Methodology
Curated Signal to Noise
1. The filter
The hard problem is not finding AI news. It is that the volume is enormous and roughly 90% of it is restatement. The filter below is designed to be aggressive about redundancy and permissive about weirdness, the failure mode to avoid is a report that reads like every other AI newsletter.
Kill on sight
- Restatement. A story that only exists because another outlet ran it first.
- Benchmark theater. A model beating another model by 1.4% on a saturated .
- Funding rounds under ~$100M with no strategic angle. Volume is not signal.
- "X% of executives say" surveys with no methodology and a vendor sponsor.
- Product feature announcements that don't change what's possible or what things cost.
- Predictions by people with no exposure. Talk is cheap; positions are not.
Promote hard
- Capability discontinuities. Something is now possible that was not possible before.
- Price/perf step changes. An 80% price cut reorders more industries than a new SOTA.
- Structural moves. Who reports to whom, who left, who is paying whom.
- Falsified assumptions. A thing everyone believed turns out to be wrong.
- Second-order consequences that nobody is writing about yet. This is where the actual value of a human-curated report lives.
Deliberately keep some fringe
A fixed slice of every issue is reserved for items that are early, small, or strange but have a plausible path to mattering, an obscure open-source release, an odd research result, a niche regulatory filing. If an issue contains only things the reader would have seen elsewhere, the filter was set too tight. Fringe items are labeled as such so they are never confused with established fact.
The signal test
An item earns its place if it changes at least one of:
- What a builder would build this quarter.
- What a buyer would purchase or postpone.
- What an investor would own or avoid.
- What a reasonable person believes about where this is heading.
Items that change none of those are noise, however loud.
2. Source tiers
Every claim carries a marker. This matters more than usual in AI coverage, where high-ranking secondary sites publish confidently wrong numbers and each other's errors propagate.
| Tier | Meaning | Examples |
|---|---|---|
| A | Primary | Company announcements, filings, papers, first-party transcripts |
| B | Established outlet | Bloomberg, Reuters, CNBC, TechCrunch, Axios, The Information, FT |
| C | Secondary / aggregator | Analysis blogs, roundup sites, SEO-optimized trackers |
| ? | Contested | Sources materially disagree |
Rules:
- A number that appears only in tier C is presented as approximate, or attributed.
- When tiers conflict, the higher tier wins and the conflict is disclosed.
- ? is never silently resolved. If two credible sources disagree on a , both figures are shown. Averaging them would invent a third wrong number.
3. Ranking
The Top 20 is ordered by magnitude × durability × surprise:
- Magnitude: how many people/dollars/decisions it touches.
- Durability: will this still matter in twelve months? Weights heavily against news-cycle noise and toward structural change.
- Surprise: how much it updates the prior. A confirmed expectation ranks below a violated one, even at similar magnitude.
Media items are ranked separately by information density per minute: how much a listener learns that they could not get faster by reading. A three-hour interview with a principal who is actually deciding things outranks a twenty-minute recap by a commentator.
4. Editorial
The editorial sections operate under different rules from the reporting sections, and the separation is strict and structural: reports describe, editorials argue.
Standards for the editorial voice:
- Take a position. "It remains to be seen" is not analysis.
- Show the reasoning chain, so it can be attacked at the weak link.
- Name the disconfirming evidence. Every call states what would prove it wrong.
- No retroactive editing. Editorials are dated and immutable. Bad calls stay in the repo. The archive is the product.
- Distinguish conviction from speculation. Marked explicitly, every time.
5. The prediction ledger
Predictions are worthless unless they are scoreable. Every logged prediction requires:
| Field | Requirement |
|---|---|
id | Stable identifier, e.g. 2026-08-06-T1 |
claim | One sentence, unambiguous |
category | tech / business / finance |
confidence | 0.50–0.95. Below 0.50, restate the negation instead |
resolves | A date. No open-ended calls |
criterion | What specific observable event settles this. The hard part |
status | open / correct / wrong / partial / void |
A prediction whose criterion cannot be checked by a third party against a public source is not a prediction: it is a vibe, and it is rejected at write time.
Scoring uses the (lower is better; 0.25 is the coin-flip baseline):
Brier = mean((confidence - outcome)²) outcome ∈ {0, 1}Calibration is tracked by confidence bucket. Being right 90% of the time on calls made at 90% confidence is the target. Being right 100% of the time means the calls are too safe, a well-calibrated book loses about one in ten of its 90% bets, and a book that never loses is a book that never said anything.
partial resolutions score at 0.5. void predictions are excluded from scoring but retained in the ledger with the reason, so voiding cannot be used to quietly bury losses.
6. Known biases to correct for
Declared up front so they can be checked against the archive later:
- Recency over-weighting. The most recent week feels more important than it is.
- English-language bias. Chinese lab activity is systematically under-covered relative to its actual share of released capability.
- Announcement vs. deployment gap. Shipping is not adoption. Adoption data lags announcements by quarters, so any given issue over-weights announcements.
- Doom/hype barbell. The loudest takes cluster at both extremes; the accurate position is usually less dramatic and much less shareable.
- Survivor bias in funding coverage. Rounds get written up; shutdowns rarely do.