AGI Zeitgeist

Document

Tip intake: how a statement I come across enters the record

A working document. It is one of the documents the site is built from, published as it is written for anyone checking the work, and it uses the project's own shorthand. The plain account of the same material is on the home page.

Tip intake: how a statement I come across enters the record

Written 2026-10-18. Status: RULED. I ruled on 2026-10-18, through the master session, "all as recommended": the four decisions at the end, the step 1 rule for quotes, and amendment A29 in docs/amendments.md, which records the change to the method. The first case run under it is Nick Bostrom's interview on The Roman Forum (pull request 18).

Changed 2026-10-21 by my ruling (amendment A32). In my words: "I want a rule that includes everything I find and gives every other person on the tracker ample search time to find things they have said too that I otherwise miss." Every tip now counts at once, in every figure, marked "added from a tip"; the tip_only exclusion of decision 1 is retired. Fairness comes from the other side instead: every person on the panel is searched every week (part 4 of docs/weekly-reading.md). The step 3 search stays, because it now measures whether the searches would have found the tip, and a miss improves the searches for everyone (step 3b).

It answers a question I asked on 2026-10-18, after sending a link to Nick Bostrom's interview on The Roman Forum (published 2026-10-16): when a prominent figure is heard discussing exactly the questions the tracker asks, what place does that have in the project, and how is it handled the same way every time?

The idea in one paragraph

A tip counts. A statement I send in becomes an ordinary row in every figure, marked "added from a tip" wherever it is shown, so a reader can see which rows came in that way and how many each person has. What keeps my reading from deciding whose words are on the record is the other half of the rule: every person on the panel is searched every week, so their new statements are found whether or not I happen to read them; and when a tip shows the searches missed something, the searches are improved for everyone. Each tip goes through the same steps whoever sends it: archive the primary source, check who is speaking against the frozen panel, run that person's logged queries to see whether the rule finds the source, extract the candidate statements, and hand them to coding under the codebook like any collected row. Separately, and only as a draft, it can produce a short dated "panel watch" note for Sentient Horizons' social channels. The data step and the commentary step never feed each other: a note can cite a row, but a row never exists because a note wanted one.

One procedure, two triggers

The weekly reading's part 4 searches every panel member every week (from 2026-10-21, docs/weekly-reading.md). Tip intake is the same job with a different trigger: I send a link instead of a schedule finding it. Part 4 follows steps 1, 2, 4 and 5 of this file for every new statement it finds, so there is one way in for a panel statement found between full passes, not two. The difference is only in the notes: a row from a tip starts tip; a row part 4 found says so, because the rule's own searches found it.

Trigger cues for a session: zeitgeist: followed by a link, or "log this for the zeitgeist tracker", or a link to a panel member speaking about AI's future sent with a request to save it.

Steps

1. Get the primary source and archive it

2. Who is speaking?

Check data/speakers.csv:

3. Would the rule have found it?

Run the person's fixed year queries for the statement's year (Y1 "<name>" AGI <year>, Y2 "<name>" "human-level" intelligence <year>, Y3 "<name>" superintelligence OR singularity OR "self-improving" <year>), the video search (V1 "<name>" interview <year> limited to youtube.com, added 2026-10-24 by amendment A35), and every search in data/added_searches.csv, plus a check of their own outlets, as the collection does. Log every search in data/search_log.csv with ids continuing the person's series.

Either way the rows count in every figure (A32). The search result decides only the notes:

A claim the tip adds to the claims ledger carries from_tip = true, and the claims page marks it "added from a tip".

3b. When the searches missed it: a better search for everyone

Added 2026-10-21 (A32). A miss is information about the searches, not about the person. When step 3 does not find the source:

  1. Say why, in the search-log row for the tip source: not indexed yet (a search for the exact title finds nothing either, as with Bostrom's interview two days after it was published), or the searches' words did not match (the person spoke on a podcast whose page never says "AGI", for example).
  2. If the words did not match, add one search that would have found it, worded for any person, not this one: a new year query (a Y4) or a new kind of own outlet to check (the podcasts a person has appeared on, for example). Run it for the person in the tip and log it, to show it finds the source.
  3. Add it as one row of data/added_searches.csv: the date, a short name (Y4, Y5 and so on), the query with <name> and <year> where the person and year go, what it reaches that the searches missed, the tip's statement or claim ids, and the search-log ids that show the miss and the test. From the next weekly reading on, part 4 runs it for every person on the panel, never only the person in the tip, and every full pass runs it too.
  4. Searches are only ever added. No row of that file is removed or narrowed, and no original search is dropped because a new one covers it.
  5. The site's changelog lists each added search, with its date and the tip that prompted it, built from the file, and every time it runs is in the search log. The reply to me names it.

This is automatic: no search waits for my approval. I ruled so on 2026-10-21, on the master session's proposal, in my words: "Yes, apply them automatically." (amendment A32, item 4).

4. Extract the candidate statements

Read the whole source. A candidate is any passage that bears on possibility, timing, present capability, self-improvement or consequences, under codebook section 0's unit rule and amendment A15's one-row-per-capability rule. Section 3's busy-year rule applies to the person's year as a whole, and amendment A26 (reworded by A34 on 2026-10-24) keeps every statement on the "now" question. Skip the cold-open montage of an episode when it repeats passages from the body; code the passage where it is said in full.

5. Code them the way the collection codes

Same codebook, same coder label and same agreement procedure as the full collection that produced the person's other rows. A tip is not coded by a lighter process because it arrived by hand. Where the person already has rows, write one line in the notes saying how the new statement compares with their latest row on the same question, quoting both ids. That is the within-person comparison methodology section 3 reports separately.

6. Draft the panel watch note (optional, drafts only)

A short dated note for Sentient Horizons' social channels, saved to docs/panel-watch/<date>-<speaker>.md: who the person is in one line, what they said (one quote under fifteen words, attributed, with timestamp), how it compares with their earlier rows, and a link to the tracker. Rules:

7. Hand over

One commit (or pull request) per tip: the archived source line in the manifest, the search log rows, the statement rows, and the draft note if any. The reply to me is a short list of what I need to rule on or post (including any better search from step 3b), and nothing else.

Tip spread

python3 scripts/panel_status.py prints how many statements and claims came in by tip, per person, and what share of each person's statements that is. The weekly reading's report carries it, and the site's changelog lists every tip row with a count per person, so anyone can see if my reading tilts the record toward the people I follow.

Each decision keeps its alternative and its stated confidence for the record.

  1. Rows the rule would not have found. Recommended: keep them, tagged tip_only, left out of figures and shares, listed on the site's changelog as "noted, outside the measured record". Alternative: code them as ordinary rows. Confidence: high. Without the exclusion, my own feed becomes a sampling frame, and it leans toward whatever the enthusiast venues are amplifying, which is the bias the vanguard sub-series is kept apart for. Ruled 2026-10-18, by me: as recommended. scripts/charts.py and scripts/site.py leave the rows out, and the validator rejects tip_only without tip (A29). Replaced 2026-10-21 by my ruling (A32): every tip counts at once, marked "added from a tip", and every panel member is searched weekly. The validator now refuses tip_only.
  2. Archived transcripts in a public repository. Recommended: keep the files local (add sources/*/ to .gitignore) and commit only sources/manifest.csv, so a reader can see what was archived, when, and its checksum, without the repository republishing a full transcript of someone else's interview. Alternative: commit the files. Confidence: medium. Ruled 2026-10-18, by me: as recommended.
  3. Where panel watch notes live. Recommended: drafts here, in docs/panel-watch/, next to the rows they cite, with the posting record kept by the Outreach process once it is ratified. Alternative: draft them in the Outreach folder from the start. Confidence: medium. Ruled 2026-10-18, by me: as recommended. The first note (Bostrom, 2026-10-18) is not one of the outreach trials; it goes to the first outreach review on 2026-10-24 as a proposal.
  4. Is the "would the rule have found it" test run again later? Added 2026-10-18 after the first run. Run on the day a tip arrives, step 3 measures how fast search engines index a new source as much as how far the rule reaches: the Bostrom interview was two days old, and a search for its exact title found nothing either (search log bostrom-127). Recommended: yes. At the person's next full pass, run the same year queries again, log them, and lift tip_only from the rows if the results now include the source or a report of the same appearance, saying in the notes on what date and by which search it was lifted. The re-run only ever lifts the tag, never adds it, and a row stays out of figures until it is lifted. Alternative: the verdict on the day of intake stands for good, which is simpler but means the sooner I send a tip, the likelier it is to be left out. Confidence: medium; the remaining question is whether "the next full pass" is soon enough, or whether a fixed wait (for example 30 days after intake) is better. Ruled 2026-10-18, by me: as recommended, at the next full pass. Moot from 2026-10-21 (A32): with no tip_only tag there is nothing to lift. The re-run survives as the weekly search of every person, and a miss leads to step 3b.

Rendered from docs/tip-intake.md at version a7bd17a.