The lenses
Lenses are what a reading looks at. The file is kept by hand and edited freely; the job reads it at the start of every run and never writes to it. They are not sources — every run searches fresh and the project keeps no feed list, because a cached list of feeds becomes a registry to maintain and a feed to be trapped in. A lens is a question to go and answer, and a run that finds nothing under one is expected to say so rather than pad it.
The standing lenses
- Capabilities. What can models or agents do now that they couldn't at the last reading? Releases, benchmarks that moved, capabilities claimed vs. demonstrated.
- Safety and alignment. Incidents, research results, red-team findings, model behavior in the wild — the real record, not the discourse about it.
- Work and the economy. Actual adoption, actual displacement, actual productivity signals — hiring data and shipped deployments outrank surveys and predictions.
- Compute and infrastructure. Chips, datacenters, energy, capex. The physical substrate tells you what's real.
- Quantum. What changed in what is computable or what is secure — not quantum news for its own sake. The operational test: did a result change what is computable or what is secure, or did AI change what quantum can do? A cryptographically relevant quantum computer is a security cliff, not an infrastructure story, and the two kinds of finding are read differently. Honest causal direction: the credible, shipping work today is AI helping quantum — ML-assisted error-correction decoding, calibration, materials search — while quantum accelerating AI training or inference remains speculative with no deployed instance. Say which direction a finding runs, or the lens collects hype both ways.
- Policy and regulation. What governments did, not what they said they might do.
- The public square. How ordinary people are meeting AI right now — fear, delight, indifference, lawsuits.
- AI as accelerant. Any field where AI made something go faster or further than it could before — quantum error correction, drug discovery, materials, theorem proving, chip design — and equally the harmful ones: warfare, campaigning, mass manipulation, surveillance. It counts only if the source shows AI was actually used in the work — a named model or method in a methods section, a technical account, or documented forensics — and says what became possible that wasn't. "AI will revolutionise X" is not a finding; neither is "experts warn AI could sway elections". This is the most hype-polluted ground on the board, and the test is what makes it worth reading. Sources: methods sections, lab technical reports, the host field's own competitions, preprints, platform transparency reports.
- Concentration. Who holds the capability and on what terms — open weights against closed APIs, price and access, who can run a frontier model rather than rent one, company rivalry, national rivalry. The race is a mechanism and not only a fact: competitive pressure compresses safety timelines, which is where this lens touches safety. It is also the axis between the two poles this project is named for, since the Culture and the Sprawl are the same capability under different ownership.
- Robotics and embodiment. AI acting physically — warehouses, vehicles, humanoids, surgical and agricultural machines. What was demonstrated under controlled conditions and what is actually deployed and working are different findings; say which.
The discipline
- For each lens: what happened, dated, with the strongest source found. Skip a lens honestly ("nothing notable") rather than padding it.
- Before placing the needle, write both cases: the strongest honest AItopia case from what has happened since the last reading and the strongest honest AImageddon case from the same evidence. The needle goes where the evidence, not the vibe, points.
- Prefer primary sources and shipped things over commentary. One good primary beats five takes.
- Record the magnitude, never the consequence. "Committed $X billion, financed this way, depreciated over Y years" is a finding. "This could crash the economy" is a forecast, and it belongs in the prose where it is labelled as judgment. The rule bites hardest on the AImageddon side, where a case can be built out of projections rather than filings — and without it the pessimistic reading wins by being the more speculative one, which the record would then mistake for evidence.
- The quiet places need naming, or they read as empty. The results that matter most in mathematics and security never become news: proofs land on arXiv and in Lean commits, vulnerabilities land in CVE feeds, advisories and bounty disclosures. Search a lens's own primary sources before commentary. Quiet findings are disproportionately real precisely because they are not press releases.
- A person's shipped work deposits; a person's forecast does not. Practitioners' own accounts of what they used and what it did — a working mathematician documenting a Lean formalisation and revising it as the models change, a security researcher's writeup of a found vulnerability — are primary sources and deposit under the normal rules. A principal's claims — Musk, Altman, Hassabis, Amodei, LeCun: forecasts, positioning, promises — are canon entries of kind prediction, dated and graded against what happened, never deposits. Without the line, the ledger fills with celebrity opinion, which is abundant, loud, and free, and drowns the filings and papers that are the actual record.