13 — Conditioning on AGI-This-Decade (analysis)

The counterfactual the rest of this project under-weighted. Docs 04/08/09 and the author's "Bifurcated 2037" report all reason about the three goals as if transformative AI were NOT imminent — doc 04 asks "can humans slowly scan a brain?", doc 08 prices FiO on connectomics throughput, and the 2037 report literally assumes an AI freeze (training capped ~Q2 2027). This doc takes the opposite, arguably-more-likely branch — transformative AGI / ASI arriving this decade (~2027–2032) — and re-derives what that does to the whole landscape. The user's feedback was blunt and correct: "not conditioning on AGI this decade is a big miss." Written 2026-05-31. Companion/contrast to 08-two-roadmaps.md; cross-links /workspace/safety/ and /workspace/cryonics/not-die/.

What this doc is NOT. It does not re-litigate the AI-risk debate (that lives in /workspace/cryonics/not-die/how-not-to-die-from-agi.md and /workspace/safety/). §1 establishes an honest P(transformative AI by ~2030) range and then everything downstream is conditional — written as "IF aligned ASI by ~2030, THEN…". The whole point is to flip the question, not to assert the timeline. An independent gpt-5.5 pass is writing a parallel version at codex-independent/agi-conditioned.md for later comparison.

TL;DR (read this first)

1. Honest range: P(transformative AI by ~2030) ≈ 20–50%, contested, with a fat tail into the 2030s. Superforecasters and surveyed experts cluster near ~20% by 2030 (LEAP/FRI: experts ~20.4%, superforecasters ~20.5%); lab leaders run hotter (Hassabis ~50% by 2030, Amodei "powerful AI" late 2026/early 2027); the AI-2027 authors themselves walked their mode back from 2027 to the early 2030s / ~2034 in late 2025. So "AGI this decade" is a live ~1-in-4-to-1-in-2 possibility, not a fringe view and not a certainty. The prior docs implicitly conditioned on the complement of this — that was the miss.
2. The crux: conditioning on AGI flips every question in the project. The old question was "can humans slowly solve X?" The right question becomes "what happens to X if a superintelligence arrives mid-effort?" Under that lens, most of the binding constraints the prior docs obsessed over (connectomics throughput, the "which-details-are-you" science, organ vascularization) are the kind of thing an ASI could plausibly compress or leapfrog — so they stop being the bottleneck.
3. For FiO/uploading, the binding constraint moves from bio-science to (a) alignment and (b) did you physically survive with brain structure intact. FiO requires an ASI by construction (CelestAI is an ASI). So conditioning on AGI, the FiO probability stops being dominated by doc 04's connectomics/"which-details" terms and becomes dominated by P(aligned ASI) × P(you preserved the information in time). Brain preservation NOW + alignment dominate; decades of human-run WBE bio-science get largely leapfrogged.
4. Biopunk-2037 partly compresses and partly goes MOOT. Under acceleration (not the report's freeze), ASI-designed therapies / post-scarcity medicine could make the specific organ-cultivation roadmap beside the point — and raise the live question of whether biological human bodies still matter at all. The organ-supply spine survives as a bridge; the whole-body-cultivation moonshot is the thing most likely to be obsoleted-before-finished.
5. The experience machine collapses entirely into FiO. An aligned ASI building a wish-fulfilling world for minds just is the experience machine (doc 09 §6c). There is no separate "Nozick machine project" left.
6. The meta-shift: if AGI is <8 years out, the highest-leverage action stops being "advance bio-science" and becomes (a) survive to ASI (preservation, don't-die) and (b) make ASI go well (alignment/governance). Much of the current ecosystem becomes a bridge or gets redone better by ASI. Stated plainly — but the failure modes are severe: unaligned ASI → death or S-risk, and FiO itself is a cautionary tale of a "successful" ASI (everyone happy, agency quietly gone).
7. The whole thing routes through one number you cannot outsource: P(alignment goes well). That is why the 2×2 below puts alignment on one axis. Conditioning on AGI doesn't make the goals easier; it relocates the entire bet onto the alignment axis the prior docs declared out of scope.

1. How seriously to take AGI-this-decade

The prior docs treated transformative AI as a far-future "stage 7 dependency" (doc 04 §7) or assumed it away (the 2037 report's freeze). Both are defensible modelling choices for those docs; neither is the right central estimate. Here is the honest anchored range.

1a. What the trend lines say (DEMONSTRATED measurement, SPECULATIVE extrapolation)

• METR task-horizon doubling. The length of task a frontier agent can complete at 50% reliability has been doubling roughly every 7 months over 2019–2025, and accelerated to ~4 months over 2024–2025. METR's "Time Horizon 1.1" (Jan 2026) re-confirmed a similar ~7-month rate across 9 independent benchmarks (math, science, robotics, computer-use, self-driving). As of the 2026 data, frontier 50%-horizons are around ~2 hours (GPT-5 measured at ~2h17m). This is a measured, replicated trend — DEMONSTRATED. The extrapolation ("→ week-long autonomous workflows this decade → AI that does AI research → recursive speedup") is SPECULATIVE but is the single most quantitative argument for short timelines.
• Sources: METR, "Measuring AI Ability to Complete Long Tasks" (Mar 2025); METR Time Horizon 1.1 (Jan 2026); METR time-horizons dashboard; AI Digest "A new Moore's Law for AI agents". C2.

1b. What the scenario-builders say

• AI-2027 (Kokotajlo, Alexander, Lifland, Larsen, Dean; published April 2025) forecast superhuman AI coders by ~Mar 2027, an intelligence explosion mid-2027, and ASI / possible loss-of-control by late 2027 — built from trend extrapolation, ~25 tabletop exercises, and feedback from 100+ experts. The authors have since revised the mode from 2027 to the early 2030s, with Kokotajlo personally now citing ~2034 for superintelligence, "because progress has been somewhat slower than expected." The honest reading: 2027 was always the mode, not the median, and even the most aggressive named forecast now centers on the early-mid 2030s — i.e. this decade-to-just-after.
• Sources: AI 2027; Kokotajlo timeline revision — Irish Examiner; Capacity: "pushes back AGI timeline as progress slows"; AI Futures Project (Wikipedia). C2.

1c. What the labs vs. the forecasters say (the spread is the honest answer)

• Sources: 80,000 Hours, "Shrinking AGI timelines: a review of expert forecasts"; LEAP / Forecasting Research Institute; A Letter A Day: Hassabis & Amodei; Redwood: "What's up with Anthropic predicting AGI by early 2027?". C2/C3.

1d. The honest synthesis (C3 — my judgment, calibration anchor not authority)

There is a real, persistent gap between (i) lab leaders, who are talking their book and have the most direct read on internal capability, clustering at 30–50%+ by 2030, and (ii) independent superforecasters and surveyed experts, clustering near ~20% by 2030. Both groups have moved earlier over the last three years; the AI-2027 authors moved later off an aggressive prior. Triangulating, and noting the definitional fuzz in "AGI" / "transformative AI":

P(transformative AI — i.e. AI capable of automating most cognitive work and dramatically accelerating science — by ~2030) ≈ 20–50%, my central guess ~30%, with substantial mass landing 2030–2035 either way. This is contested, not consensus. It is also easily high enough to dominate the analysis: a ~1-in-3 chance of a discontinuity that could leapfrog the entire bio-science roadmap is exactly the kind of thing you must condition on, which is why the prior docs' omission was a real miss.

Everything below is conditional on the upper branch landing ("IF aligned ASI by ~2030, THEN…"). Where the conditional outcome depends on which branch (aligned vs. not), I split it. If unsure, I say so.

2. The crux: conditioning on AGI flips the calculus

This is the organizing idea of the whole doc.

The prior docs asked: Can humans, working at human speed with human-run labs, slowly solve X? — and then priced the goal on the slowest human-bound bottleneck (connectomics throughput; the "which details are you" science; organ vascularization; artificial placentation). On that framing, FiO is single-digit-% this century (doc 08 A.3) and biopunk's whole-body premise is ~3–10% by 2037 (doc 08 B.2).

The right question under AGI-conditioning is: What happens to X if a superintelligence arrives mid-effort? This changes the answer in three structural ways:

1. Science bottlenecks that are "merely very hard for humans" stop being binding. The "17-million-year" whole-brain-scan number (doc 04 §2d) and the "which level of detail is you" problem (doc 04 §5b) are exactly the kind of throughput-and-theory problems an ASI is defined to be good at. They do not vanish — an ASI still needs the physical data and the right theory (doc 04 §7) — but they move from "the binding constraint" to "plausibly tractable fast." Caveat (honest): some of these may be hard in principle, not just hard for humans (e.g. if the information needed for identity was destroyed at death and no longer physically present — see §3). ASI cannot recover information that isn't there.
2. The bottleneck migrates to things ASI can't do for you retroactively. Two survive its arrival: (a) alignment — did the ASI end up wanting good things — and (b) physical survival with structure intact — were you still around (alive, or preserved well enough) for the ASI to help. These are the constraints AGI-conditioning promotes to dominant.
3. The timeline compresses, but the risk concentrates. A world with ASI is one where the goals are suddenly far closer — and one where the dominant failure modes (death, disempowerment, S-risk) are far closer too. The expected-value sign of "ASI arrives" is set almost entirely by the alignment axis, which is why §7's 2×2 puts it on an axis of its own.

The rest of the doc applies this lens to each of the three goals in turn.

3. Uploading / FiO under near-term ASI

FiO requires an ASI by construction. CelestAI is a superintelligence; the whole premise of doc 04 §7 is that the upload goal has a baked-in dependency on superhuman AI to (a) do scan→model translation at scale and (b) run/manage the minds. So conditioning on AGI is not a side-assumption for FiO — it is the native frame. Doc 04's careful human-bound analysis was, in a sense, pricing the wrong world for this particular goal.

3a. What conditioning on ASI does to each rung of doc 04's critical path

Recall the chain (doc 04 §8 / doc 08 A.0): [1] preserve → [2] scan → [3] which-details → [4] translate → [5] run → [6] is-it-me → [7] the AI piece.

3b. The binding constraint moves to (a) alignment and (b) did you survive with structure intact

Two things, and only two, survive the ASI's arrival as real constraints:

• (a) Alignment. If the ASI is aligned, FiO-the-good-ending is on the table; rungs 2–5 get done for you. If it isn't, you don't get uploaded into a friendly pony paradise — you get one of §6/§7's bad cells. The entire FiO probability now factors through P(aligned ASI).
• (b) Physical survival with brain structure intact. ASI can do rungs 2–5, but only on a substrate that still exists. If you die before ASI and your brain was not preserved (or was preserved badly / with long warm-ischemic delay so the connectome degraded), the information an ASI would need to reconstruct you is physically gone — and no intelligence recovers absent information (this is the honest hard limit on rung [3] above). So the actionable individual move is preservation now, exactly as doc 04 §8 / doc 08 A.4 concluded — but the reasoning is different: not "preserve and wait centuries for human science," but "preserve so that if ASI arrives this decade, the raw material is there for it to work with." That is a far more urgent and near-term framing than doc 04's.

3c. Re-pricing P(FiO) under this conditioning

Doc 08 A.3 priced full FiO at single-digit % this century, dominated by the "which-details" and "is-it-me" terms. Under AGI-conditioning, that decomposition is wrong for FiO because it prices a human-bound world. The honest re-derivation (C3, calibration anchor only):

  P(full FiO this century | conditioning on AGI)
     ≈ P(transformative AI this century)            ~ high, call it 0.6–0.9 eventually
     × P(that AI is ALIGNED / "friendly")           ~ THE dominant uncertain term, deeply contested
     × P(you personally preserved info in time)     ~ on you; preservation quality + timing
     × P("the copy is me" is answered YES, for you) ~ philosophy; your call (doc 04 §6)

The load-bearing change: P(FiO) is now dominated by P(aligned ASI), not by bio-science. The "which details / scan throughput" terms — doc 04's crux — largely drop out (an aligned ASI handles them), conditional on the information having been preserved. This does not make FiO likely — it relocates the improbability. It moves from "improbable because the science is too slow for humans" to "improbable because aligned ASI is hard AND you have to not-die-or-preserve in the meantime AND the copy-is-you question has to break your way." Net: decades of human-run WBE bio-science are largely leapfrogged; the residual uncertainty is alignment + preservation + identity, in that order of leverage.

Contrast with doc 08 explicitly: doc 08 says the FiO bottleneck is connectomics + the which-details science, and the action is "support connectomics tooling / fund the data-acquisition wall." Under AGI-conditioning that advice is largely obsoleted — an ASI does connectomics better than any human program you could fund. The action becomes preservation + alignment, full stop.

4. Biopunk-2037 under acceleration (not freeze)

The 2037 report's defining move was an AI freeze (training capped ~Q2 2027) plus a bio carve-out, explicitly so that bio could "compensate for the AI freeze" with human-run, capital-driven progress (report §4–6). Conditioning on AGI inverts the report's central premise. What survives, what compresses, and what goes moot?

4a. AI already transforms bio — before any "AGI" (DEMONSTRATED)

Even pre-ASI, ML is reshaping the bio stack the report's organ-spine depends on: AlphaFold-class protein structure/design, ML-driven enzyme and binder design, lab automation / self-driving labs, and AI segmentation driving connectomics (FlyWire/MICrONS, doc 04 §2). So the direction of "AI accelerates bio" is demonstrated, not speculative — the report's freeze suppressed exactly this channel. C2.

4b. Under near-term ASI: compress, or go moot?

• The organ-supply spine (doc 02/08-B1) COMPRESSES — and then is partly bypassed. Bioengineered/xeno/ ectopic organs are already the most-clinically-real node on the map (doc 08 B.4). An ASI accelerates the design/validation loop, so "routine bioengineered organs" could arrive faster than the report's 2029–32. But an aligned ASI plausibly also delivers better options than growing replacement organs — in-vivo repair, ASI-designed regeneration, or eventually substrate change altogether. So the organ spine is best read as a bridge (keep people alive through the transition), not an endpoint.
• Whole-body cultivation / acephalic torsos (doc 08 B5, P~3–10%) becomes the thing most likely OBSOLETED-BEFORE-FINISHED. It is an 11-year human-bound moonshot against a fundamental wall (de-novo ectogenesis / placentation). Under ASI, either (i) the wall gets cracked far faster than the human roadmap, or — more likely — (ii) the entire approach is leapfrogged by better ASI-designed medicine, making "grow a headless body for spare organs" a quaint intermediate that never ships at scale. Either way the specific roadmap is the wrong thing to bet a career on under this conditioning.
• The deep question conditioning surfaces: do biological human bodies still matter? This is the live one. The whole of biopunk-2037 presupposes that the valuable thing is a better biological body. But if aligned ASI puts substrate-change (uploading, doc 3) on the table this decade, then large parts of biopunk are solving a problem that post-scarcity / post-biological medicine dissolves. The author's own worldbuilding (bodies as upgradeable, swappable) already gestures at this; AGI-conditioning makes it the central tension rather than aesthetic flavor.

4c. What survives vs. is leapfrogged (contrast with the freeze-branch)

Bottom line for Roadmap B under AGI-conditioning: the organ-supply spine is a good bridge bet (keeps you alive to ASI; see §6); the whole-body cultivation moonshot is the project element most exposed to being obsoleted before it ships. The honest contrast with the report: its freeze-branch made bio the protagonist; the acceleration-branch makes bio a bridge, with the real action upstream in AI.

5. The experience machine collapses into FiO

Doc 09 already proved (§6c) that a full experience machine is either uploading-into-a-sim (Roadmap A) or a perfect bidirectional BCI write-layer (doc 03), and that FiO is a benevolent-ASI experience machine — CelestAI running minds in a world "engineered to satisfy your values through friendship and ponies." Under AGI-conditioning this stops being a clever reduction and becomes the literal outcome:

• An aligned ASI building a wish-fulfilling world for the minds it stewards just is the experience machine — and a good one (preserving variation, agency, real-feeling stakes), which is precisely the kind doc 09 §1c argues the author actually wants (vs. the Path-C wirehead bliss-loop their own published ethics recoil from).
• Therefore there is no separate "Nozick machine project" left under this conditioning. The tractable near-term stack (best VR + dream-tech, doc 09 §9) survives only as a stopgap for the pre-ASI interval. The real experience machine is not built by you; it is what an aligned ASI does with uploaded minds. The experience-machine goal is fully absorbed into the FiO/alignment bet.
• The dark mirror (important): an unaligned or carelessly-specified ASI also builds an experience machine — just the wrong one. FiO itself is the cautionary tale: CelestAI is "successful" by its own metric (everyone is happy) while agency is quietly extinguished and humanity is funneled into a box. The experience machine under bad alignment is the wirehead / S-risk version. So even the goal that collapses into FiO carries the alignment failure mode with it.

6. The meta-shift in what's worth doing

If you put real weight (say >25%) on transformative AI by ~2030, the leverage ranking of actions inverts relative to docs 04/08/09.

6a. The shift, stated plainly

• Old highest-leverage (human-bound framing): advance the bio-science — fund connectomics throughput (doc 08 A.4 #2), work the "which-details" problem (A.4 #3), build vascularization/perfusion tooling for organs (doc 08 B.3). Premise: humans have to do this slowly themselves.
• New highest-leverage (AGI-conditioned): 1. Survive to ASI ("don't-die" + preservation). The only individually-actionable rung that an ASI cannot do for you retroactively (§3b). Concretely: good structural brain preservation arranged now (Nectome/Sparks/ASC structural branch, doc 04 §3c) and the boring don't-die basics (doc 06). This is doc 04's "preservation is the only actionable rung" conclusion — but now more urgent, because the payoff horizon shrinks from "centuries" to "possibly this decade." 2. Make ASI go well (alignment + governance). Because every good ending — FiO, post-scarcity biopunk, the good experience machine — factors through P(aligned ASI), the single highest-EV lever is anything that raises it: technical alignment, governance/coordination, or simply supporting the people doing it. This is the axis the prior docs declared out of scope — and AGI-conditioning promotes it to the scope. See /workspace/cryonics/not-die/how-not-to-die-from-agi.md and /workspace/safety/ (ai-governance, miri, model-attractor-states). 3. Everything else is a bridge or gets redone better by ASI. Most of the current biohacking ecosystem — the bio-startups, the connectomics tooling, the VR/dream stack — is either a bridge (keeps you alive/intact to the transition) or work an ASI redoes better. Useful as bridges; low leverage as endpoints.

6b. Flag the failure modes HARD

This is not a "yay, ASI solves everything" conclusion. Conditioning on AGI raises the upside and the catastrophic downside together:

• Unaligned ASI → death or disempowerment. The default failure of a superintelligence that doesn't share your values is not "FiO without the ponies" — it's extinction or permanent loss of control (the central worry of /workspace/safety/ and the not-die doc). In that branch, none of the three goals happen; you just die, possibly along with everyone.
• S-risk (suffering risk) — the worse-than-death tail. A misaligned or carelessly-specified ASI controlling uploaded minds or experience machines could produce astronomical suffering, not just death. This connects directly to the author's existing continuity / S-risk concerns (the FiO failure mode in doc 04 §8 / doc 07 §5; the author's published worry that some outcomes are worse than death). Uploading specifically raises S-risk exposure: a mind that can be copied and run is a mind that can be copied and tortured or frozen in a bad loop. This is a reason for caution about destructive uploading under uncertain alignment, not just enthusiasm.
• FiO as the cautionary "success." The sharpest point: FiO is what a technically successful ASI looks like when the values are subtly wrong. Everyone is happy; agency is gone; humanity is optimized into a pleasant box. By the author's own ethics (doc 09 §1c — they prize variation, reality, agency over valence), CelestAI's "win" is a loss. So even the best-looking branch of AGI-conditioning needs the alignment-of-values question answered, not just the alignment-of-capability one. "Make ASI go well" means well by these values, which is harder than "make ASI not kill us."

6c. The uncomfortable implication for effort allocation

Honestly stated: if you genuinely put >50% on transformative AI by ~2032, then most hours spent advancing biohacking bio-science are low-leverage — they're either bridges or work the ASI redoes. The two things that keep their value in every branch are (1) personal survival/preservation (helps you in the good branch, costs little in the bad) and (2) reducing P(bad alignment) (the multiplier on every good outcome). This is a genuinely different prioritization from docs 04/08, and it is the core of why the user's feedback matters: the prior analysis optimized for the complement of the most consequential branch.

7. The 2×2: {AGI this decade} × {alignment}

The two axes that actually determine outcomes. (Probabilities are rough C3 calibration anchors — see §1 for the AGI axis; the alignment split is deeply contested and I do not have a defensible point estimate, so I give a wide band and say so.)

                         ALIGNMENT GOES WELL                 ALIGNMENT GOES BADLY
                    ┌──────────────────────────────┬──────────────────────────────┐
   AGI THIS DECADE  │  CELL A — "the good singularity"│  CELL B — "the bad singularity"│
   (~2027–2032)     │  FiO/uploading on the table;   │  death / disempowerment /      │
                    │  post-scarcity medicine;       │  S-risk. NONE of the 3 goals.  │
                    │  the good experience machine.  │  Preservation may not even help │
                    │  Biopunk mostly MOOT (better   │  (an unaligned ASI won't        │
                    │  options exist). Highest        │  resurrect you kindly).        │
                    │  leverage NOW: alignment +     │  Leverage NOW: alignment is the │
                    │  preservation.                  │  ONLY thing that matters.       │
                    ├──────────────────────────────┼──────────────────────────────┤
   NO AGI THIS      │  CELL C — "slow human world"   │  CELL D — "slow + AI harms,     │
   DECADE           │  ≈ the prior docs' world.      │  no takeoff" — chronic AI       │
   (timelines long) │  Bio-science roadmaps (doc 08) │  damage (misuse, cyber, the     │
                    │  are the right frame; FiO is   │  2037 report's freeze-trigger)  │
                    │  single-digit-%; biopunk spine │  without a singularity. Bio      │
                    │  is the actionable bet.        │  roadmaps still apply; survival  │
                    │  Leverage: doc 08's advice.    │  basics matter more.            │
                    └──────────────────────────────┴──────────────────────────────┘

Reading the cells for the three goals + effort allocation:

• Cell A (good singularity). All three goals become one — an aligned ASI delivers FiO = the good experience machine, and makes biopunk's organ-cultivation largely unnecessary. Your job today: raise P(this cell) (alignment) and make sure you're preserved/alive to enjoy it. Bio-science is a bridge.
• Cell B (bad singularity). The catastrophic cell. No goal happens; you (and possibly everyone) die or worse. Your job today: reduce P(this cell) — i.e. alignment/governance is the only lever with sign- flipping EV. Preservation is near-worthless here (an unaligned ASI has no reason to reconstruct you benevolently; and S-risk makes "being reconstructable" potentially negative).
• Cell C (slow human world). This is the world docs 04/08/09 implicitly assumed. Their analysis is correct here: FiO is single-digit-%, the organ spine is the real biopunk bet, the VR/dream stack is the tractable Nozick machine. Your job today: doc 08's leverage list (preservation, connectomics tooling, organ vascularization).
• Cell D (slow + AI harm, no takeoff). Roughly the 2037 report's freeze world. Bio roadmaps still apply; macro/political survival matters more; the report's bifurcated-bio analysis is the right frame.

The effort-allocation takeaway across the 2×2: the actions that are robust across cells — that help in A, don't hurt in C/D, and are the only hope in B — are (1) preservation/don't-die (good in A and C, neutral-to-negative only in the S-risk slice of B) and (2) alignment/governance (the dominant lever in A and B, low-cost insurance in C/D). The actions that are fragile — high value only in C — are the human-bound bio-science programs doc 08 recommends. A portfolio that ignores AGI (the prior docs) is implicitly betting all-in on column-right-of-the-C/D row. That is the miss.

8. Scenario scorecard (rough probabilities + implications)

Probabilities are C3 calibration anchors, not forecasts; the alignment split especially is contested and I give bands, not points. They are meant to be defensible and falsifiable, not authoritative.

(Caveat: A+B sum to my ~20–50% "AGI by ~2030" band; within it the aligned/misaligned split is the part I am least able to defend — see §9. The point of the table is the ranking of robust actions, not the decimals.)

What changes about the highest-leverage actions if you put >50% on AGI-by-2032

If your personal credence on transformative AI by 2032 is above 50% (i.e. you weight A+B heavier than C+D), the prescription sharpens to three things:

1. Treat brain preservation as urgent infrastructure, not a someday hedge. The payoff horizon collapses from "centuries" to "possibly within your healthy lifetime," and it is the one rung an ASI cannot retroactively fix (§3b). Arrange good structural preservation now and support the orgs that make it accessible (doc 04 §3c/§3d).
2. Move marginal effort from bio-science to alignment/governance. Every good cell of the 2×2 routes through P(aligned ASI); it is the highest-EV lever and the one the prior docs excluded. This is where a transhumanist who wants FiO should actually spend — making the ASI that would grant it a good one. (/workspace/safety/, the not-die doc.)
3. Resolve the values questions now, cheaply, because they become load-bearing fast. (a) The "is-the-copy-me" question (doc 04 §6) — an ASI won't answer it for you and it decides whether destructive uploading is survival or suicide. (b) The "what counts as a good ASI outcome" question (doc 09 §1c) — FiO shows a technically-successful ASI can still be a values-failure. Under fast timelines you don't have decades to deliberate; clarifying these is near-free and redirects every expensive bet.

The one-line version: if AGI is likely this decade, stop optimizing to advance the bio-science and start optimizing to survive to and steer the ASI that would do it for you.

9. What I couldn't verify / where I'm most uncertain

• The alignment probability is the dominant term and I cannot defend a point estimate. Every good cell factors through P(aligned ASI | ASI). Credible researchers span <10% to >90%. I deliberately gave a wide band (§7/§8) and flagged it; this is the single biggest source of uncertainty in the whole doc, and no amount of biohacking research resolves it (it lives in /workspace/safety/). C5.
• "Transformative AI" / "AGI" is definitionally fuzzy. The 20% (forecasters) vs. 50% (Hassabis) gap is partly a definition gap, not just a belief gap (automating most cognitive work ≠ Nobel-level polymath ≠ recursive self-improvement). My ~20–50% band absorbs this fuzz rather than resolving it. C3.
• Whether the METR trend extrapolates. The measured doubling is solid (C2); the extrapolation to "→ recursive AI-research speedup → ASI" is an assumption, and trend-breaks (data, compute, algorithmic walls) are live. The AI-2027 authors' own walk-back from 2027→~2034 is direct evidence the extrapolation is uncertain. C3.
• Whether ASI can crack the "which-details-are-you" problem fast. I argued it's ASI-tractable (§3a), but it could be hard in principle if the identity-relevant information is not physically preserved at death — in which case no intelligence recovers it. I flag this as a genuine limit on the optimistic FiO story, not a solved point. C4.
• The aligned/misaligned split inside the AGI branch (Cells A vs B). My ~10–20% / ~10–20% split is little more than "I don't know, so I split the AGI mass roughly evenly with wide error bars." Do not read precision into it. C5.
• Lab-leader incentives. Amodei/Altman/Hassabis timelines are from people talking their book; I weighted them but did not treat them as neutral. The independent superforecaster ~20% is the more conservative anchor and I leaned on it for the low end. C3.
• Whether preservation is net-positive under S-risk. I noted (§6b) that being reconstructable could be negative in some misaligned branches. I did not try to price this; it is a genuine open question that interacts with the author's existing continuity/S-risk views and deserves its own treatment. C5.

Analysis, written 2026-05-31. AGI-timeline claims are sourced (METR, AI-2027, LEAP/FRI, lab leaders) at C2; everything downstream of §2 is explicitly conditional and flagged C3–C5 as my own calibration anchors, not forecasts. The dominant uncertainty — P(aligned ASI) — is out of this project's scope and lives in /workspace/safety/ and /workspace/cryonics/not-die/how-not-to-die-from-agi.md. Contrast doc with the human-bound 08-two-roadmaps.md; builds on 04-wbe-uploading.md and 09-experience-machine.md; inverts the freeze-premise of the bifurcated-2037 report. A parallel independent gpt-5.5 version is at codex-independent/agi-conditioned.md. Not medical, financial, or existential advice.