The best AI work is buried
Every week brings thousands of new papers, talks, podcasts and posts. Most repeat each other or oversell. Finding the handful that matter takes hours of reading that most people never get back.
What it does, in one minute
It does the hunting, then explains
Software reads research listings, subreddits, newsletters, developer platforms and social timelines. It groups repeat mentions, filters clickbait and grades candidates for usefulness. Each shelf offers a short list with a note explaining why a pick matters.
You can check every step
Shelf placement cannot be bought. Reader suggestions enter discovery and face the same selection checks as other candidates. Cards show their scores. New paper commentary is reviewed against its abstract; new Insights are checked against the source excerpts the writer actually read. Those checks can miss things. Inspect the originals where available and the dated audits, including criticism and withheld grades.
Read the full method
What you actually get
Shelves where every item had to earn the slot
Podcasts, papers, books, video, music and short written insights. Automated picks pass selection checks; Brendon’s pinned books are identified on their cards. Reader suggestions can enter discovery and face the same checks as other candidates.
Papers you can listen to
When narration is available, a paper starts with a spoken introduction to its claims and why it may be worth your time, followed by the text extracted from its PDF. Extraction can miss figures, tables or equations. Open the original to inspect those details; audio availability can vary by paper.
It is a synthetic voice. A good one. You will still know.
Two questions per 24 hours
The chat carries my interests and how I read this industry, so you can argue with the reasoning instead of reading a verdict. You get two questions. Your allowance resets 24 hours after your first accepted question; the chat shows when. Each answer calls a model, and those calls cost money.
It speaks for me. It is not me, and it will be confidently wrong sometimes. So am I.
Automated picks are selected without waiting for a person to approve them. Pinned books and original writing have separate paths, and mistakes can be corrected. The method below explains what the checks can establish, where evidence may be missing, and how to inspect a pick yourself.
agreement, not volume
One loud voice is not a signal
Mentions from separately registered communities can earn a pick a bonus. That is a signal to inspect, not proof of independent discovery: communities may repeat the same announcement. It is a bonus, not an entry requirement. A pick can also earn its place on quality and source standing alone, and its card shows which one carried it.
Think of restaurant recommendations. Hearing about a place from several groups can make it worth a look, but those groups might all have read the same review. The system groups registered sources and discounts repeats; it cannot establish that people never influenced each other.
The corroboration score
In words: every mention counts for something, a trusted source counts for more, and a mention loses value the older it gets. Add up what is left. That is the whole idea — the line below is just that sentence written precisely enough to run.
value(s) = authority × e−ageDays / 7 × (0.5 + 0.5 · strength)score = min(10, Σ value(s) / 2.5) over distinct groups
Calibrated so that one fresh, top-authority sighting reaches about 3.6 and three of them reach 10. A single source cannot climb into corroborated territory alone, however loudly it shouts.
- decay constant
- 7 daysa sighting is worth 37% of itself a week later — last month’s buzz is not this week’s shelf
- signal retention
- 21 daysolder sightings are pruned outright
- entity cap
- 2,000newest entities win
- repeat sightings
- never stackthe strongest one per source is kept
Where the same item wears two titles
Exact fingerprints require identical normalised text, so a formal paper title and a community’s paraphrase never matched — 13 of 1,080 signals converged. Fuzzy title matching closed that gap deterministically, at no API cost. The threshold was measured, then tightened when a test found it too loose.
Measured Jaccard overlaps: different-item pairs at 0.33 and 0.36, a scaffolding-only pair at 0.571, genuine same-item pairs at 0.75, 0.80 and 0.82. The threshold sits at 0.70.
0.70 threshold
0.00.51.0
different items
same item, two titles
the pair 0.5 would have merged
Jaccard overlap of content tokens. Genuine and false pairs separated cleanly with nothing in between; 0.5 sat inside the gap but hugged the wrong edge.
Tap or hover any point to see the pair it came from.
deliberately not transitive
A fuzzy match contributes no keys, so it can never seed another round of matching. Chained approximate matches drift, and a false merge here manufactures convergence — which is the failure this whole page exists to avoid. A containment rule was tried and rejected: it merged a model announcement with a benchmark analysis of the same release.
what actually got built
Eight stages, and a loop back to the start
Eight steps take something from “somebody mentioned it” to “it is on the shelf” — and the last step changes the first. Automated picks follow this pipeline. Pinned books and original writing have their own selection and review paths.
-
Read communities that do not talk to each other
Seven feeds are polled independently: research paper listings, a link-voting board, subreddits, named editors writing newsletters, developer platforms, and two social timelines. Each sighting is recorded with who saw it, how strongly, and when — nothing is judged yet.
-
Work out when two mentions are the same thing
The same paper might arrive as an arXiv link, a video about it, and a code repository. Each mention is reduced to a handful of identifiers —
arxiv:2601.01234,yt:…,gh:owner/repo, a tidied-up URL, a tidied-up title — so the three can be recognised as one. A strong identifier outranks a bare URL, and addresses likeyoutube.com/watchare refused outright: on their own they are the same string for every video on earth. -
Group those mentions into one entry
Anything sharing an identifier is merged. Titles that are obviously the same thing worded differently get exactly one chance to join — and a match made that way contributes no identifiers of its own, so it can never set off a further merge. Chains of approximate matches drift, and a wrong merge here would manufacture the very agreement this site exists to detect.
-
Decide what may count as agreement
One community arriving over three transports collapses to one vote. Sources declared non-independent collapse together. Publication venues and first-party announcements are shown and never counted. An undeclared source contributes zero — the registry fails closed.
-
Apply hard floors that no model can argue with
Course, listicle and clickbait title shapes are blocked before anything is scored. An unregistered source must show exceptional organic traction instead. Title and structural checks run in code. New paper commentary also needs an independent model review against its source abstract, including factual claims in the deeper take. Missing evidence, an unavailable reviewer or a failed verdict withholds the new recommendation. New Insights also need readable source excerpts and an independent review of their title and every paragraph. Missing or unsupported evidence withholds the new pick. These checks reduce unsupported claims; they do not guarantee factual perfection.
-
Score against a calibrated rubric
The model grades only the dimensions the prompt actually defines. Its quality grade becomes the starting point; corroboration, standing, freshness and measured traction are bonuses that can only push a score up, never drag one down. Nothing below 8 out of 10 is published as though it cleared the bar. In a thin week the best of what is left can appear rather than leaving a shelf broken, marked 'best this week' and showing its real score — the number on the card is never adjusted to fit.
-
Cross the one-way boundary
A single module copies a named list of files from the private side to the public snapshot, stripping anything personal on the way. It is the only thing that writes published content, and a test fails the build if another module ever starts. Every public route — and the chat — reads only the snapshot, never the private side.
-
Grade the result and re-weight the sources
Every published shelf gets an audit attempt against its exact public snapshot. Grades require verified prose authorship and an independent judge; missing evidence or failed judgments produce a withheld mark and an explanation. Separately, each source is re-measured on how often its picks were independently confirmed. A strong record can raise a source’s standing, one step per review. A weak one cannot lower it — “nobody else mentioned its picks” is a fact about everybody else, not about the source. Only going quiet for weeks lowers a tier.
One real run, all the way down
The same journey as numbers, from the most recent video run — read live from /api/content/youtube while this page loaded. Five points where the count actually drops; each bar is what survived the one above it.
reading /api/content/youtube…
08 → 01. What the field independently confirmed feeds back into how much each source is worth on the next sweep. Nobody edits that number by hand — and it is deliberately a ratchet upward, because a source going unmentioned says more about the room than about the source.
one number, one meaning
A score you can inspect
Current automated picks use a quality anchor with measured bonuses. Some books, music and older records retain earlier rubrics; each card uses its stored calculation. A thin shelf may include a labeled below-band pick at its actual score.
Anchored, then lifted
A weighted average diluted strong items: a genuinely good video whose freshness and engagement were merely average got dragged down to about 5, while shelves whose displayed score is essentially the quality grade sat at 8–9. Same axis, incomparable distributions.
the quality grade
the model’s one judgement — everything below the bonuses
8.0 — publish floor
The whole bonus budget is the top +2.00 of the scale. Everything under it is the grade. Bonuses have limited reach: a grade of 6 needs every bonus at its maximum to reach 8.0. A score is an estimate of usefulness, not a certificate that every claim is true.
Every component is a bonus, so nothing gets gutted for lacking views it has not had time to bank — and measured corroboration earns a positive lift until the score reaches its ceiling. Each bonus names where it comes from: corroboration from independent communities (decayed), authority from the source’s earned tier, engagement from real measured traction only, freshness computed from the publication date and never asked of the model.
- quality anchor floor
- 5.5a separate gate, not part of the sum: the raw quality grade must clear this before lift is even considered
- publish band
- 8.0 – 10below 8.0 ships only to keep a shelf from breaking, and only labelled as such — never silently, and never with a flattered number
- so, in practice
- quality ≥ 6.0both conditions must hold, and +2.00 is the most lift can ever add — so nothing graded below 6 can reach the band however well it corroborates
What the model is never asked
The model is asked one thing: how good is this, against a rubric it is shown. It is not asked how fresh something is — a publication date is arithmetic, and a model guessing at it from a title is a made-up number in a slot that looks measured. Freshness, corroboration and engagement are all computed in code from things that were actually counted.
The division is the point. A model is good at judgement and bad at arithmetic, so it is given the judgement and none of the arithmetic. The rubric and the JSON it is asked to return share one definition, so the question and the answer format cannot drift apart — they did once, and every sub-score on three video cards came back as the same number wearing four labels.
deterministic floors
Rules you can inspect before a pick is scored
A persuasive title can mislead a model. Some checks therefore run in code rather than relying on its judgement: title patterns and measured thresholds run before scoring, and prose checks run after generation.
Title hard floor
Title patterns filter course-farm, introductory, listicle and clickbait packaging before quality scoring. That does not exclude every educational implementation: a serious source-backed build outside these patterns can qualify on depth and usefulness. A well-produced “Zero to Hero Full Course” can legitimately score 7.9 on depth and still be study material rather than curation. Two classes: most block unconditionally, and the ambiguous ones (marked) block only for sources without earned authority — a Dwarkesh interview titled “Part 2” is exactly what this engine is for.
full coursezero to herofor beginnerstop N tools / top N AIyou won’t believepart N
Six of nineteen shown. the marked one blocks only for sources without earned standing. Every block is logged with the pattern that caused it, so the record explains why a title died rather than just that it did.
Authority or traction
Unregistered video sources face authority-or-traction checks. An authoritative guest can qualify an interview. A missing view count is never treated as a measured zero.
- traction floor
- 100,000 viewswithin the window below
- traction window
- 270 dayswas 14 — which blocked a video with 159,000 genuine views for being 43 days old
- authority tier
- ≥ 7the registry tier that exempts a source
- guest authority
- countsa primary-source interview is authoritative on an unregistered show
- no view counter
- blocks, and says soan RSS podcast has no counter — reporting “0 views” would be a fabrication
the case that built it
“Agentic AI Full Course with LangGraph … (Zero to Hero) — Part 1”, unknown channel, 15,000 views, published to the video shelf at 7.9. The rubric was not wrong about its production values. The rubric was the wrong instrument.
The speculation gate
Commentary is written by a model in Brendon’s voice. A podcast curator works from publisher descriptions and available transcript excerpts. When the model wants to characterise the episode anyway, it hedges and keeps going. Two published lines, live on 2026-07-30:
“This episode likely cuts through the hype better than most…”
“The conversation with Sarah Guo likely cuts past the usual AV hype…”
A reader cannot tell that apart from a genuine assessment — it is a confident claim the evidence does not support, carrying a person’s name. Nine patterns now block that shape after generation, alongside three other families that catch invented biography, narration of the rubric, and source name-dropping — all applied to the published prose after generation. Having an opinion is the job; pretending to have watched something is not.
published unedited
Each daily audit attempt is public, including withheld grades
The judge reads the exact published snapshot. A grade needs verified prose authors from different labs; missing evidence stays visible.
Who is doing the grading
Three separate model passes — lenses — read each finished shelf and score it 1–10. The questions are not generic, because a shelf can only be judged against what it is for. Focus music graded on novelty would fail for being repetitive, which is precisely what focus music should be.
- podcasts, papers, video
- practitioner-value · anti-slop · freshness-and-mixis it worth a professional’s time, does anything offer only shallow course material rather than source-backed practitioner depth, is it current and varied
- music
- focus-quality · production-quality · mood-diversitygraded on whether it helps you concentrate, not whether it is new
- the written blog
- human-voice · specificity · worth-publishingthe first lens asks one question: does this read like a person, or like AI prose
the obvious objection
Yes — a model is grading work a model helped produce. Two things stop that being a rubber stamp.
The judge is not the author. It runs on a model from a different lab than the one that wrote the published prose, and the model that actually answered is recorded on every grade. Full prose hashes and actual serving receipts bind that check to this snapshot. It verifies prose-author independence; it does not certify independent discovery, selection or orchestration. Selection calls have separate private receipts for diagnosis. When the two collide anyway, the grade is withheld rather than published — a missing mark is a gap someone can chase, a flattering one is a lie with a footnote. The day this separation was introduced, insights fell from 9 to 3.3 and discourse from 9 to 3.7 on unchanged content: the self-graded numbers had been running about five points generous.
A judge cannot grade what it was not shown. Asked to score something it cannot see, a model invents a number rather than admitting it. So every lens declares the evidence it needs — the summary, the date, the source, the duration — and a test renders each shelf’s real judge input and fails the build if a lens was not shown what it requires. A freshness lens once scored four shelves it had never been given a date for. It cannot happen silently again.
What it has changed about itself
Three loops revise the machine's own inputs from its own record. The counts below are read live from /api/content/self-improvement — a loop that stops running shows a frozen number here rather than prose that keeps claiming it works.
reading /api/content/self-improvement…
Ask it yourself, the way a machine would
The site publishes an agent card at /.well-known/agent-card.json and answers JSON-RPC at /a2a. Every skill it advertises is derived from the same table that serves it, so it cannot advertise something it does not do.
Latest recorded audit attempt
reading /api/content/audit…
reading /api/content/audit…
The other half runs on every push
The weekly grade is the slow check. The fast one runs on every code change: a set of labelled examples — including the real failures that caused this engine to be built — is pushed through the actual gates and the actual scoring, using recorded model answers so the result is identical every time and costs nothing to run. Three things must hold, or the change does not ship.
- slop stays out
- alwaysnothing labelled bad may ever come back published
- order is preserved
- alwaysa clearly-strong example must outrank a borderline one
- corroboration counts
- measuredadditional independent evidence earns lift; the score remains capped at ten
the data is open
Inspect the public evidence
These feeds show what the system tracks, what it published and what its audits could establish. Open any of them.
/api/content/pulse
what is being tracked right now, which sources count as independent of each other, and why — written by the same code that does the counting, so it cannot describe a rule it is not running
/api/content/audit
recorded audit attempts, their dates, grades or withheld marks, and the criticism behind them
/api/content/ledger
the track record of what was published and what happened next
/api/content/topics
a deliberately weaker claim: what is being discussed, not cross-verified
The hunting is done. The reading is the part worth keeping.
That is the whole trade. The hunting — feeds, tabs and half-finished episodes, spent mostly finding out what not to read — runs in software, against thresholds printed on this page. What is left on the shelves is short on purpose — some days a shelf adds nothing, and would rather stay short than pad itself. The reading is not automated, and it should not be: it is the part worth keeping.
The simplest way to use it is to let it come to you. Every Monday at 7am ET: what cleared the bar that week, each item linked to its own page here with the score breakdown attached. No digest of a digest, and one click to leave. If you would rather argue with it, the chat takes two questions per 24 hours.