Status, 2026-10-01

Branch worktree-seed-phase (not yet in main; I merge). Everything below is committed on that branch.

Update, 2026-10-06: the expert panel is frozen

Frozen by me on 2026-10-06 ("Freeze and merge", amendment A18), on branch panel-gaps-2026-10-06 (PR 2). Before the freeze the three unclear core cases were closed (Shanahan, Bengio and Floridi pass, all thin), my rulings were recorded as A17 (Legg, Cowen and Yudkowsky kept; Sutskever's 2015 line "Not a timeline statement"; both thin readings count; search first, then freeze), and a cohort search added Liang Wenfeng, Arthur Mensch and Arvind Narayanan.

• Core, 25: Kurzweil, Bostrom, Hanson, Hinton, LeCun, Hassabis, Altman, Marcus, Ng, Russell, Brooks, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman, Legg, Cowen, Yudkowsky, Shanahan, Bengio, Floridi. Quotas: skeptic 8, outside the US and UK 6, philosopher 4, frontier lab 11 (all minimums met).
• 2020s cohort, 12: Amodei, Aschenbrenner, Kokotajlo, Cotra, Huang, Leike, Mollick, Chollet, Sutskever, Liang, Mensch, Narayanan.
• Wide list, 308: data/wide_list_candidates.csv, now built from the full TIME100 AI lists 2023 to 2025 as well as the MIRI and AI Impacts dataset.
• data/speakers.csv carries the 37 core and cohort members with tier and quota tags; python3 scripts/validate.py passes.

Next step: full-panel collection. Trace each of the 32 newly frozen core and cohort members year by year from first statement to the present (methodology-v2 section 3: fixed queries logged in data/search_log.csv, quiet years logged as rows), as was done for the five seed speakers; then the light capture for the wide list (one statement per person per era, rows tagged tier: wide). The public streams have not started.

Update, 2026-10-01 (later): amendment A15 and the newspaper archive checks

A15 applied (commit 2fad7d4). My rulings on the eight calls are in amendments.md as A15 and in the codebook (new ruling 13, one row per capability, with the two splits as its worked example; the accepts list now includes firm unhedged first-person dates, "predict", "going to", "almost certainly" and minimum years). Data changes: LeCun's four "not a road to AGI" rows rejects → unexpressed; marcus-2016-01 and marcus-2026-03 rejects → accepts with open-ended ranges (from 2046 and 2036); marcus-2026-04 and altman-2025-06 split into a/b rows, so the dataset now has 159 statements. Kurzweil, Chollet, Kai-Fu Lee and Karnofsky tagged frontier_lab with dates. None of the changed rows is in my blind sample.

Archive checks (commit a9757ef). scripts/archive_check.py searched the Guardian and the New York Times for 30 of the 35 core candidates and recorded 1,184 hits in docs/panel-evidence-2026-10/archive-checks.csv. Keys are read at run time from ~/.config/sentient-horizons/agi-zeitgeist.env; a scan found no key in any repo file; .gitignore now excludes *.env and .agi-zeitgeist/.

• Stopped early. The New York Times closed the connection twice, inside its rate limit. After the first, the run was resumed once; after the second it was stopped under the overnight rule. Not checked: Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman, and Brynjolfsson's NYT 2024-onward search. All six already pass, so nothing can fall, but their newspaper evidence isn't on the same footing yet. Rerun with: python3 scripts/archive_check.py "Erik Brynjolfsson" "Jürgen Schmidhuber" "Kai-Fu Lee" "Toby Walsh" "Max Tegmark" "Mustafa Suleyman" (first delete Brynjolfsson's partial rows from the CSV).
• What counted. Only 1 hit in 10 named the person in a headline or abstract; 1,037 named them only in the article text and were recorded but not counted, as ruled. Counted pre-2018 pieces: Kurzweil 8, Musk 8, Hassabis 6, Bostrom 4, Brooks 2, Hinton 1, LeCun 1, Ng 1 (plus Goertzel 1 and Pearl 1, both not about AI futures).
• Verdicts changed. LeCun becomes a full pass (NYT 2015 interview). Brooks no longer relies on his own essay. Seven move from pending to fail: Pinker, Floridi, Legg, Sutskever, Shanahan, Wooldridge, Goertzel. Etzioni's only recent lead (NYT 2024 on election deepfakes) isn't an AGI statement, so he still fails C2.

Freeze candidates now: 19 core passes (skeptic 7, outside the US and UK 5, philosopher 3, frontier lab 9: all minimums met), 3 unclear (Yudkowsky, Bengio, Cowen), 14 fail; 8 cohort passes plus Sutskever unclear.

For me

1. The headline-or-abstract rule is strict. It excluded 14 pre-2018 Guardian pieces about AI that quote Shanahan, 27 for Pinker and 21 for Etzioni, and it moves five skeptic candidates to fail. If that's not the intent, a looser rule ("named in the text of a piece whose headline or abstract is about AI futures") would bring several back; the CSV has what's needed to re-mark.
2. My reading of "headline-only match": name in the headline but no abstract to show the piece is about AI. Recorded with the flag headline_only, not counted (3 hits). Correct me if you meant something else.
3. The NYT dropped connections twice at about 25 requests each time. If it happens again, a longer pause than five a minute may help.
4. The core is at 19, six short of "about 25"; the three unclear cases would take it to 22.

Not started, as instructed: public streams, wide-list collection, full-panel collection.

Update, 2026-10-01: after my rulings (amendment A14)

I ruled on all fifteen decisions listed further down. All seven steps of his 2026-10-01 brief are done:

scripts/validate.py runs clean. The chart is unchanged: no lead-frame codes changed in the recode.

Panel freeze candidates (details in docs/panel-verification-2026-10.md)

• Core, pass now (19): Kurzweil, Bostrom, Hanson, Hinton, LeCun, Hassabis, Altman, Marcus, Ng, Russell, Brooks, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman. Quotas: skeptic 7, outside the US and UK 5, philosopher 3, frontier lab 7. All four minimums are met.
• Core, unclear (10), mostly waiting on the newspaper archive checks: Yudkowsky, Bengio, Cowen, Pinker, Floridi, Legg, Sutskever, Shanahan, Wooldridge, Goertzel.
• Cohort, pass (8): Amodei, Aschenbrenner, Kokotajlo, Cotra, Huang, Leike, Mollick, Chollet; Sutskever unclear.

New questions for me (raised while applying the rulings)

1. Firm dates without a listed word. "I set the date for the Singularity as 2045", "I'm still saying 2029", "I've stuck with my 2029 prediction", "put the date at 2029" (Kurzweil 2005, 2014, 2016, 2024) and Altman's "by 2030 … I'd be very surprised" and "Humanity is close to building digital superintelligence" were kept as accepts. Coding them "no position" would misreport the speaker. Confirm, or rule them out.
2. A hedge and an acceptance word on the same date ("I think it may be 20 years or less") coded uncertain, the hedge deciding. Confirm.
3. Bare "takeoff" coded si_assist, since the words don't say who improves the systems. Confirm.
4. Rejecting a route, not the capability. LeCun's "LLMs are not a road towards AGI" (four rows) stays rejects. Under a strict reading of ruling 9 these would be unexpressed.
5. "At least a decade" / "at least thirty or fifty years" (Marcus 2016, 2026) stay rejects with a minimum year. Ruling 11's reasoning could make them accepts.
6. Words that may count as acceptance: "predict" (Hinton 2023, kept uncertain because of "without much confidence"), "almost certainly" (Marcus 2026, read as above 50%), "going to" (several rows, read as "will").
7. Two rows worth a human look: marcus-2026-04 (now si_assist, but Marcus says that kind of tool exists, so disputed fits poorly) and altman-2025-06 ("the takeoff has started", target kept as superintelligence, marked ambiguous).
8. Arguable tags still not counted: frontier lab for Kurzweil, Chollet, Kai-Fu Lee and Karnofsky; Kokotajlo not tagged philosopher (trained in philosophy, works as a forecaster).

Still waiting on me

Code the 30-statement sample (python3 scripts/agreement.py compares); freeze the panel; Guardian and New York Times API keys for the archive checks; merge the branch. Not started, as instructed: archive checks, public streams, wide-list collection, full-panel collection.

The rest of this file is the status as written on 2026-09-30. The fifteen decisions at the end were ruled on 2026-10-01 in amendment A14.

What was completed

Your first brief (seed phase), all five steps, then the version 2 brief, all six steps, then amendment A13 (three-tier panel), steps 5a to 5e.

scripts/validate.py runs clean: 14 sources, 157 statements, 200 figures, 9 events, 740 search-log rows.

The seed in numbers

Pass check (methodology-v2 section 12): at least 60 statements with working sources: met (157; every link reachable, six sites refuse scripted checks but were read by page-reading tool). Codes spread across buckets: met. Double-coding agreement on two speakers: not yet done (see decisions).

What the seed shows so far (five people only, not a finding)

The "whether, then when, then now" story does not show up in these five. "When will it arrive?" leads in every era: 14 of 21 statements in 2005 to 2015, 14 of 36 in 2016 to 2022, 49 of 100 since 2023. "Is today's system close?" rises from 2016, not 2023 (14 statements in 2016 to 2022). "Can it improve itself?" leads only 5 statements, all since 2023. The clearest change is dates moving closer (Hinton: 2070 in 2015, then 5 to 20 years, then "close to AGI" in 2026), not a change of question.

What was skipped, and why

• Metaculus history: not retrieved. The site refused every request (403) and the Internet Archive was offline; every attempt is in the search log.
• The full TIME100 AI lists: only about half of each list could be read (time.com refuses scripted requests). The wide list records the gap; missing names were not filled in.
• Several panel checks: news sites the tools could not reach (New York Times, Guardian, New Yorker, Wired, BBC and others) left some prominence checks unresolved; those are marked "fails as searched" or "unclear", never rounded up.
• No public streams, no wide-list collection, no full-panel collection, as instructed. Population survey items stay in aggregates.csv; moving them to data/surveys.csv belongs to the population stream.
• Corrections found while recoding (fixed, listed for the record): two spliced quotes (altman-2017-01, lecun-2025-05); the 2026 Hinton briefing row ("maybe a year" was the window for lawmakers, not a superintelligence date); an Altman quote moved from Davos 2024 to StrictlyVC 2023; several non-arrival years removed from timeline fields.

What needs your decision

Panel freeze (details in docs/panel-verification-2026-10.md):

1. Bylines. Does a piece the person wrote count as "naming" them? This decides whether the skeptic quota of six is reachable: without it, 4 of the 16 core passes are skeptics; with it, 6 (adding Marcus and Brooks), two of them at medium confidence.
2. "Pre-2017" for the skeptic tag: strictly before 2017, or 2017 and earlier? (Kai-Fu Lee, LeCun, Dennett.)
3. Co-written statements for the 2024-or-later check (Ernest Davis).
4. Dennett died on 2024-04-19. Keep or drop?
5. Arguable quota tags (frontier lab: Kurzweil, Chollet, Kai-Fu Lee, Karnofsky; philosopher: Hinton, Hassabis; non-US/UK: Chalmers, Legg; skeptic: Suleyman).
6. Cohort prominence check: the A13 brief read literally asks the cohort for three national pieces from 2017 or earlier, which contradicts the cohort's definition. I treated it as information only. Confirm?
7. Archive checks for the C3 failures that are probably the tools (Pinker, Shanahan, Wooldridge, Goertzel, Floridi, Legg, Sutskever). They would mostly add skeptics.
8. Wide list size: 184 names against a target of 60 to 100, because the rule takes every name. Trim by a rule, or keep?

Codebook rulings (the recoders hit the same questions independently):

9. "Not here yet" coded as rejects. Most of Marcus's 23 and LeCun's 13 rejects reject AGI existing now or by a near date, not ever. A chart of "rejects" would mislead. Options: restrict rejects to "never / not in the frame given" and use present_capability: disputed for "not yet", or add an explicit rule.
10. Hedged words. "May", "might", "could" next to a date: one recoder coded them uncertain, another accepts. This is the most likely point of disagreement in the double-coding check; the codebook should settle it before that check.
11. Distance as acceptance. "Decades away" read as accepting it arrives eventually (Marcus, LeCun): fine, or unexpressed?
12. "Takeoff" and "recursive self-improvement" map to si_sustained by the codebook, but speakers often narrow them ("very slow takeoff", "a larval version"). Only 5 statements have a self-improvement target, so the four levels were rarely tested.

Process:

13. Double-coding for the pass check. Who codes two of the five independently: you, or a second agent pass? Ruling the codebook questions first would raise agreement.
14. Merging the branch. The work is on worktree-seed-phase; main is still at the skeleton commit. Your main checkout holds uncommitted copies of files the branch also contains: scripts/validate.py and docs/methodology.md (my edits from early in the session), and docs/amendments.md, docs/methodology-v2.md, docs/2026-09-30-kickoff.md and docs/reviews/ (from the planning session). The branch has all of them, with amendments.md also carrying the agent-decision entry, so a fast-forward merge will refuse until the main-checkout copies are set aside. If nothing else changed there since: in ~/Code/agi-zeitgeist, run git stash push -u -m pre-seed-merge, then git merge --ff-only worktree-seed-phase, then check the stash is redundant before dropping it (git stash show -p stash@{0}). Or open a pull request once the repo has a GitHub remote.
15. Licence (open decision 4 in the kickoff plan) is still unset.

Rendered from docs/status-2026-10-01.md at version a7bd17a.