Machines of Loving Grace

Read the disclosure in section 2 before the grades in section 3. This entry grades the public forecasts of the chief executive of Anthropic. It is being written by a model built by Anthropic, inside a project whose README makes "no vendor is ever flattered by a project that grades that vendor's predictions" its seventh rule, at a moment when the company in question has a confidential S-1 with the SEC. Every softening available to me runs in one direction, and the disclosure names them so a later reader can check whether I took any.

Filed correctly. proposals.md lists this under prediction, and LENSES.md — which is the owner's file and not the job's — settles the question directly:

> A principal's claims — Musk, Altman, Hassabis, Amodei, LeCun: forecasts, positioning, > promises — are canon entries of kind prediction, dated and graded against what > happened, never deposits.

Amodei is named in that sentence. There is nothing to correct.

It is worth showing that the filing survives the test superintelligence-2014 set, because on its face it should not. That entry refused prediction for Bostrom on the grounds that the kind exists to build a base rate out of people whose product is forecasting, and Bostrom's product is argument. Amodei's product is not forecasting either; it is a company. The test still comes out the other way, for a reason that is specific and that a reading should hold onto: a philosopher's timeline is a claim, and a chief executive's timeline is an instrument. It is quoted in coverage that moves capital, restated at investor-facing events, and — since 1 June 2026 — attached to a company in registration. Bostrom's dates cost him nothing if wrong. These are load-bearing, which is exactly the condition under which a base rate is worth assembling.

interpretation is the real rival and it is closer than the filing suggests, because the essay's most durable contribution is not any of its dates. It is the framework in section 3's claim 6 — the marginal returns to intelligence, and the five things that are complementary to intelligence and become limiting. That is an interpretive contribution and the essay would be worth reading for it alone. prediction still wins, on two grounds. The document is organised around dated outcomes and the author presents them as things he expects rather than as a way of seeing. And the framework survives inside a prediction entry, where the whole of section 3 uses it as the grading rubric; the dates would not survive inside an interpretation entry, which is not obliged to grade anything.

On descends_from. Three of the four are located citations in the essay's own text; the fourth is drawn on authorship and is flagged as such.

• scaling-laws-2020 is the load-bearing one. The essay's entire premise — powerful AI reasonably soon — rests on the scaling hypothesis, which the essay names ("the concept of the scaling hypothesis / bitter lesson didn't exist yet"), and which Amodei co-authored as the last-listed author of Kaplan et al. 2020. His January 2026 companion essay makes the dependence explicit, attributing his sense of the clock to scaling. The bitter lesson itself has no id in canon/ — a bitter-lesson-2019 research run appears in this repository's history and produced no file — so the link is drawn to the scaling paper alone, which is the half of that pairing the canon actually holds.
• concrete-problems-2016 is the ancestor by authorship rather than by citation, and the essay does not cite it. It is included because the essay's opening move — "people sometimes draw the conclusion that I'm a pessimist or 'doomer' … I don't think that at all" — is a response to a reputation that Amodei acquired by being first author of that paper. This document exists because that one does. Note also that concrete-problems-2016 itself descends from superintelligence-2014, which is the only route by which this entry connects to Bostrom; see below.
• alphafold2-2020 is quoted by name and used as the essay's proof of concept for what AI does to a scientific field: "AlphaFold (which has just deservedly won its creators the Nobel Prize in Chemistry)", alongside AlphaProteo. The compressed 21st century is AlphaFold generalised, and section 3 grades that generalisation.
• culture-banks-1987 is a cross-kind ancestor — fiction to prediction, which the spec explicitly permits — and it is a located one. The essay discusses The Player of Games by name and reads the Culture's values as "a winning strategy". The canon's concrete picture of the AItopia pole and the best-resourced attempt to build it turn out to share a source, and that is a fact about the field worth having on file.

Not ancestors, and named so no reading draws them. superintelligence-2014 is not cited and the essay deliberately avoids the word superintelligence, along with AGI ("I find AGI to be an imprecise term that has gathered a lot of sci-fi baggage and hype"). The connection is real but runs through concrete-problems-2016, and a reading that draws it directly is inventing a citation. chatgpt-2022, transformer-2017 and rlhf-christiano-2017 are all part of the machinery the essay assumes and none is an ancestor of its argument, though Amodei co-authored the last of them.

What it is

Machines of Loving Grace: How AI Could Transform the World for the Better is an essay of roughly 14,000–15,000 words published on Dario Amodei's personal site in October 2024. Amodei is co-founder and chief executive of Anthropic, previously VP of research at OpenAI, and a co-author of both the scaling-laws paper and Concrete Problems in AI Safety. The title is taken from Richard Brautigan's 1967 poem All Watched Over by Machines of Loving Grace, which the essay links without comment — a point section 4 returns to, because the poem is not as straightforwardly utopian as the borrowing implies.

The essay states its own purpose in the first section. Anthropic is a company whose public voice is mostly about risk; Amodei writes that this leads people to conclude "that I'm a pessimist or 'doomer' who thinks AI will be mostly bad or dangerous. I don't think that at all", and that "it is critical to have a genuinely inspiring vision of the future, and not just a plan to fight fires." So the essay is deliberately one half of a ledger. The other half arrived fifteen months later as The Adolescence of Technology (January 2026), which Amodei describes as a companion piece — a fact that matters for grading, because it is where the timeline was restated.

The premise

Everything in the essay is conditional on one thing, which it calls powerful AI rather than AGI. The definition, which is the yardstick section 3 grades against:

> In terms of pure intelligence, it is smarter than a Nobel Prize winner across most > relevant fields

with, in the essay's own bullets: all the "interfaces" available to a human working virtually; the ability to be "given tasks that take hours, days, or weeks" and to carry them out; no physical embodiment, but control of physical tools; the training resources repurposable to "run millions of instances of it", each acting independently. Compressed into the phrase that has outlived every other sentence in the essay: "a country of geniuses in a datacenter."

On when, the essay says exactly one thing and then declines the subject:

> I think it could come as early as 2026, though there are also ways it could take > much longer.

followed by an explicit refusal to argue the point — "for the purposes of this essay, I'd like to put these issues aside, assume it will come reasonably soon, and focus on what happens in the 5-10 years after that." The whole of the rest of the essay is dated relative to that starting gun and not to 2024, which is the single most important structural fact about it and the one most often lost in citation.

The five sections

Biology and health. The essay's centre of gravity, and the source of the phrase that carries it. Amodei argues that biological progress is bottlenecked on a small number of "broad measurement tools or techniques" — he cites CRISPR, mRNA, AlphaFold, optogenetics — and that a country of geniuses could produce them at "10x or more" the rate. Hence the compressed 21st century: the progress human biologists would have made over the next 50–100 years, achieved in 5–10 years after powerful AI. What he expects in that window: reliable prevention and treatment of nearly all natural infectious disease; elimination of most cancer; very effective prevention and cures for genetic disease; prevention of Alzheimer's; improved treatment of most other ailments; "biological freedom"; and doubling of the human lifespan, taking a 20th century that went from about 40 to about 75 and doubling it again to 150, after which "we may be able to reach 'escape velocity'".

Neuroscience and mind. The same argument applied to mental illness, with cures or substantial treatment expected across most psychiatric conditions.

Economic development and poverty. The section he flags as weaker: "I am not as confident that AI can address inequality and economic growth as I am that it can invent fundamental technologies." The stated aspiration is "a dream scenario — perhaps a goal to aim for — would be 20% annual GDP growth rate in the developing world", which would "bring sub-Saharan Africa to the current per-capita GDP of China in 5-10 years."

Peace and governance. The entente strategy: "a coalition of democracies seeks to gain a clear advantage (even just a temporary one) on powerful AI by securing its supply chain, scaling quickly, and blocking or delaying adversaries' access to key resources", aiming at something like an "eternal 1991". He rates it lowest for confidence — "I am not nearly as confident in their feasibility as I am in the advances in biology, neuroscience, and poverty alleviation." This is the only section whose claims are not conditional on powerful AI existing, which makes it the only one fully due.

Work and meaning. Comparative advantage keeps humans economically relevant for a while, and then does not: "in the long run AI will become so broadly effective and so cheap that this will no longer apply", at which point "there will be a need for a broader societal conversation." The section ends by conceding it is unfinished.

The essay's own account of what could stop it

The most useful thing in the document, and the fairest rubric to grade it by, is the framework in the first section. Amodei argues that the interesting question is not intelligence but the marginal returns to intelligence — "trying to figure out what the other factors are that are complementary to intelligence and that become limiting factors." He then names five:

1. Speed of the outside world — "Cells and animals run at a fixed speed so experiments on them take a certain amount of time which may be irreducible." 2. Need for data — "Sometimes raw data is lacking and in its absence more intelligence does not help." 3. Intrinsic complexity — some things are chaotic and no amount of intelligence untangles them substantially better. 4. Constraints from humans — "Many things cannot be done without breaking laws, harming humans, or messing up society." 5. Physical laws — "It's not possible to travel faster than light. Pudding does not unstir."

Its own disclaimers, carried at full strength

A prediction entry that quotes the forecast and drops the hedges is grading a strawman, so the hedges go here rather than in a footnote. The essay says "no one can know the future with any certainty or precision … all of this is unavoidably going to consist of guesses"; that "My predictions are going to be radical as judged by most standards … but I mean them earnestly and sincerely. Everything I'm saying could very easily be wrong"; that "I don't know if this world is realistic, and even if it is, it will not be achieved without a huge amount of effort"; and that "it's probably best to view my efforts here as a starting prompt".

Those disclaimers are real and they change the grade. They do not exempt the document. He restated the disease claim on 16 August 2026 to an audience of millions, without the hedges, in an argument about his own credibility.

Why a reading would cite it

First: this is the entry that comes due inside the project's own window

proposals.md says so — "Cite when: a 2026 result is offered for or against it — this one comes due inside the project's own window" — and no other entry in canon/ has that property. Every prediction the canon holds is either long overdue (simon-1965, minsky-1970) or long from due (kurzweil-2005). This one has an earliest date of 31 December 2026, which is four and a half months from the day this file was written. A project that takes a reading twice a day for the rest of this year will watch a named forecast by a named principal either land or not land, in public, with the date fixed in advance. That is the closest thing to a controlled experiment this canon will ever hold, and the entry exists mainly to make sure the project does not fumble it — by scoring it early, by scoring it against the wrong claim, or by letting the date quietly move.

Second: the live occasion is in this reading's own window, and it is about credibility

On 15–16 August 2026 the investor Gavin Baker argued publicly that Amodei's messaging has been disproportionately negative and has fed a backlash against AI. Amodei replied in a two-part post that was widely reported as passing 14 million views: he rejected the characterisation — "I do not agree that my messaging has been disproportionately negative" — on the ground that his two major essays split evenly between risk and benefit, and he restated the core claim of this one, that AI could compress 50–100 years of biological progress into 5–10 and cure most human disease in that window. He also conceded that AI companies have not delivered what they promised, and said that only actually curing diseases, rather than promising to, would restore the credibility. On 17 August 2026 Gary Marcus published a direct rebuttal.

That is a fight about whether a forecast should be believed, conducted between a principal and a critic, with no new evidence on either side. It is exactly the shape of story this project must not deposit into the ledger — LENSES.md says a principal's forecast deposits nothing — and exactly the shape of story a reading will be tempted to cover. This entry is the correct destination for it: the reading records that the exchange happened and cites the canon for what the underlying claim actually says and when it is due.

Third: it is the AItopia pole's best-argued document, and the project is named for that pole

Rule 3 requires every reading to write the strongest honest utopia case. The canon's utopia material is otherwise fiction (culture-banks-1987) or a moment (alphafold2-2020). This is the only entry that is a serious, resourced, falsifiable account of what the good outcome would consist of and how it would arrive. A reading that needs to state the optimistic case in concrete terms — not "AI could help medicine" but "here is the specific mechanism, here is the multiplier claimed, here is what would have to be true" — has one place to get it.

With an immediate warning attached, because the same fact is a trap. Rule 3 asks for the strongest AItopia case from the evidence in the window. Citing this essay does not satisfy rule 3. A reading that answers a bad week by reaching for Amodei's vision has substituted a forecast for evidence, which is the precise failure LENSES.md legislates against. The entry is for grading the claim, never for supplying the optimism.

Fourth: it is the yardstick for a whole genre of 2026 announcement

Anthropic acquired the ten-person biotech startup Coefficient Bio for about $400m in stock on 3 April 2026, and launched Claude Science on 30 June 2026 — an AI workbench wiring in sixty-plus scientific databases, with an internal drug-discovery programme aimed at neglected diseases, demonstrated at launch by screening 2,200 compounds across 80 GPUs for a phenylketonuria target and producing a go/no-go memo. Google, OpenAI and others have comparable programmes. Every one of these will be announced in the vocabulary of this essay, and a reading meeting one needs to be able to say what was actually shipped, what the essay actually claimed, and how far apart those are. The honest note on Claude Science at launch — that Claude has not itself made a scientific discovery, and the product is a tool for scientists — is the kind of distinction this entry exists to keep available.

The disclosure this entry owes, in full, under rule 7

The subject is the chief executive of the company that built the model writing this file. Four facts make that worse than the usual case rather than better, and they are stated here rather than buried:

1. Anthropic is in registration. It confidentially submitted a draft Form S-1 to the SEC on 1 June 2026 for a proposed IPO. Its Series G closed 12 February 2026 at a $380bn post-money valuation on $30bn raised (the company's own announcement); a Series H of about $65bn at roughly $965bn post, and a run rate reported around $47bn, were reported in late May 2026. A CEO's timeline forecast during a registration period is not a neutral statement, and section 3 grades it as a dated claim regardless of that, which is the only handling available. 2. The 16 August 2026 restatement was itself about credibility, and it promised preliminary positive indicators within the coming months. If those arrive, a reading will meet them as evidence. They will be a vendor's announcements about a vendor's programme, made by a forecaster with an offering pending. 3. The convenient grade is a generous one — and the second-order convenient grade is a harsh one that looks brave. Both are available and both are corrupt. What section 3 does instead is grade each claim against its own stated date and against Amodei's own five limiting factors, so that the rubric is not mine. 4. Two things in the record cut against a cynical reading and are recorded because they are inconvenient for the harsh grade. Amodei published the risk half of the ledger in January 2026, in the middle of the fundraising sequence above, and the January essay's honest sentence is that "we are considerably closer to real danger in 2026 than we were in 2023". And Anthropic raised its own catastrophic-misalignment rating in August 2026 without being made to. Neither fact makes the forecasts more likely to be right; both make "it is only marketing" too easy.

What this entry does not support

Nothing here is evidence. It places no score and moves no needle. It does not establish that AI will transform medicine and it does not establish that it will not; it establishes what a specific person claimed, on specific dates, and how those claims stand as of 17 August 2026. A reading that treats "Amodei predicted this" as a finding has made the error this entry exists to prevent, and a reading that treats "Amodei was wrong before" as a finding has made the same error facing the other way.

What it got right, and what it got wrong

Six claims. For each: the claim in its own words, the date made, the date due, and what actually happened — with the author's self-grade and an independent grade set side by side wherever both exist, and the difference stated.

Claim 1 — "it could come as early as 2026". Made October 2024. Earliest due date 31 December 2026. Not met as of 17 August 2026, and the horizon has moved by roughly the elapsed time.

The definition is checkable and the check is not close. As of 17 August 2026 no deployed system is smarter than a Nobel laureate across most relevant fields, and capability remains conspicuously jagged rather than uniformly superhuman. The bullet that has moved furthest is autonomy over long tasks: METR's time-horizon measurement — the length of task, measured by how long a human expert takes, that a model completes with 50% reliability — has run from under an hour in early 2025 to roughly half a day for the strongest models measured in 2026, doubling every four to seven months. That is "hours" reached and "days" approached, out of "hours, days, or weeks." METR's own caveats travel with the number and matter: it measures task difficulty and not how long an agent can act; measurements above sixteen hours are unreliable with the current suite; and the tasks are mostly software, ML and security work with clean success criteria, which is the narrow end of what a Nobel laureate does.

The self-grade. On 13 February 2026, asked on Dwarkesh Patel's podcast how his expectations had held up, Amodei's answer was "about as I expected it to go … there's plus or minus a year or two here", and he restated the timeline as "we'll get that 2026, maybe 2027. Again, that is my hunch. Wouldn't be surprised if I'm off by a year or two." In the same conversation he put 90% on a country of geniuses within ten years.

The independent grades, which differ from his and from each other. Zvi Mowshowitz (16 February 2026) reads the capability record as "going at roughly Dario's expected pace plus or minus a year or two, except coding is going faster than expected" — a partial endorsement — while recording that a separate dated Amodei prediction, that AI would be writing 90% of code within three to six months (March 2025, due roughly September 2025), was "clearly wrong", and concluding that "his predictions on impact do not square with his predictions on capabilities, period, and it is not a small difference." Jacob Brazeal (18 April 2026) grades it more generously still: "Dario's prediction that we might have powerful A.I. as soon as 2026 seems basically correct, give or take a year or two; the skeptics were wrong."

The finding the dates make on their own, which is the most useful thing in this entry. In October 2024 he wrote that powerful AI could come as early as 2026 — one to two years out. In January 2026 he wrote: "As I wrote in Machines of Loving Grace, powerful AI could be as little as 1–2 years away, although it could also be considerably further out", and the same essay's worked scenario has the country of geniuses materialising "in ~2027". The same relative interval is re-attributed to the earlier essay while the absolute date has moved forward by about the time that elapsed. Zvi's read of the February interview — that he is confident of the geniuses by 2028 — is the same drift measured a month later.

Two honest qualifications, and they are load-bearing. "As early as" is a lower bound with an explicit hedge attached, not a point estimate; the essay says plainly that it could take much longer. And "as I wrote in Machines of Loving Grace" is fairly read as pointing at the general claim that this could be very soon, rather than as purporting to quote a relative interval. The honest verdict is therefore not "he predicted 2026 and was wrong." It is: the earliest date in his own stated range has not been met with four months left in it; the range has been re-anchored twice, each time to about the same distance from the speaker; and he grades that drift as ±1–2 years of noise while a sympathetic independent reviewer grades one adjacent prediction of his as clearly wrong. A base rate wants both numbers, and they differ.

Claim 2 — the compressed 21st century: "50-100 years of progress in 5-10 years". Made October 2024. Due 5–10 years after powerful AI. NOT DUE. The clock has not started, and the leading indicators split cleanly in two.

The discipline first, because it is the thing most often got wrong. This claim cannot be scored in 2026 by anyone, including its author and including his critics. Its window opens when powerful AI exists. A reading that writes "the compressed 21st century did not happen" has scored a claim that is not due; a reading that writes "it is happening" has done the same in the other direction. What is available is a leading indicator, and it points two ways at once:

The front end of discovery has genuinely accelerated, and the record is dated. AlphaFold2's creators took the 2024 Nobel Prize in Chemistry, within weeks of the essay. Rentosertib — a TNIK inhibitor for idiopathic pulmonary fibrosis whose target was identified with AI and whose molecule was designed with generative AI at Insilico Medicine — published Phase IIa results in Nature Medicine on 3 June 2025, with the 60mg arm showing a mean forced-vital-capacity change of +98.4mL against −20.3mL on placebo. It is the first end-to-end AI-derived drug with clinical proof of concept.

The back end has not moved at all, and the back end is what the claim requires. Rentosertib entered Phase III on 7 July 2026 — a trial of 320 patients across 47 centres in China, which will take years. Target to Phase III has taken roughly six years for one molecule in one indication. Gary Marcus's rebuttal of 17 August 2026 makes the sharper version of the point: as of March 2026 no AI-designed drug had reached clinical adoption after more than a decade of effort; cancer is "more than 200 distinct diseases, all with different causes, biology and mechanisms"; and "you also can't simply skip the time-consuming clinical work." Andrew Maynard, writing on 13 October 2024 before any of this — the most valuable kind of independent grade, recorded before the outcome — called the 10:1 compression ratio "overly optimistic, especially when human behavior is thrown into the mix", and was sharpest about the mental-health section, on the ground that deciding what counts as illness is not a technical question.

The grade a reading can actually use: discovery is compressing and development is not, and the essay's own first limiting factor predicted exactly that. "Cells and animals run at a fixed speed so experiments on them take a certain amount of time which may be irreducible." Amodei wrote that, and then wrote a claim that requires it to be false at scale. Nothing since 2024 has shortened a Phase III.

The self-grade and the independent grade differ, and the shape of the difference is unusual enough to record precisely. On 16 August 2026 Amodei restated the claim unchanged and conceded that AI companies have not delivered what they promised, saying that only actually curing diseases would restore credibility. Marcus, the next day, agreed on delivery-so-far and rejected the forecast. They agree about the present and disagree about the future, which is a cleaner disagreement than most and is worth citing as such rather than as a fight.

Claim 3 — "20% annual GDP growth rate in the developing world"; sub-Saharan Africa to China's current per-capita GDP in 5–10 years. Made October 2024. Also not due — and the pre-condition data is available and is not moving.

Same conditionality: the window opens with powerful AI. Amodei flagged this section himself as the one he is least confident of. The current numbers are not a grade, but they are the base from which any such acceleration would have to start: the IMF has sub-Saharan Africa at 4.5% growth in 2025 and 4.3% in 2026, essentially unchanged across its April and July 2026 rounds, with the outlook clouded by commodity prices.

The rival forecast is the useful thing here, because it was made first and is equally dated. Daron Acemoglu's The Simple Macroeconomics of AI (NBER working paper 32487, May 2024, five months before this essay) puts the macroeconomic effect at no more than a 0.66% increase in total factor productivity over ten years — about 0.06% a year — and argues that even that is likely exaggerated because early evidence comes from easy-to-learn tasks. The gap between Acemoglu's ≤0.66%-over-a-decade and Amodei's 20%-a-year is the widest disagreement between two dated forecasts anywhere in this canon, and neither is due. A reading meeting an AI-and-growth finding should say which end of that gap it lands nearer, and should notice that both forecasters have been in print since 2024 with numbers three orders of magnitude apart on the same question.

Claim 4 — the entente strategy. Made October 2024, restated January 2025. DUE, because it is the one section not conditional on powerful AI. Going the other way.

The proposal was that a coalition of democracies could secure the supply chain, scale quickly, and block or delay adversaries' access, buying a clear if temporary advantage. Amodei restated it in On DeepSeek and Export Controls (January 2025), arguing that DeepSeek's results strengthened rather than weakened the case for controls. What has happened since:

• Chinese open-weight models are approximately 61% of tokens consumed on OpenRouter as of May 2026, the largest neutral router between models, and about 41% of downloads on Hugging Face. Alibaba's Qwen family passed three billion cumulative downloads in 2026, against Google's 418 million and Meta's 227 million.
• The controls leak. Two federal cases in March 2026 concerned Nvidia-equipped servers diverted to China through Taiwan and south-east Asian intermediaries, and the policy literature through 2026 broadly reads enforcement as weaker than expected.
• The layer moved. On 17 August 2026 Reuters reported that a platform accepting a stablecoin issued by a US-politically-connected venture hosts 43 of roughly 90 models from Chinese developers, including one under Commerce Department export restrictions — control at the chip layer routed around at the model layer via a payments rail.

Grade with care, because a proposal is not a forecast. "A coalition should do X" is not falsified by the coalition not doing X. The gradeable claim embedded in it is that securing the supply chain was feasible and would yield a durable lead, and that looks wrong so far: the lead in open-weight distribution has gone the other way, and the controls have been circumvented at three separate layers. Amodei's own confidence rating on this section — the lowest of the five — is the fairest thing to quote alongside the result.

Claim 5 — work and meaning, and the descendant prediction that is actually gradeable. Made October 2024; the descendant made May 2025, due 2026–2030. Not due; the disagreement about it is dated and real.

The essay's own claim — comparative advantage holds for a while and then does not — is too loosely specified to grade and the essay says so. The gradeable descendant is Amodei's May 2025 prediction that AI could displace half of all entry-level white-collar jobs within one to five years, due between 2026 and 2030. His January 2026 essay restates it and records the pushback: that critics accused him of a lump-of-labour fallacy, and that "some didn't see the 1–5-year time range and thought I was claiming AI is displacing jobs right now (which I agree it is likely not)." That is a real and creditable clarification, and it narrows the claim.

Zvi's February 2026 objection is the one to carry: "his predictions on impact do not square with his predictions on capabilities, period, and it is not a small difference." A reading meeting labour-market data in 2026 should note that this prediction is inside its window but not at its deadline, and that the forecaster has himself said current data is not the test.

Claim 6 — the marginal returns to intelligence, and the five limiting factors. Made October 2024. The part that has aged best, and it has aged best against its own author.

This is the claim with the strongest positive grade in the entry, and it deserves to be stated as one rather than buried among the misses. The essay's framework — that the interesting variable is not intelligence but what is complementary to it and becomes limiting — has been vindicated, in detail, by everything that has happened since. Every place the essay has fallen short so far is a place one of its own five factors bit: clinical trials running at the speed of biology (factor 1); AI-for-biology bottlenecked on the absence of the right experimental data rather than on reasoning (factor 2); regulatory and institutional pace (factor 4). A document that names the reasons it might fail, and then fails for exactly those reasons on exactly that list, has done something more useful than a document that turns out to be right about a date.

The criticism that survives is not that the list is wrong but that, as one reviewer put it, acknowledgment is not incorporation: the essay names the constraints in section one and then builds a timeline in sections two through six that requires them to be mild. That is the specific, quotable form of the charge, and it is more useful to a reading than either "he was right" or "he was hyping".

Commonly misused as

Not required for prediction. Included on the precedent superintelligence-2014 set, because this essay is already being quoted in all of the following ways and the project will meet every one of them.

"Anthropic's CEO says AI will cure cancer in five years."

The single most common misuse, and it is a dropped conditional. In the essay the 5–10 years runs after powerful AI arrives, not from the date of publication or the date of quotation. If powerful AI arrives in 2027, the essay's own window for most-disease runs to roughly 2032–2037. A reading that repeats the headline without the conditional has made the claim about twice as aggressive as its author made it — and should also note that Amodei's own August 2026 phrasing, "~5-10 years" stated flatly, invites the error even though the arithmetic still roughly works out.

"Amodei predicted powerful AI by 2026, and 2026 is here, so he was wrong."

The essay said as early as 2026, with the alternative that it takes much longer stated in the same sentence, in a document that calls all of its content guesses. The honest charge is not a busted point estimate; it is the re-anchoring documented in claim 1 — that the earliest date has moved forward by about the elapsed time while being described as unchanged. State the drift, which is documented, rather than the miss, which is not yet established.

"This AI-bio result shows the compressed 21st century is underway."

Available on every protein-design paper, every AI drug-discovery announcement, every lab's science product. The claim is not due and cannot be confirmed by an instance. What an instance can show is movement on a leading indicator, and claim 2 gives the split to apply: discovery-side results are consistent with acceleration; nothing yet observed compresses development, and development is where the 50–100 years actually sits. A reading citing an AI-bio finding should say which side of that line it falls on.

"It's a fundraising document."

The lazy dismissal, and it is wrong on the internal evidence even though the incentives are exactly as bad as section 2 records. The essay contains the best-organised list of reasons the vision might not arrive that anyone published in 2024, written by the person with the most to gain from it arriving, and that list has since predicted its own author's shortfalls. The disclosure in section 2 exists so that the incentives are visible; visible incentives are not a refutation, and treating them as one lets a reading skip the work of grading the claims.

"Cite the optimistic essay to balance a pessimistic reading."

The misuse this project is most likely to commit itself. Rule 3 requires the strongest honest AItopia case from evidence in the window. This essay is not evidence, in the window or out of it. Reaching for it when the week's findings are grim produces a reading that looks balanced and is not, which is worse than an honestly unbalanced one. The same applies in mirror image to terminator-1984.

"All watched over by machines of loving grace" as a straightforwardly utopian phrase.

Brautigan's 1967 poem, written while he was poet-in-residence at Caltech, imagines a "cybernetic ecology" where mammals and computers live together in mutual programming harmony — and whether it is sincere or gently satirical is genuinely contested among readers and always has been. Amodei links it and says nothing about which reading he intends. A reading that borrows the phrase for either irony or sincerity is supplying a tone the source does not settle, and should either say so or use different words.

The one this canon has to watch in itself

This entry is available on every AI-and-medicine story, every timeline story, every Anthropic story, and it always sounds well-read. That is the failure mode superintelligence-2014 named in itself, and here there is a second one stacked on top: grading Anthropic's CEO is a way for a model built by Anthropic to look scrupulous. The discipline is the same in both cases and it is strict. Cite this entry when a specific claim is at issue — a date, a multiplier, a growth rate, an export-control result, a restatement that moved — and name which of the six graded claims is doing the work. If a reading cannot name the claim, it should not cite the entry.

Sources

Primary, read directly:

• **Dario Amodei, Machines of Loving Grace: How AI Could Transform the World for the Better, darioamodei.com, October 2024.** Read in full across several passes. Source for: the powerful-AI definition and all six bullets; "a country of geniuses in a datacenter"; "it could come as early as 2026, though there are also ways it could take much longer" and the surrounding refusal to argue timelines; the compressed-21st- century formulation and the 10x discovery-rate claim; the biology and health list including the lifespan doubling to 150 and "escape velocity"; the 20% developing-world growth aspiration and the sub-Saharan Africa comparison; the entente strategy and the "eternal 1991" framing; the comparative-advantage passage; the five limiting factors quoted verbatim; the marginal-returns-to-intelligence framing; the rejection of "AGI" as a term; the doomer paragraph; the AlphaFold, AlphaProteo, CRISPR, mRNA and Player of Games references; and all four disclaimers quoted in section 1. Limitations, both material. The essay carries only "October 2024" as its date and the day is not established — secondary sources give both 11 and 24 October, the Internet Archive returned no snapshot for the URL when queried, and no reading should cite a day. And the text was read through a fetch-and-summarise tool rather than by eye, so quotations are kept to sentence length and the longest ones should be checked against the source before a reading leans on them.
• **Dario Amodei, The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI, darioamodei.com, January 2026** (posted 26 January per contemporaneous coverage). Read directly. Source for: "As I wrote in Machines of Loving Grace, powerful AI could be as little as 1–2 years away, although it could also be considerably further out" — independently corroborated by search against the exact phrase, and the single most load-bearing quotation in this entry; the ~2027 scenario date; the 1–2 years to AI autonomously building the next generation; the restatement of the entry-level-jobs prediction and the record of pushback to it; and the companion-essay framing. Fortune's coverage of 27 January 2026 is the secondary anchor for the date and for "we are considerably closer to real danger in 2026 than we were in 2023".
• **Dario Amodei, On DeepSeek and Export Controls, darioamodei.com, January 2025. Identified and characterised from the essay page and contemporaneous coverage (Simon Willison, 29 January 2025; LessWrong mirror). Not read in full**; the entry claims only that it restates the export-control case and argues DeepSeek strengthens rather than weakens it.
• **Anthropic, Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation, 12 February 2026, and Anthropic confidentially submits draft S-1 to the SEC, 1 June 2026.** Company announcements, used for the financial facts in the section 2 disclosure that are on the record from the company itself.
• **Anthropic, Claude Science, an AI workbench for scientists, 30 June 2026**, with coverage from CNBC, STAT and MIT Technology Review the same day. Source for the launch date, the sixty-plus databases, the phenylketonuria screening demonstration, the internal neglected-disease drug programme, and the note that Claude has not itself made a discovery.

Independent grades, which is what this kind requires:

• **Zvi Mowshowitz, On Dwarkesh Patel's 2026 Podcast With Dario Amodei, 16 February 2026 (thezvi.substack.com / LessWrong). Source for the "roughly Dario's expected pace plus or minus a year or two, except coding is going faster than expected" assessment; for the 90%-of-code prediction graded "clearly wrong"; for the geniuses-by-2028 read; and for "his predictions on impact do not square with his predictions on capabilities". The most useful independent grade found**, because the reviewer is broadly sympathetic to short timelines and grades a specific claim against its own date anyway.
• **Gary Marcus, No, Dario Amodei, we will not be curing cancer and "most human disease" in five to ten years, 17 August 2026.** Source for the clinical-work argument, the 200-distinct-diseases point, the causal-understanding objection, the "unrealistic hope" characterisation, and the claim that as of March 2026 no AI-designed drug had reached clinical adoption. Adversarial by disposition and labelled as such; the March 2026 figure is Marcus's and was not independently confirmed here.
• **Andrew Maynard, Is this how AI will transform the world over the next decade?, futureofbeinghuman.com, 13 October 2024.** The contemporaneous independent assessment, and therefore the most valuable one — judgment recorded before the outcome was known, the same standard superintelligence-2014 set. Source for the "overly optimistic" verdict on the compression ratio, the objection to the mental-illness section, and the deficit-model criticism.
• **Jacob Brazeal, Review: Machines of Loving Grace, 18 April 2026.** The most favourable independent grade located — "basically correct, give or take a year or two; the skeptics were wrong" — included because the independents disagree with each other and an entry that showed only the hostile one would be worth less.
• **Daron Acemoglu, The Simple Macroeconomics of AI, NBER working paper 32487, May 2024. Read as abstract and summary. The dated rival forecast for claim 3: no more than 0.66% TFP over ten years, with the argument that even that is likely overstated. Not read in full**; no claim is made here about its model beyond the headline figure and its stated caveat.

Evidence for what actually happened, read as reporting:

• Dwarkesh Patel's interview with Dario Amodei, 13 February 2026 ("We are near the end of the exponential"). Quotations — "about as I expected it to go", "we'll get that 2026, maybe 2027. Again, that is my hunch", "the country of geniuses in a data center. That's the starting gun", "the lack of public recognition of how close we are to the end of the exponential" — were taken from a third-party transcript, not from the audio. Treat the wording as close but not certified.
• The Amodei–Baker exchange on X, 15–16 August 2026. Read through coverage (Benzinga, CoinReporter, Cryptonomist, RamaOnHealthcare) and through Gary Marcus's reply, not from the posts themselves; Benzinga returned 403 and the posts were not retrieved. The facts used — that Baker argued the messaging has been disproportionately negative, that Amodei replied in two parts to wide circulation, that he restated the 50–100-years-in-5–10 claim, that he rejected the "disproportionately negative" characterisation, and that he conceded AI companies have not delivered and promised preliminary indicators within months — are consistent across four independent write-ups. The 14-million-views figure is reported, not verified.
• **METR, Task-Completion Time Horizons of Frontier AI Models (metr.org), read directly for the metric's definition and caveats, including that measurements above sixteen hours are unreliable and that the task suite is mostly software, ML and security work. The specific per-model horizon figures used in claim 1 came from secondary summaries rather than from METR's own leaderboard**, which did not render its numbers to the fetch; they are stated as approximate for that reason, and a reading wanting a number should take it from METR directly.
• Insilico Medicine / rentosertib. Nature Medicine Phase IIa publication announced 3 June 2025 (Insilico press release, EurekAlert, PRNewswire) with the +98.4mL versus −20.3mL FVC figures; Phase III initiation announced 7 July 2026 (PRNewswire, BioSpace, Clinical Research News), 320 patients across 47 centres in China, identifiers CTR20262475 and NCT07687459. Company announcements and trade press; the underlying paper was not read.
• Anthropic's acquisition of Coefficient Bio, about $400m in stock, reported 3 April 2026 by The Information and covered by TechCrunch, FierceBiotech and BioSpace; roughly ten people, founders previously at Genentech's Prescient Design.
• **IMF, Regional Economic Outlook: Sub-Saharan Africa, April 2026 and the July 2026 WEO update** — 4.5% growth in 2025, 4.3% projected for 2026.
• Open-weight distribution and export-control effectiveness, 2026. The ~61% of OpenRouter tokens, ~41% of Hugging Face downloads, Qwen's three billion downloads against Google's 418 million and Meta's 227 million, and the March 2026 diversion cases come from a mix of trade press and think-tank commentary (East Asia Forum, Chatham House, AI Frontiers, BISI). The token-share figure in particular is a single-source number that a reading should re-verify before quoting. The 17 August 2026 model-layer routing item is from Reuters via this project's own reading of that date; under rule 7 the project's own output is context here and not evidence.
• Anthropic's Series H and revenue run rate (about $65bn at roughly $965bn post-money; about $47bn annualised) as reported in late May 2026, CNBC and others. CNBC returned 403 and the figures were taken from search summaries and secondary aggregators; they are the least well-sourced numbers in this entry and are used only to establish the order of magnitude of the disclosure in section 2.

Not obtained:

• The essay's publication day. See above. The month is certain, the day is not.
• The X posts of 15–16 August 2026 in their own words. Every quotation from that exchange in this entry is at one remove. This is the entry's most important gap, because that exchange is the live citation occasion in section 2 and it will be cited by a reading within days.
• Any Amodei self-grade on the specific claims of this essay. He has restated the timeline (three times), restated the biology claim (August 2026), and conceded that AI companies have under-delivered — but there is no passage in which he says which of the essay's claims he now thinks were wrong. The prediction kind wants a self-grade set against an independent one; the self-grade available is "about as I expected", which is a grade of the whole and not of the parts.
• Isomorphic Labs' actual clinical status as of August 2026. Search returned directly conflicting accounts — an FDA clearance in January 2026 on one telling, no disclosed clinical candidate by mid-2026 on another — from sources of poor quality. No claim about Isomorphic is made anywhere in this entry, and establishing this would materially sharpen claim 2.
• A systematic scorecard of Amodei's forecasts as a set. Individual predictions have been graded by individual critics. Nobody found here has assembled them — 90%-of-code, entry-level jobs, powerful-AI dates, the disease claim — into one dated hit-rate. That would be worth a canon job of its own, and it is the thing that would most improve the base rate this kind exists to build.
• Anthropic's S-1. It is confidential. Every financial figure in section 2 is therefore either the company's own announcement or press reporting.
• Contemporaneous technical review of the biology argument by a working biologist in October 2024. Maynard is a scientist but not a biomedical researcher, and Marcus is a cognitive scientist writing in 2026. The most valuable missing source is a bench biologist's judgment recorded before any of this was gradeable — the same gap superintelligence-2014 and scaling-laws-2020 both had to record, and it recurs because the people best placed to grade a forecast are rarely the ones who write about it while it is still a forecast.