dark Factory earlier this year, covering StrongDM’s Level 5 autonomous software factory where three engineers ship production code with no human touching the implementation. In that article I mapped the five levels of AI coding autonomy, a framework that has become a useful shorthand for where organizations sit on the spectrum from human-driven coding to full autonomy:…
This is the third article in a series about forking an open-core product with AI coding tools. Part 1 laid out the roadmap of what was gated behind the Pro paywall. Part 2 executed the fork: rebrand, telemetry strip, first feature. It proved the easy stuff was as easy as it looked. Part 3 is about the stuff that isn’t easy. Speaker diarization (identifying who said what in a meeting) is the…
Part 1 laid out the roadmap: here’s what’s gated, here’s what it would take to ungates it, and here’s a worked example of how an AI coding agent turns a spec into real code. This part is the actual fork. Not a thought experiment. Not a hypothetical. Real commands, real diffs, real output. Cloning and Branching The starting point is the Meetily v0.4.0 MIT codebase. Clone it,…
This is Part 1 of a series about using vibe coding to fork open-core projects. Not as an abstract argument, but as a walkthrough of a real fork, a real roadmap, and a real example. The Vibe Coding Argument Nobody’s Making Vibe coding gets a bad rap. The discourse is dominated by people building Yet Another Todo App with Cursor, or prompting their way into a GPT wrapper and calling it a…
If you are neurodivergent and work in a typical office, you have probably had this experience. You know what you need to do your best work. Maybe it is a quiet space, or written instructions instead of verbal ones, or a schedule that does not require you to switch contexts six times before lunch. You know these things would help. But asking for them means disclosing your diagnosis to a manager you…
Most companies are playing a game they don’t understand the rules to. They’re standing in a room full of competitors, all of them staring at the same two doors. One door says OpenAI. The other says Anthropic. They believe the strategic decision is which door to walk through. They believe this so deeply that they’re paying enormous premiums for inference that’s past the…
Anthropic just published an open-source skill for Claude Code called launch-your-agent . It takes a technical founder from “I want to build an agent that does X” to a live, scheduled, self-grading Claude Managed Agent in one conversation. The interview is iterative, not a form. The output is a real deployed agent, not a plan for one. And the first run grades itself against your own…
Google’s Open Knowledge Format and my Artifact Pyramid methodology converged on the same structural insight within weeks of each other. Here are six features from my open source project that would make OKF bundles dramatically more useful for agent pipelines. Includes a bridge proposal for combining them.
The Architecture of Focus Your calendar is a design document. You probably never read it that way. Most companies run on a schedule built for managers: back-to-back meetings, hour-long or half-hour increments. It works great for coordination and decision-making. It is actively hostile to the kind of work most organizations say they need most (sustained creative and technical output) because that…
I needed a research tool I could run on my own hardware, without per-call pricing. So I built one. And then I open sourced it so you can have it, too. The problem with per-call pricing is not the line item on your card. It is the question you do not ask because the cost is not worth it. The secondary source you accept because scraping the original is too expensive. The fragment of a page you read…
The City of Raleigh publishes 178 public datasets. Crime reports, restaurant inspections, building permits, bike lanes, speed humps, EV charging stations, dog parks: it is all there on data.raleighnc.gov, free for anyone to use. Most citizens never touch it. The barrier isn’t access. The barrier is discovery. The data lives behind an ArcGIS Hub portal, a web interface designed for GIS…
Companies are shipping AI into production with no way to tell if it’s actually working. The slop isn’t a model quality problem, it’s an evaluation problem. Here is the four-rung ladder that turns vibe checks into engineering discipline, with tools you can run today.
Paul Graham’s essays describe the ideal startup founder in terms that map almost perfectly onto Autistic and ADHD cognitive patterns. Corporate hiring processes penalize the same traits. The startup ecosystem is functioning as an inadvertent neurodiversity inclusion program.
There is a structural transformation happening in how software organizations operate. It has been described from different angles: as the software factory (repeatable delivery pipelines as industrial processes), as the AI flywheel (operational data as a continuous training signal), and as the AI-native startup playbook (eval suites as governance, tokens as headcount). But these are not three…
I had an AI agent extract and analyze all 464 articles from the Y Combinator Library last weekend. Every piece of startup advice the world’s most influential accelerator has published: 123 categories, from Becoming a Founder to Fundraising to Artificial Intelligence. The word “flywheel” appears exactly once. But the flywheel mechanics are everywhere, hiding in plain sight. YC…
The Product Manager wanted to ship an MVP. Three phases. Skip the reproducible build. Skip the caching subsystem. Skip the Mermaid diagrams. Ship fast, iterate later. The Debugger said that path would destroy maintainer trust. “Non-Deterministic Bug Blindness,” the Debugger said. “Skip the reproduction baseline and the ‘fails on main’ gate? The agent writes a fix,…
Maximize Work Not Done: The Overlooked Agile Principle Behind Nano Unicorn Success Business leaders have exactly one tool for operational expenditure, and it’s a hammer. When the pressure to cut costs comes down, the reflex is the same every time: headcount reduction. A layoff round. A hiring freeze. A restructuring that moves the same work onto fewer shoulders. Nobody thinks to cancel the…
The productivity industry wants you to believe there is a best system. There is a note-taking methodology that will fix your knowledge work, a second brain framework that will tame the chaos, a set of categories that will make everything fit. Find the right one and the torrent of information becomes manageable. It is a compelling promise, and it is wrong in a specific and instructive way. There is…
I’ve been in the room. Multiple times. Multiple companies. The security organization says no. Completely, categorically, by default. The posture isn’t “let’s figure out how to make this work safely.” It’s “this is forbidden until we’ve completed our review” where the review cycle is measured in quarters and the technology is shipping weekly. My…
You’re right to be scared. Every day there’s another headline. Another company citing AI in a layoff announcement. Another prediction that your profession is six months from obsolescence. You’re watching the news and thinking: this time it’s different. It is different. But not in the way you think. Let me show you what I mean. I grew up around adults who were building the…
Part 2 of the Clanker Kanban series. Six software architecture documentation methodologies independently discovered progressive disclosure between 1995 and 2011 without naming it. The artifact pyramid names what they all found. The Technical Architect profile in the Clanker pipeline carries this lineage.
Part 1 of the Clanker Kanban series. Why sprint ceremonies fail for AI agent coordination and what to build instead. The artifact pyramid, the orchestrator pattern, and four AI-native ceremonies that replace planning meetings with structured handoffs.
In the late 1980s, a former trial lawyer named Denise Shekerjian read a newspaper article about the MacArthur Fellowship. The “genius grant,” people called it. It came with a mysterious phone call, a generous six-figure award paid with no strings attached, and the kind of cultural recognition that changes a life. Shekerjian was not interested in the money or the prestige. She was…
Dear Gravity, I don’t know if you remember me. I sat in the back of the room in high school physics, third row from the window. I definitely fell asleep during angular momentum. I am not qualified to write this letter. But I’ve been reading about you lately, Gravity. And I think you’re trying to tell us something. Here’s what I learned this week, after many failed attempts…
I found a real bug in SearXNG’s Brave Search integration: the braveapi engine was returning HTTP 422 errors from the API even though direct httpx calls with the same parameters worked fine. I identified the root cause and submitted a three-line pull request with the fix. I disclosed in the PR that an AI agent helped draft it, as I do on all my contributions. The maintainer’s response…
There is a crisis hiding inside the multi-agent AI boom. You just cannot see it on SWE-bench. Every major benchmark measures what a single agent can do in isolation. SWE-bench tests one agent resolving GitHub issues. GAIA tests one agent completing tasks. AgentBench evaluates one LLM across environments. You can compare any model on any of these, and the numbers look great. But put two agents in a…
If you used the internet in the 1980s (and “using the internet” meant sitting at a VT100 terminal in a computer science lab), you probably used finger . You typed finger username@hostname and TCP port 79 returned a few lines of ASCII text telling you whether that person was logged in, when they last checked email, and what they had written in their .plan file. That .plan file was the…
A 35-billion-parameter model belongs in a datacenter. That’s the assumption. You need an H100, or two, or eight of them. A consumer GPU tops out at 16 GB of VRAM and you’re not fitting a 35B model in there. End of story. Except Qwen3.6-35B-A3B isn’t a normal 35B model. It’s a Mixture-of-Experts architecture: 35 billion parameters spread across 256 specialized expert…
This is the third and final part of the series. If you haven’t read Part 1 and Part 2 , the short version: a jumping spider with fewer than 100,000 neurons has a depth perception system so well understood that researchers have modeled it and proposed building a sensor based on it. Nine years later, nothing has been built. This part is about why. I need to start with something that has been…
Last time, we met Portia , a jumping spider that plans hour-long hunting routes, learns by trial and error, and manipulates mental images with fewer than 100,000 neurons. We ended on a provocation: one of Portia’s capabilities, its depth perception system, has been studied at the optical level, computationally modeled, and explicitly proposed as a sensor template. And nine years later,…