A daily practice for noticing small, checkable events that hint at where technology, work and society are heading before the change is obvious, and for keeping honest score of how well that works.
"Talent hits a target no one else can hit; Genius hits a target no one else can see." — Arthur Schopenhauer
For anyone. Five minutes, no jargon.
Most news tells you what has already happened to lots of people. Signals looks for the opposite: small, odd, checkable events that few people have noticed yet but that could matter a great deal later. A county quietly rewriting its zoning rules for data centers. A research result that contradicts the hype. A one-person software project doing something the big companies said wasn't possible.
Each morning the practice gathers these from many corners of the internet, checks every one against its original source, and publishes the handful that survive as a short daily report. Over time it connects the dots into longer-running questions, writes down specific predictions, and checks later whether it was right. It is a personal project by Roger Meike, with Claude, Anthropic's AI model, doing most of the reading, checking and writing.
A typical morning narrows like this. The numbers change every day, and on a quiet day the report is simply shorter. "Three solid signals beat ten shaky ones" is a rule, not a slogan.
Roughly: a couple of dozen leads, a quarter to a third survive checking, three to five are headlined.
The dashboard packs a lot into a few tabs. Here is one good daily route through it, the one Roger uses each morning. Each "Go" button opens that exact spot.
The dashboard's five tabs, and where each stop of the tour lives.
Each line is one scenario's probability over time. A rising or falling line means the evidence is pushing it; a widening shaded band means recent days are unsettling the estimate. Start here to see what changed.
Start with Today's Synthesis (what connects the day), then Mainstream Context (the big news and what it will cause), then the headlined signals that caught your eye.
Every dot is an event from the last 30 days, placed by how much evidence it carries against how much coverage it got. Today's events are numbered. The top-left corner, labelled Contrarian (lots of evidence, little coverage), is where weak signals live; the bottom-right, Hype, is the opposite.
A small green square next to a tracking signal means it received evidence today. Open one to see the new entry and how it sits against everything logged before.
Futures holds the scenario radar, the evidence explorer (which evidence moved which scenario) and the predictions ledger. Simulation shows where the economic model thinks today's evidence leads over the next 25 years.
Longer reading lives under Deep Dives, linked from the dashboard header.
A signal is something small that, if it spreads, would change how things work. The test is not "is this important?" but "is this early, is it real, and would most people be surprised by where it leads?"
A major AI lab releases a new model. Every tech site covers it the same day.
Kept as backdrop. The practice still asks what it will cause that nobody is discussing yet.
A rural county drafts rules requiring 250-foot setbacks around data-center electrical substations.
Tiny on its own. One of many local moves that, over months, added up to a fight in Congress over who pays for data centers' power.
Single events are easy to over-read. So the practice builds upward in three steps, and each step asks for more evidence than the one below it.
A few concrete "change stories" (focus futures) worked on every week, plus six broad scenarios that bound what could happen. The evidence shows which way things are leaning. It is never a single forecast.
And a simulation: the latest readings on work and AI capability are plugged into an economic model, which plays out the next 25 years (jobs, wages, incomes, who owns the gains) under each explanation of how automation works, weighted by how well today's evidence supports it.
About fifty measurable questions, each with a clear line that would count as "this has really happened." Evidence for and against piles up in a dated log.
The individual, verified events in each morning's report. Mostly noise, sometimes revealing.
Spotting things early only matters if you can show you did. Three habits keep the practice honest.
Four outside AI systems (ChatGPT, Perplexity, Grok and Gemini) work as AI agents. Each runs overnight in its own app with its own assignment. They suggest leads; nothing they say is trusted until it has been checked against the original source.
Claude does the rest each morning. It scans small corners of the web itself, checks every lead, writes the report, updates the logs and the numbers, and runs the weekly review. An automated "conductor" on Roger's computer starts each step, confirms it actually produced what it should, and retries when something fails.
Roger sets the direction: which questions to follow, which predictions to make, when to retire or rewrite something. Decisions that need him wait in a queue, each with a cautious default that applies if he doesn't answer.
One level deeper: the morning run, the rules, and the machinery behind the dashboard.
The daily cycle runs unattended on Roger's Mac. A small program, the orchestrator, runs each stage in order. A stage counts as done only when the files it was supposed to produce exist and look right; the AI's own word that it finished isn't enough. Failed stages are retried, a second pass at 8 AM picks up anything left over, and if the computer was asleep or offline, a catch-up run publishes the day's report as soon as it's back.
ChatGPT, Perplexity, Grok and Gemini each run their scheduled assignment in their own apps.
The agents' overnight output and a set of mainstream news pages are saved to disk, so every later step works from files.
Claude searches places the agents rarely look (§09) and verifies what it finds.
Claude reads the agents' leads and Roger's own tips, checks every link and date, separates weak signals from mainstream news, and writes a ledger.
Claude writes the report, adds evidence to the tracking logs, scores each piece of evidence for the probability engine, rebuilds the dashboard data and runs a consistency check. If the check fails, nothing is published.
The site is updated, and the live page is checked to confirm today's report is really there.
After the report is live: the weekly review, the simulation update and the focus-futures session (§14). None of these can hold up the daily report.
A missed day is acceptable. The method says that if there is nothing solid to report, report less rather than reach.
The AI agents are useful but correlated: they read overlapping parts of the web and drift toward what is already popular. So each has a narrow job only it can do. Since September 2026 their assignments also rotate by kind of change (money, rules, work, infrastructure, science, culture, wildcard) rather than by technology, ask that at least a third of leads come from outside AI software, flag any figure that comes from the party making the claim, and state what most people currently assume, so that "nobody's talking about this" is measured rather than assumed. The agents find; they don't interpret. What a lead means is Claude's call, argued from the primary evidence.
Cybersecurity incidents are currently out of scope for the daily report.
A tracking signal is a question with a finish line, for example whether running very large AI models on ordinary computers becomes routine, measured by a specific speed on a specific memory budget. Each belongs to a broader theme (a meta-signal) and keeps a dated evidence log. Every entry is marked confirming, disconfirming or complicating, and a complicating entry must say which way it cuts. When entries conflict, the log says whether the conflict is real (the same thing measured two ways with different answers) or only apparent (different populations, periods or definitions). Thresholds measure outcomes, such as how many people actually use something, rather than activity, such as how many announcements were made.
| Tier | What it is | Examples |
|---|---|---|
| Tier 1 | The thing itself, or a direct measurement | Peer-reviewed paper, benchmark result, regulatory filing, shipped product, enacted law, first-hand experience |
| Tier 2 | The originator's own account | Company blog, press release, preprint, a conversation with the person who did it |
| Tier 3 | Someone else's report | News article, analyst note, social post, hearsay |
Lots of Tier 3 talk with little Tier 1 behind it is flagged as possible hype. Quiet, steady Tier 1 evidence is the opposite, and it's what the practice hopes to catch.
Six scenarios bound the space of possible futures: The Bazaar (a thousand workshops, open doors), The Citadels (walled gardens, rented intelligence), The Symbiosis (where does the human end?), The Rewilding (choosing less, deliberately), The Divide (the enhanced and the rest) and The Cambrian (faster than anyone can track). Eight axes on the Futures tab show where the evidence currently sits, and they are reviewed weekly.
Every scored piece of evidence nudges each scenario's odds up or down. How hard it nudges depends on its quality tier, how directly it bears on the question, whether it is genuinely new or a repeat of known information, and whether the source was speaking against its own interest (a skeptic conceding progress counts for more than a booster claiming it). A map declares how each tracking signal relates to each scenario: necessary, sufficient on its own, one path among several, or incompatible.
The six probabilities are computed independently and do not add up to 100%. The scenarios overlap, and pieces of several can come true at once. The dashboard shows how each one moved over the past seven days.
The Simulation tab runs an agent-based model of an economy under six competing explanations of how automation actually substitutes for work, from "no AI wave" to a flywheel in which AI speeds up AI research. Each week the weights on those explanations are reviewed against four tracked labor and capability signals, and they move only when new primary data justifies it. Unlike the scenarios, these weights do sum to one: they say how much the evidence favors each explanation.
Grade predictions that have come due, check the stalest questions, run the bias audit with counter-searches, ask "what stopped being talked about?", re-read the axes, and re-measure head start against the news archive.
Each explanation's weight gets a dated verdict, moved or held, with the evidence named.
Work the focus futures: the week's evidence, indicator readings, a check against each one's kill criteria and one consequence nobody has written down yet. Plus a collision (what if two happen together?) and an inversion (what if one goes the other way?).
The current focus futures, chosen in September 2026, are bounded change stories that the evidence supports but popular discussion hasn't framed yet:
| Horizon | Focus future |
|---|---|
| Mid-2027 | Frontier AI models streamed from SSD: storage, not graphics memory, becomes the binding input for private AI |
| Late 2027 | The open web becomes metered: networks and platforms, not regulators, decide who may read public information |
| End of 2027 | AI computing becomes interruptible: flexibility is bought with on-site power and sold on as tiered service levels |
| Early 2028 | Agent commerce runs on card networks and "know your agent" checks; liability, not payment, is the market |
Deep dives, about two a week, take one question further than a daily report can: the state of the art, the players, how close a threshold is, and what follows if it's crossed. Opinion-divergence reports map who thinks what about a contested question (researchers, industry, regulators, practitioners, the public) and whether those views are converging or splitting, which often says more about timing than events do. When a tracked threshold is crossed, a futures wheel traces first-, second- and third-order consequences and turns them into new things to watch. Quarterly retrospectives audit what fired, what stalled, which sources earned their place and how well calibrated the practice has been.