Late Lessons, Jensen Huang and AI

Revision log: 06 (and 05) after the fidelity, fairness, usefulness and public-framing reviews#

Working file, 26–27 September 2026. Records each issue raised in fidelity.md (F), fairness.md (Fa), usefulness.md (U) and public.md (P), whether it was verified against the sources, and what was done. Section numbers refer to the revised 06 unless marked 05. Backups of 05 (pre-revision) and 06 (mid-revision) are in the session scratchpad only.

Outcome codes. FIXED = verified and changed; PARTLY = verified, part applied, rest noted; NOT CHANGED = checked and left, with reason; INTERNAL = a pre-publication gate or question, moved to INTERNAL-notes-for-essay.md §D.

Size. 06 grew from about 18,500 to about 29,000 words (the usefulness review asked for 3,000–4,000 words of additions; the fairness and fidelity fixes added context in many places). Nothing substantive was cut. Candidate cuts are listed in INTERNAL §D.


Fidelity (fidelity.md)#

# Issue Verified Outcome
F1 “Apparent agency is ‘illusory’” misstates 2026-01-31 (object is self-awareness; hedged) Yes (corpus l.38) FIXED. §4.6 now quotes the hedged self-awareness sentence; functional “stochastic” agency added as real (§3.4, §4.6 Application and Limits).
F2 Open-weights “two strands” attribute others’ views to him (2025-07-23; 2023-07-12) Yes (both posts) FIXED. §9.2 row rewritten as “record is thin”, reporting both texts accurately; §7 item 8 drops 2023-07-12 from basis, open-weights application labelled [Inferred, low].
F3 NANO 2026 “no governance solution” truncated Yes (NANO l.64–68) FIXED in §3.6, §7 preamble and §8.4 (claims outcome “evidence”, no measured outcomes).
F4 Pause alignment omits 2023-04-04 call for urgent coordinated action Yes FIXED. §3.2 bullet, §4.8 heading and Limits, In brief item and limits sentence.
F5 Hedges dropped: FFTF p.281 “not a zero probability”; broad definition of catastrophic risk Yes (FFTF-F p.281; 2023-05-31 l.50–52) FIXED in §4.1, §3.2, In brief.
F6 Containment “almost word for word” overstated; Maynard treats containment as temporary Yes (2025-07-06 l.76–78) FIXED. §4.4 relabelled [Inferred, medium], “shares Huang’s premise”, temporary-condition quote added; §11 Symmetry “overlaps with”.
F7 “Careers” claim miscited to NN 2014-03 Yes (S1: Maynard & Aitken 2016 pp.999–1000) FIXED. §6.4 cites Maynard & Aitken 2016 p.999 (co-written), cf. NN 2014-03 p.160.
F8 2026 [mixed] “threat to something it values” cited to 2019 Yes (2026-07-16 l.126) FIXED in §8.3; “Garbee lesson” renamed.
F9 In brief labels too high; antitrust, eminence, liability items Yes FIXED. Header “[Implied or Inferred]”; items reworded; liability “(low-medium confidence)”.
F10 Labels too high or in wrong scope (§4.6, §5.1, §5.8, §6.1 item 3, §4.3, §5.7, §4.2) Yes FIXED in each place as proposed (incl. nearer Huang on personal motive, nearer Klein on structural pressure; “regards that depiction as incomplete”).
F11 April 2026 essays’ provenance not disclosed; 30Y quote truncated Yes (S6 provenance) FIXED. New §1.4 bullet; §3.1 full 30Y phrase plus balancing next sentence; §11 note. Same 30Y fix applied in 05 §4 (C3 area).
F12 Series introduction quoted from draft 3 as if published Yes (draft 3, dated 27 Sept) INTERNAL (gate). Citation left as planned; all quoted phrases listed for re-check.
F13 Moltbook n.4 misread (subject, tense, Moltbook-specific) Yes (l.78) FIXED in §9.3; “distinctive concern” relabelled [Implied].
F14 Page and slug errors; “Weighing 2009” undefined Yes (FFTF-D page markers; slugs.txt; S3) FIXED. p.166; NN 2015-12 p.1006; four full slugs; Weighing 2009 defined in §1.5 and added to 05 App. C.
F15 Clipped or softened quotations (9 items) Yes, each FIXED: severity floors + “hidden consequences”; “a potential risk”; “trying hard” + “lack the breadth”; “at worst” smoke-and-mirrors + “valuable role”; “Constitutional AI approach”; first movers “not yet”; PEN 2006 p.32 wording + Testimony 2008 p.8; “may be linear” as the transforming function; “federal capacity” (also §3.6).
F16 Co-authorship and [mixed] flags uneven Yes FIXED. Toxicol. Sci. 2011, Nature 2006, ILSI 2005 flagged; §1.4 Trojan bullet widened to pp.11–14; §11 wording corrected.

Fairness (fairness.md)#

# Issue Verified Outcome
Fa1 Proxy table calls Huang outlier on open weights; drops “company practice” finding Yes (leaders comparison §3; 03 §9.5) FIXED. Supplier row split (open weights “Partly”; China/chips “No”); company-practice row added; Reading and In brief distinguish statements from practice.
Fa2 “Denies competitive pressure” misstates [40:21]; Meta omitted Yes (transcript l.609–612; NBC 24 Sept) FIXED in §4.8, table, In brief (minority with Meta, Andreessen, Mistral).
Fa3 “Safety aperture” rests on inference from absence; “I completely agree” omitted; labs’ discretionary tools omitted Yes (transcript l.350; 2026-07-16 l.86) FIXED. §5.2 adds the exchange and Klein’s non-asking; paper’s concession added; table row renamed and downgraded to [Inferred, medium]; In brief caveated.
Fa4 Narayanan and Kapoor missing; only interested parties cited on “what AI is” Yes (02 §9.2; E4 §2.3) FIXED in §5.1 (“Who sits where”), §5.4 (“We were wrong”), §9.3, §9.6.
Fa5 Huang’s several auditors ignored in §4.5 Yes (02 §4.2 All-In) FIXED in §4.0, §4.5, §5.3, §7 item 3.
Fa6 Conditional lab commitments support Huang’s moral-hazard point; Karnofsky’s affiliation missing Yes (2026-07-16 l.106; transcript l.817–820) FIXED. New paragraph in §4.9; §5.4; §9.3 names Karnofsky’s role; §6.3 Reading no longer says “does not touch Huang”.
Fa7 “Inverse error” mischaracterises Huang; unmarked single [mixed] source Yes (transcript l.495–499) FIXED. Comparison deleted; replaced with deflationary-mechanism vs non-deflationary-effects reading (02 T9). §3.2 keeps the lecture point with a secure 2020 source added.
Fa8 “Before any product existed” narrows Huang’s containment rule Yes (transcript l.500–503) FIXED in §4.4 Limits.
Fa9 In brief compresses “we understand it, obviously” Yes FIXED in In brief and table (full clause).
Fa10 Deficit-model reading of water line lacks context, sits in “value” list Yes (transcript l.1532–1540) FIXED. Moved to §5.9, full turn quoted, low-medium, veto noted.
Fa11 Interests section one-sided Yes (02 §8.4 table; 02 §10.2; 03 §9.2) FIXED in §6.4 and §6.1 item 4 (“the same is true of other self-held stop rules”).
Fa12 02 wrongly said to apply the Late Lessons lens Yes FIXED in §8.5.
Fa13 04 used as factual source; “Huang’s reading” of restart Yes (only 04 and M9 carry the 27 June detail) PARTLY. Row now cites OpenAI’s report and METR, notes sources disagree and that the 27 June detail is unchecked; “consistent with Huang’s containment diagnosis”. Direct check of OpenAI’s report is INTERNAL.
Fa14 Series intro quoted selectively (negative) Yes (draft 3) FIXED in §1.6, §11, In brief, §8.5 (“informative because of its depth and breadth”).
Fa15 Hinton “wild guess”; critics welcomed Huang’s bar Yes (02 C124; 02 §7.1, §9.2) FIXED in §4.1 and §4.0.
Fa16 “Promoter and overseer” row places state concern in Huang’s column; unverified authority Yes FIXED (recast; “reported; not verified”).
Fa17 §7 “Direction” lines read as prescription Yes FIXED. All recast as “Where his work points” / “On his reasoning”.
Fa18 Other unmarked single [mixed] sources Yes FIXED (§6.1 item 7; §9.2 internal-sequence row; EO 14409 row reduced to what the paper records, no pre-emption position attributed).
Fa19 Ex Machina comparison primes the reader Yes FIXED. Film removed from §5.3; principle cited from FFTF p.162. Not moved to §3.7 (unnecessary).
Fa20 Minor overstatements (antitrust; pause; “fantastic”; C083 grade; §5.1 label) Yes (transcript l.778) FIXED each.
Fa21 Omissions from Huang’s conditions Yes (transcript l.1222; 02 §4.2; E1 Scotland; 02 §2.3 Bessent) FIXED in §4.0 and §4.7.
Fa22 Only OpenAI’s incidents listed Yes (leaders comparison timeline) FIXED. Gemini row added to §9.2.

Usefulness (usefulness.md)#

# Issue Verified Outcome
U1 “Broader landscape” not built Yes (2026-09-15 lists; 2024-08-25 quadrant; 2025-05-04; 2026-07-19) FIXED. New §8.6 with quadrant table. Correction to the review: “Job replacement and redistribution” is on his 2018 list, not the “risen in significance” list; 06 states it correctly.
U2 Purpose and justice missing Yes (FFTF pp.287–290; 2019-03-31; 2020-11-12; 2024-07-13; 2025-01-07) FIXED. New §3.0; justice paragraph in §3.6; new §4.11; distribution folded into §5.10; “neither camp” (oxygen; FFTF p.290). Excluded-book sentences of 2026-05-21 avoided.
U3 Jobs nearly absent Yes (transcript l.137–189; 03 §4.6; leaders comparison) FIXED. New §5.10; jobs and energy rows in §6.3; divergence in In brief.
U4 Record of answering leaders unused Yes (2023-05-15; 2023-05-25; 2023-10-02; 2024-08-04; 2024-03-21; FFTF p.249) FIXED. New §6.5; GMO lesson in §4.2; FFTF p.249 in §5.3.
U5 “Complex, not complicated” and social risk missing Yes (2024-08-07 l.108; 2018-09-03) FIXED in §3.1, §5.3, §7 item 7. 2016-03-02 and 2017-04-10 not added (not verified in time; lower value).
U6 Integration losses (Evo 2, Asilomar, 2025-05-04, Kapoor/Narayanan, independence concessions, Waymo/Cruise, which history, consciousness) Yes, each FIXED (§6.1 item 9; §3.6; §3.4/§5.2/§5.5/§7 item 9; §5.4; §4.0/§5.3; §5.4/§6.2; §5.7; §3.4).
U7 Future Rising and 2016–2022 underused Yes (FR PDFs checked for all quotes) PARTLY. FR pp.72–73, 148, 150, 152–153, 166–167 added (§3.2, §5.1, §5.8, §6.1, §10.3); 2019-03-31, 2020-10-15, 2020-11-12 added; period note in §11. 2022-02-12 and 2022-06-13 not added.
U8 Load-bearing claims on 2026/[mixed] texts Yes FIXED. §5.2 rebuilt on 2018–2025 sources; In brief caveat; §5.4 and §7 item 7 lead with secure sources.
U9 RSI absent Yes (02 §3.9; transcript l.1110, 1159–1165) FIXED. §9.2 row and new §9.5, flagged thin.
U10 Reception not read through the lens Yes (02 §9.2) FIXED. New §9.6 (Marcus, Mowshowitz, Hashim, Narayanan/Kapoor, Yudkowsky).
U11 §10 gaps (radiology as hype; leaded petrol; safety-culture claim; fairness to Huang; claim table) Yes (04 l.33, 49, 69; LL2-03 p.61; 03 §4.6) FIXED. §10.2, §10.3, new §10.5 with table.
U12 §7 omits recurrent proposals Yes FIXED. Items 13 (expertise) and 14 (pathway mapping) added; GMO/benefits into item 12; social-risk reboot into item 7; 2025-05-04 into item 9.
U13 Findings not prioritised Yes PARTLY. Three strongest alignments and divergences marked in In brief. Optional six-lens-question table not added.
U14 Tilt negative in §1.6/§11 Yes FIXED (with Fa14).
U15 Leaders beyond Huang thin Yes (leaders comparison; 02 §9.2; 2025-03-02) FIXED. §9.4 expanded (Zuckerberg, Nadella, Musk, Altman, Bengio).
U16 Uncovered interview segments Yes FIXED. New §9.7, with silences stated.
U17 Housekeeping (uncited App. A items; 391-post dating; draft intro; length) Yes FIXED for App. A and dating; draft intro INTERNAL; length not cut (see Size).

Public framing (public.md)#

# Issue Verified Outcome
P1 Appendices point to unpublished working files (06 and 05) Yes FIXED. 06 App. B and §1.2–1.3 replaced with neutral note; 05 §1 Sources, crosswalk, §6 preamble, T10/T11 headings, App. A preamble, App. B and App. C paths cleaned. Decision on publishing any working layer is INTERNAL.
P2 Draft series introduction quoted as published Yes INTERNAL (gate).
P3 04 identified by draft filename Yes FIXED in 06 (filename removed). Renaming and final version of 04 INTERNAL.
P4 06 provenance incomplete (05 and 01–04 AI-prepared; wording; model named) Yes FIXED in §1.1 and §1.4. “Reviewed by him” kept per the brief’s provenance note; “reviewed and edited” question INTERNAL.
P5 05 provenance omits model, vendor, LL2-22 Yes FIXED (05 subtitle, Method, Limits).
P6 “Clarification” wording signals off-page correction; repeated Yes FIXED. 06 run-ins renamed; §1.5 cites personal communication. 05 heading renamed, “cautioned against” sentence removed, eight repetitions replaced with “(§1)”.
P7 Process language in 05 Yes (corpus confirms Modem Futura is co-hosted) FIXED (“at Maynard’s request”, “Batch readers”, “free and legitimate”, “compiled before”, “concept index Family”, “Theme file” column).
P8 In brief on the article unlabelled, “endorse” Yes FIXED (label; “consistent with”; §10.1 retitled; his reservation added).
P9 Series intro attributions stretched Yes (draft 3 text) FIXED in In brief and §8.5.
P10 Unresolvable identifiers and code collisions Yes FIXED (Weighing 2009; E4 citation; 02 P/T/A/C explained; prefixes; Mirror defined; Garbee lesson renamed; Raine and Centaur glossed; “Late Lessons diagnostic lens”). “S3 2026” key kept, since supplement groups are no longer referenced publicly.
P11 05 authorship sentence implies ghost-authorship Yes (S5) FIXED (05 §7 item 10).
P12 Garbee ruling wording could read as judgement on co-author Yes FIXED in 06 §1.4 and 05 §1.
P13 Excluded-book residue in 05 D3 (“drift”) Yes (2026-07-16 cites the book for “values drift”) FIXED (05 D3 reworded to the paper’s own documentation).
P14 Timestamp convention; URLs Yes FIXED (§1.2, §1.5).
P15 “Post-recording” dated from release Yes FIXED (“post-release”).
P16 Inconsistency on published view of Huang Yes FIXED (§1.1).
P17 “Writing skill trained on” Yes FIXED (§1.2, §10). Whether to match the introduction’s “train” INTERNAL.
P18 §1.1 overstatement and dating Yes (05 §7 Phase 0) FIXED.
P19 05 “we/our” voice Yes FIXED at the six flagged lines (quotations kept).
P20 Loose §1.4 wording Yes FIXED.
P21 Unspecified conflict of interest in §10.4 Yes (04 n.4) FIXED.

Not changed, or outside the reviews#