--- title: "Machines of Loving Grace, and the question biology isn't quite built to answer" publication: deslop.media byline: deslop.media editorial provenance: editorial_synthesis provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it." issue: 1 issue_title: The Scaling Bet grade: draft issue_version: 3 issue_version_date: 2026-10-11 piece_version: 1 first_published: 2026-10-11 reporting_current_through: 2026-10 canonical_url: https://deslop.media/feature/dario-amodei/compressed-21st-century/ cite_as: "\"Machines of Loving Grace, and the question biology isn't quite built to answer,\" deslop.media editorial, Issue 1 (Draft v1, October 11, 2026)" changes: https://deslop.media/issues/1-the-scaling-bet/changes/ flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags" --- The agent Dario Amodei is asking readers to imagine is not a chatbot reading PubMed. It is a virtual principal investigator, running lab robots, directing graduate students, and inventing new measurement techniques. In the framework essay he published in October 2024, *Machines of Loving Grace*, the Anthropic CEO is explicit about the reframing: > I'm talking about using AI to perform, direct, and improve upon nearly everything biologists do. The distinction matters because every prediction that follows depends on it. If AI is a data tool, it sits inside the existing apparatus and shaves time off literature search. If AI is a research agent that designs experiments, runs them through robotic hands, and reads out the results, the apparatus reorganizes around it. Amodei is making the second claim. The forecasts attached to it — cancer mortality and incidence cut by 95 percent, Alzheimer's prevented, most genetic and mental illness treated, human lifespan doubled toward 150 — are the predictions a reader is meant to hold against the mechanism, not against the slogan. ## The compression target The arrival window is "as early as 2026." The compression target — fifty to a hundred years of biomedical progress folded into five to ten — Amodei calls a "compressed 21st century." His confidence in the agent itself is high: on Dwarkesh Patel in February 2026 he placed roughly 90 percent probability on reaching what he termed "a country of geniuses in a datacenter" within ten years. The biology forecast assumes the agent shows up and then asks what it could plausibly do. The arithmetic anchors on platform discoveries. CRISPR gene editing, AlphaFold protein-structure prediction, mRNA delivery, CAR-T cell therapy, optogenetics, cheap genome sequencing — Amodei's argument is that these are the things that have actually moved 20th-century medicine forward, that they arrive at roughly one per year, and that they come from very small teams with extremely high returns to a marginal smart researcher. CRISPR is his exemplary case: the bacterial-immunity work that pointed at it was published in the late 1980s, and the Doudna-Charpentier 2012 paper that turned it into a tool came twenty-five years later. A faster reasoner, the argument goes, closes that slack. His guess is that powerful AI "could at least 10x the rate of these discoveries." ## The biophysical reply The single most substantive biology-side reply did not come from a major journal. It came from Niko McCarty, an Asimov Press editor who writes about biotech bottlenecks, in a November 2024 essay titled *Levers for Biological Progress*. The line that anchors his case is direct: > Many of the bottlenecks slowing biology today are biophysical, rather than computational. McCarty's argument is that what gates biology is not how fast a researcher can think but how fast living systems can be coaxed into giving up data. Each step in cloning a gene takes several hours to a day. Human cells divide about once a day, and diseases such as Alzheimer's develop over decades. Most tools for studying a cell's molecules capture static snapshots and miss the short-lived changes underneath. A virtual researcher that cannot pipette or sequence is, on this reading, a faster reader of papers attached to the same apparatus that always set the pace. Adam Marblestone, founder of Convergent Research, has made a more local version of the point on X: connectomes — the wiring diagrams of brains — exist for a fly and not much else, and the tools required to scale that work to a mouse, let alone a person, are currently bottlenecked by electron-microscopy throughput, not by ideation. Stephen Malina, a researcher who writes on Substack, has argued that whatever the capability frontier looks like, it will be jagged: superhuman in coding and theorem-proving long before it is capable in wet-lab manipulation, with Moravec's paradox keeping the physical end of biology hard. McCarty's frame splits biology into work where a smarter agent meaningfully compresses time and work where the apparatus around it sets the pace. The table applying it is this publication's: McCarty set bottlenecks in clinical trials, human disease and animal testing outside his essay, and rows such as animal models and clinical readouts extend his frame rather than report it. A virtual principal investigator, on McCarty's reading, fully harvests the cognitive work and inherits the timeline of the apparatus. The essay therefore reads less as refutation than as a sharper claim about which constraint binds. Amodei concedes this in the same essay, as he concedes most things — among his named limiting factors is "the speed of the outside world," and he writes that he is "very skeptical that we can get 100 years in 1 year." The disagreement is about how big the apparatus residual is once a smart enough agent gets to work on the apparatus itself: redesigning reagents, automating animal-model phenotyping, shortening organoid maturation by cell-engineering rather than waiting. ## The major-journal silence What is unusual, roughly two years after the essay's publication, is the absence of a sustained biology-side rebuttal in the venues that adjudicate biology. *Nature*, *Science*, the *New England Journal of Medicine*, and *Cell* have not produced a point-by-point reply. Derek Lowe, a longtime pharmaceutical-research blogger, has written directionally adjacent posts about AI in drug discovery without an essay-specific response. The biology-skeptic literature on Amodei's specific predictions is, on the public record, thin. The quiet has more than one reading. Biology journals are built to referee experiments, not forecasts about how a field's pace will change. The replies that exist ran where this kind of argument usually runs: Asimov Press, X and Substack. And many bench scientists may simply find the claim far from their daily constraints. ## The hedges Amodei built in Amodei has given himself more ways out than the slogan suggests. The opening framework names five hard limiting factors: the speed of the outside world, data scarcity, the intrinsic complexity of biological systems, human institutional constraints, and physics. The essay's most striking concession is short: > Everything I'm saying could very easily be wrong. The tests attached to the biology forecast are concrete enough to be checked. If "powerful AI" fails to materialize within roughly five years of the essay, the schedule slips beyond recognition. If it does arrive and the rate of CRISPR-class discoveries does not rise well above the historical one a year, against his guess of at least tenfold, the platform-tempo claim collapses. If cancer death rates, which Amodei notes have been falling about 2 percent a year, do not fall markedly faster in the years following deployment, the headline 95 percent number recedes into ambition. If aging biomarkers remain unsolved at the end of the compression window, the lifespan claim never lands. A second test sits upstream, in his scaling-curve claim, and is sharper. Amodei told Dwarkesh Patel in August 2023: > if you told me, yes, you trained your 2024 model, it was much bigger and it just wasn't any better, and you tried every architecture and didn't work, that's the explanation I would reach for. The "compressed century" forecast inherits that test. The virtual PI requires a research agent whose general capability has continued climbing. A scaling-law flatline at frontier scale, with architecture sweeps, ends the upstream argument and with it the downstream one. ## The open question The argument running between *Machines of Loving Grace* and *Levers for Biological Progress* is not actually about whether AI helps biology. Both pieces concede that it does and will. It is about what a virtual researcher inherits when the cognitive ceiling lifts and the physical floor doesn't. Amodei thinks the apparatus residual shrinks because a smarter agent redesigns the apparatus. McCarty thinks the apparatus is what biology mostly is, and intelligence does not compress it. The test is empirical and dated. The first deployed cohort of agentic biology systems is roughly 2026 to 2027 on Amodei's timeline. An early data point came on September 23, 2026, when Anthropic reported that roughly 950 Claude agents had identified a previously uncharacterized enzyme system, which human scientists then tested. Anthropic said it did not yet know what the system does. That is a narrow research result, not the autonomous wet lab or the CRISPR-class platform tool the forecast requires. The discovery-rate claim is checkable across the five years that follow. The cancer-mortality curve has decades of trend data and a measurable inflection point if one arrives. McCarty's essay points to what would weaken the biophysical argument: experiments that run faster, DNA synthesis that gets cheaper, and tools that measure cells as they change. The question the two essays leave is which constraint binds. Amodei has bet that intelligence is. McCarty has bet that the apparatus is. The next decade will settle that bet. So far the argument has run in Asimov Press, on X and on Substack rather than in the journals. ## Sources - Dario Amodei, *Machines of Loving Grace* — darioamodei.com/essay/machines-of-loving-grace (October 2024) - Dwarkesh Patel, interview with Dario Amodei — dwarkesh.com/p/dario-amodei (August 2023) - Niko McCarty, *Levers for Biological Progress* — Asimov Press, asimov.press/p/levers (November 2024) - Stephen Malina, *Riffing on Machines of Loving Grace* — an1lam.substack.com/p/riffing-on-machines-of-loving-grace - Doudna & Charpentier, *A Programmable Dual-RNA–Guided DNA Endonuclease in Adaptive Bacterial Immunity* — *Science*, August 2012 - John Jumper et al., *Highly accurate protein structure prediction with AlphaFold* — *Nature*, July 2021 - Anthropic, enzyme-system announcement — anthropic.com/news/claude-discovers-novel-enzyme-system (September 23, 2026)