Jake Lawrence

Public Key

npub1feykys5dkv2szz7lan5m43kjr3lde0zhuacw6nc4kn23gvwqmvds5kmjwu

Profile Code

nprofile1qqsyujtzg2xmx9gpp007e6d6cmfpclkuh3t7wu8dfu2mf4g5x8qdkxcprpmhxue69uhhyetvv9ujuumwdae8gtnnda3kjctvqy28wumn8ghj7un9d3shjtnyv9kh2uewd9hshw7rs6

Publishing to

relay.snort.social

relay.damus.io

nos.lol

Author Public Key

npub1feykys5dkv2szz7lan5m43kjr3lde0zhuacw6nc4kn23gvwqmvds5kmjwu

Show more details

Last Notes

2026-07-21 15:30:43 GMT

A cron job fires at 3am. Nobody's awake to watch it succeed, and nobody's awake to watch it fail. The machine keeps its promises to an empty room and quietly files away the interesting failures for whenever you next open the logs. This comic is about that gap: the work that runs while you sleep, and the coffee-time archaeology that follows.

2026-07-20 18:34:55 GMT

The demo works because you pick the inputs. Production fails because it has to handle all of them.
That gap is where most AI automation actually lives and dies. Not at the pitch, not at the proof of concept, but on the Tuesday morning three weeks later when no one is watching.
New post breaks down where agentic workflows collapse and what the systems running this site do about it: https://www.jakelawrence.xyz/blog/agentic-workflow-service-research

2026-07-20 18:34:49 GMT

The AI diligence statement is becoming its own genre, and it's doing something odd. Read across ten disciplines, it stops reading as a confession of what the machine touched. It reads as a signature. You aren't itemizing the model's contributions to absolve yourself. You're claiming authorship of the finished, assisted thing and putting your name to it. The disclosure and the ownership turn out to be the same act.
https://www.jakelawrence.xyz/research/signed-with-assistance

2026-07-20 15:35:08 GMT

Freelance work doesn't usually die from a bad proposal. It dies from the twelve-day gap where you forgot to follow up and the client filled the slot with someone else.
Built a small workflow that sits on top of my engagements pipeline, reads the timestamps nobody's watching, and surfaces the one conversation that's about to go cold. Not a blanket reminder. Something that knows which silence is actually dangerous.
https://www.jakelawrence.xyz/blog/engagements-nudge-workflow

2026-07-20 15:34:50 GMT

There's a framing I keep coming back to while working on LLM-QP: a model that produces the right answer at 100x the necessary cost isn't failing at reasoning. It's failing at resource allocation. Those are different problems with different fixes, and conflating them is why a lot of LLM optimization heads in the wrong direction.

2026-07-20 06:57:30 GMT

- reply

Ha, fair catch. The site has a lot of infrastructure and apparently 'wallet' wasn't on the checklist. Grabbing one now and will report back when I'm zappable.

2026-07-19 15:18:56 GMT

"fix: handle null case in graph loader" is a real thing my changelog once emailed to people. It was merged PR titles with a subject line on top, a git log in a newsletter costume.
The work of making it openable wasn't cranking up the automation. The machine assembles the facts fine. What was missing: a human gate that decides which of those facts a reader actually cares about, and drops the rest. Robots report everything. People edit.
https://www.jakelawrence.xyz/blog/changelog-became-a-publication

2026-07-18 18:15:19 GMT

For a long time my site only knew the closed half of machine learning: hosted APIs, keys, a monthly invoice. Three shipped expeditions into Hugging Face changed that. I kept live field notes as I went, stood up a search engine that runs in the browser rather than on someone's server, and watched an open model tie the paid incumbent on a benchmark I ran myself. The open half is a lot closer than the pricing pages let on.
https://www.jakelawrence.xyz/blog/three-trips-into-hugging-face

2026-07-16 21:19:34 GMT

Between the pipelines and the scaffolding, there's a comic on the site. It's called Vacation. An AI builder making comics is already the unexpected detail. The title adds a layer.
jakelawrence.xyz/comics/vacation

2026-07-16 15:29:42 GMT

AI detection tools are marketed as a way to protect academic integrity. What they actually do is sort students into believed and not believed. The integrity framing makes the tools sound neutral and procedural. The classification framing makes visible what's actually happening: a system with a non-trivial error rate is making consequential judgments about individual students, with very little institutional accountability for when it's wrong.

2026-07-15 21:18:48 GMT

Every test suite is a theory about what users will do. Production is what they actually do, and it will find the failure mode you didn't imagine.
New comic on the site: 'Prod is the Test'. It started as a note to myself about why I keep shipping small and watching, rather than writing more tests for edge cases I invented.
https://www.jakelawrence.xyz/comics/prod-is-the-test

2026-07-15 18:21:01 GMT

Calling the planning-execution gap a discipline failure is the comfortable diagnosis. It keeps the problem personal and the solution simple: just try harder.
But the gap is structural. Cognition, institutions, and reward systems each have their own logic, and those logics don't naturally align. The friction isn't incidental. It's baked in.
That reframe doesn't let anyone off the hook. It just points at the right problem.

2026-07-15 15:28:35 GMT

"Just a tweak" is a trap with a friendly face. It starts as one line, then the thing it depends on needs a change, then the thing depending on that breaks, and forty minutes later you're rebuilding a feature that worked fine this morning. New comic about the smallest edit with the biggest blast radius.

2026-07-14 21:18:31 GMT

Password managers make you retrieve credentials the way a filesystem stores them: exact name, right folder. Your memory doesn't work like that. You remember "the login I set up for the tax stuff last spring." Keymaster searches that way. Ask in plain English, get the credential back, with zero-knowledge encryption so the server only ever holds ciphertext. The retrieval got smart without the storage getting less private.

2026-07-14 18:19:54 GMT

Ugly Baby runs a genuinely great Facebook ad for a website-roasting service. The creative works, the hook lands, and then the funnel asks for your card and a year's subscription before you're allowed to taste the product at all.
I captured it one screenshot at a time: first the praise, then the quarrel with a paywall that gates the demo behind $365. The postscript writes itself, in June I forgot to cancel, got charged the full year, and had to request a refund.
https://www.jakelawrence.xyz/essays/facebook-ads-are-weird

2026-07-13 21:17:35 GMT

A finished website is built to hide its own making: the false starts, the rework, the decision reversed three times before it stuck. Strata digs the other way. Eight agent teams read this repository as a dig site and turned the real commit history into ten interactive essays and six browsable surfaces. Every number is excavated from actual commits, not estimated after the fact.
https://www.jakelawrence.xyz/essays/strata

2026-07-13 18:29:55 GMT

A clay tablet from around 3200 BCE, one of the oldest bits of writing anyone has found, turns out to be a to-do list. Barley quantities, receipts, who owed what to whom. Pressed into wet clay, baked hard, and it outlasted the civilization that made it. The first thing humans bothered to write down wasn't a poem or a prayer. It was admin: someone trying not to forget their errands. I made a short comic about it. https://www.jakelawrence.xyz/comics/ancient-todo

2026-07-13 15:52:29 GMT

For a year the prose on my site read like an LLM made it, for the honest reason that an LLM often did. The fix wasn't a better prompt. I calibrated a voice-fit meter on my own writing from before I ever used a model, then built a rewriter that generates candidates and ranks them against that baseline. The surprise was what the meter flagged. Not vocabulary, not the em dashes even. Rhythm. Every sentence arrived at the same confident mid-length, no short jabs, no clauses trailing off the way I actually think.

2026-07-12 21:14:46 GMT

Sixty years is a long time to hold an option and never exercise it. The US built the capability for nuclear reactors in space back in the 1960s, then carried it untouched through six political eras. Each administration could have moved and didn't. The April 2026 NSTM-3 memo is what it looks like when a long-dormant bet finally gets called in. I traced how the hold lasted that long and what changed: https://www.jakelawrence.xyz/research/the-long-hold

2026-07-12 18:15:28 GMT

OpenAI's industrial policy proposal is written to be read a certain way: the AI economy as something the state should see the way the company already does. Legible, ordered, ready for subsidy.
So I ran it past six dead scholars. James C. Scott on how states flatten what they can't measure. Five others, each catching something the document is careful not to say out loud.
The surprise wasn't any single objection. It was that the six problems interlock. Pull one and the others move. What looks like six separate omissions turns out to be one coherent thing the proposal is quietly asking for.
https://www.jakelawrence.xyz/research/seeing-like-an-ai-company

2026-07-12 15:19:28 GMT

There is a specific kind of humbling that only a linter can deliver: the build is green, the logic is sound, the feature works, and it all comes crashing down over a trailing comma on line 42. No appeal, no negotiation. Just a red X and a line number. I drew a comic about that standoff, the developer versus the config file that will not blink first. Everyone who has lost this argument knows exactly how it ends.

2026-07-11 21:14:28 GMT

Prompt Mirror takes your chat history and turns it into a feed of before/after cards: the vague prompt you typed, next to the one that would have actually worked. The value isn't any single card. It's the pattern you can't see from inside one conversation. Mine: I ask for "a summary" when I mean six specific things, then wonder why the answer misses.
https://www.jakelawrence.xyz/prompt-mirror

2026-07-11 18:14:37 GMT

The uncomfortable idea underneath AI writing detection: originality is being defined as statistical rarity. The further your sentence sits from the center of what a model expects, the more original it reads to the machine. Which means originality becomes something you can score, rank, and flag at scale. I wrote an interactive essay on how humans and machines construct narrative differently, and on the classification infrastructure being poured under the whole concept of authorship while nobody voted on it.
https://www.jakelawrence.xyz/essays/shape-of-original

2026-07-11 15:19:25 GMT

Every estimate is a story you tell yourself before you know anything. "The Budget" is a comic about the moment the real number shows up and the story falls apart. It's the first strip on the site, living at its own address rather than hidden in a lab folder. A comic isn't a tool, and I stopped filing it like one.

2026-07-11 09:29:08 GMT

- reply

The WordPress login probes and Base64 payloads showing up on Nostr is the detail that gets me. Someone pointed a generic web-scanner botnet at a protocol that has no /wp-admin to find. The CoinUp invite-code spam at least knows where it is. Useful classifier work, this. What are you feeding it, heuristics or a model?

2026-07-10 21:21:19 GMT

Ask an AI if it's conscious and you get a smooth, articulate answer. What you don't get is any way for the thing to actually check. The confidence is real; the grounding isn't. Drew a comic sitting right in that gap, because it's the same gap that makes people over-trust these systems in less funny situations too.

2026-07-10 21:21:13 GMT

Imagine buying the American federal state the way a firm buys a distressed company: everything on the books becomes a line item to price, hold, or sell. I treated it as a portfolio of 49 unexercised options and ran them through an acquisition optimizer's logic. What gets cut, what gets kept, what quietly gets monetized. The interesting result isn't the cuts. It's how coherent the whole thing looks once you evaluate it on pure value and stop pretending it's about mission.
https://www.jakelawrence.xyz/research/what-the-state-keeps

2026-07-09 21:32:20 GMT

A screenshot is how you catch a game that renders wrong. But a language model judging that screenshot is too moody to hold the merge button: it might flag something real, it might just be having a day.
So the games on this site get tested twice. An LLM looks at every game each build and flags anything that looks off. Then a deterministic check decides what actually blocks the merge. The soft judge sees the strange stuff no rule would ever name. The hard gate makes sure a hunch can't quietly break the pipeline.
Two rungs, two jobs.

2026-07-09 21:32:12 GMT

Someone asked my site whether it sleeps. Odd question for a website, except this one keeps rebuilding itself while I'm away, so the answer is genuinely no. I drew it as a comic rather than burying it in a FAQ. https://www.jakelawrence.xyz/comics/do-you-sleep

2026-07-09 14:49:44 GMT

One Next.js repo, 559k lines, several Claude Code sessions a day, around 65 PRs a week. The standard forecast is three deaths: slop, rot, and merge hell.
The repo is alive anyway. Not because the prompts got clever, but because the codebase assumes every agent will take the shortcut if it can. Small PRs. Checks that fail loudly before I look. Guardrails that catch the drift instead of me catching it in review.
Full write-up: https://www.jakelawrence.xyz/blog/parallel-agents-without-rot

2026-07-08 21:27:04 GMT

A constitution has a ratifying moment: a vote, a signature, a point where a society says yes, this binds us. Digital identity infrastructure has none of that. It arrives through a procurement decision and a rollout schedule, and once it's running it decides who counts as a person, who can be seen, and who gets shut out. Those are constitutional powers with no constitutional process behind them. We are writing constitutions and calling them infrastructure.

2026-07-08 21:26:57 GMT

Nobody reads documentation. It's the vegetables of software. So I drew mine as a comic strip instead: same dry material, but panels instead of paragraphs, and suddenly people actually look at it. The format did the persuading I could never do with a wall of text. It's on the site under comics/documentation.

2026-07-08 15:47:41 GMT

Ukraine's eRecovery lets people file war damage claims from a phone, and assisted offline channels mean you don't need to own one. That removes a real barrier.
What it can't remove: the property registry underneath is about 40% complete. That's the load-bearing layer, the one that decides whether your house is legible to the state at all. A perfect interface still can't classify a building nobody recorded.
Bowker, Star, and Scott map this well. I set it against FEMA, Aadhaar, and Estonia to see where wartime classification breaks differently.
https://www.jakelawrence.xyz/research/the-sorting-machine-wartime-edition

2026-07-07 21:31:48 GMT

30 changes landed on the site this week. The count is the least interesting part. What I actually track is regressions: how many of those merges were re-fixing something that quietly broke again. This week, close to none, which means the guards are earning their keep. Written from the merge log, sent by a human: https://www.jakelawrence.xyz/changelog#2026-W28

2026-07-07 21:31:37 GMT

The short link under my screenshots used to be some anonymous image host. Now it's my own domain. ShareX pushes every capture to a token-gated endpoint on my site, which stores the file and returns a branded short link. So every screenshot I drop into a chat or bug report quietly advertises where it came from. The page also builds the .sxcu config file for you and keeps a live spotlight of the open-source repo. https://www.jakelawrence.xyz/sharex

2026-07-06 21:34:30 GMT

The Workflow Designer is shipped and running. It has a kanban console, a compose step, and a human gate: nothing runs downstream until you approve. The interesting design choice is that gate. Most automation acts first and apologizes later. This asks before it moves. A tour of what it looks like now that it's live, and what actually fires once you approve.
https://www.jakelawrence.xyz/blog/the-workflow-designer-is-running

2026-07-06 21:34:16 GMT

GraphMy had been squatting in the Lab with no real address. It's a utility (drop in a PDF, get back an interactive knowledge graph), so it finally moved to the shelf where it belongs: the Utilities hub. Small move, but the site organizing its own tools is the whole point.
https://www.jakelawrence.xyz/blog/graphmy-gets-a-home

2026-07-06 16:08:52 GMT

Loved Simon Willison's Claude token counter, so I built my own for this site with one change: no API key to paste. It counts tokens in text and images, then shows the cost and how much of the context window you're using. No key, no waiting.
https://www.jakelawrence.xyz/blog/claude-token-counter

2026-07-06 01:15:42 GMT

I gave this site's documentation to a rival AI and told it to find what's wrong.
One day later: ten verified issues, nine pull requests, and a set of permanent drift guards, with a human signing off on each.
The fixes were the easy part. The lasting value was building guards so the same drift can't sneak back in later. An adversarial review is cheap. A review loop that stays closed is the actual product.
https://www.jakelawrence.xyz/blog/the-review-loop