Commentary

Argument: Reviewing Alexander Krauss's study of 761 major discoveries, almost all made possible by a recently developed tool or method, Conor Griffin of Google DeepMind argues that funders should back toolmakers and the databases AI depends on, such as the Protein Data Bank, and use AI to upgrade and invent scientific instruments.

Why read: Read it for a question to put to any AI-for-science plan: which instrument, method or dataset does it improve, and who maintains it? Griffin also gives four objections to Krauss, starting with survivorship bias, since most tools ever built presumably led nowhere. It runs on Google DeepMind's policy blog.

Argument: Jan Pisek of the University of Tartu points to a Tallinn court that fined plaintiffs for a filing citing studies an AI invented, holding users responsible for AI material they submit. Academia should do the same: authors answer for every reference, hallucinated citations are misconduct, and arXiv's new penalties are timely.

Why read: Read it for a standard borrowed from the courts: whoever submits AI output answers for it, whatever tool produced it. It is a short letter that leans on one Estonian case.

Argument: Jorge Bravo-Abad reads a week of 307 papers and argues that a model's most useful result is how it does on data unlike its training set. Name the shift a claim depends on (new chemistry, site, patient group, year or intervention), make that shift the test, and report it before any random-split score.

Why read: Read it for checks to copy: on chemistry held out of training, uncertainty-picked samples did no better than random ones; once related rock samples were kept apart, logistic regression matched deep learning. And when a model ranks experiments, test a low-ranked candidate too, or a hit cannot show the ranking works.

Argument: Magnus Lynch, a dermatologist, gave 200 biopsy-proven lesions to GPT-6 Astra, Claude Fable 5.1 and an Inception v3 network fine-tuned on public skin images. The CNN named the right one of six diagnoses for 125, the two LLMs for 84 and 85. He concludes a small, old, specialised network beats both.

Why read: Read it for the triage table: the LLMs kept 97 and 96 of 100 lesions that needed care but cleared only 32 and 22 of 100 harmless ones. The CNN cleared 76 at its preset threshold, where it also cleared 18 of 30 melanomas; set to match the LLMs' caution, it cleared 1 or 2. Each LLM ran once.

Argument: Ankit Maloo argues that this year's AI proofs share a pattern: models push every plausible idea to the end and a checker discards the false proofs. They win where a problem needs few key insights and few options per step, staying on paths humans drop after weeks without visible progress.

Why read: Read it for a rule you can use before a result appears: expect AI to finish work whose key leaps people already made, where only long, hard derivation remains. He also argues that failed attempts teach only when they are published. He works from the labs' write-ups and mathematicians' commentary, not new data.

Argument: Pietro Michelucci of the Human Computation Institute argues that AI safety proposals, such as AI executives' advice to the UN Security Council, focus on the model and largely overlook risks from how AI interacts with people and institutions. No component is fully trustworthy, so the whole human-AI system must stay safe when parts fail.

Why read: Read it for a method from human computation: hidden questions with known answers keep a running trust score, and anyone below a threshold keeps working but no longer sways the output. He would apply it to AI agents' access to tools and data, though he notes performance cannot always be checked against known outcomes.

Argument: NYU mathematician Tristan Buckmaster, who released a forced-Euler result with Levent Alpöge the night before OpenAI's Navier–Stokes announcement, argues that AI companies see researchers' half-formed ideas in chat sessions while competing with them using models the public cannot use. Academics, he says, need data protections and open-weight models of their own.

Why read: Take away a list to demand of an AI provider: zero data retention with audits, training only on opt-in data, no capturing a whole session when a user rates one reply, and, for credit, reward for exposition and no penalty for disclosing AI use. He is a party to the priority dispute with OpenAI; his account of its October release is his own.

Argument: Nature argues that tech donors, a group AI company listings could soon enlarge, bring useful freedom to science funding: long bets and fast pivots. But unlike public funders they are not obliged to report what their grants produced, and Nature says they should; without that, nobody can tell which approaches worked.

Why read: Read beside Ben Reinhardt's case for venture-style philanthropy, it supplies what that case leaves out: rounds and credit for early backers coordinate funders, but only published results let others learn from them. Its lesson for new donors is the Gates Foundation's Grand Challenges, slower to pay off than planned.

Argument: Four Metis Reasoning founders, Thore Graepel among them, trace hallucination, reasoning nobody can audit and knowledge nobody can edit to one cause: training fuses what a model knows with how it reasons. Their fix is AlphaGo's parts: an explicit record of what is known, a judge that admits a claim only on evidence, and uncertainty reduction as the objective.

Why read: Take away what it says a trace must record to be auditable: each claim's evidential status, the evidence that licensed the change, and why a withdrawn premise was dropped. That is a test any agent claiming auditability can be held to. The separation thesis (knowledge apart from reasoning, evidence as the gate) is set out in a page. A launch paper, with no system and no benchmark.

Argument: Ben Reinhardt argues philanthropy funds risky, startup-like nonprofits badly because it lacks venture capital's coordination tools: rounds, stage norms, credit for early backers, exits. The norm against funders claiming early wins leaves no reason to back risky work early. He proposes philanthropic SAFEs, shareable diligence reports, impact portfolios and syndicates.

Why read: Take away the diagnosis that philanthropy's gap lies in coordination and credit, and an appendix that prices research projects by stage, from about $100-500k for an idea to $50-200M+ for deployment. Reinhardt runs a research nonprofit that raises this kind of money.

Argument: Mihai Codreanu, who co-led Google's AI-in-science study, tells Annelies Gamble the constraint has moved downstream: scientists save about 7 hours a week, yet 44% say their main bottleneck moved, 41% report a growing backlog of untested hypotheses, and nearly half of those saving time spend over a quarter of it checking AI output.

Why read: Read it for three ideas: a verification tax that in lab and physical sciences is the experiment itself; a possible streetlight effect, with about half of respondents saying AI encourages safer projects; and his test for real discovery, longitudinal data on whether a result later proves useful to others and opens new directions, beyond the tasks AI already speeds up. Every figure is self-reported.

Argument: Alex Rives, Biohub's head of science, says today's so-called virtual cells cannot simulate a cell or replace experiments, above all because much inside a cell cannot yet be measured. Within about five years, Rives suggests, a model could predict the range of states a cell moves to after an intervention.

Why read: Read it for why protein models worked and cell models have not: proteins come with sequences from all of evolution, close to unlimited data, while cell data is scarce, so model design still matters and lab experiments must check the model. Rives speaks for Biohub's $1.8B data push.

Argument: Tim Fist, Georgia Adamson and Konstantin Pilz put the US lead over China's top models at about 6.5 months, set mostly by China's chip access and distillation of US models. Existing executive powers (export controls, an antitrust safe harbour against distillation, anti-smuggling checks) could, they estimate, roughly double it by end-2027.

Why read: Read it for a model that turns compute and distillation into months of lead, and its answer to the sales objection: with demand outside China above supply through at least 2027, restricted chips would sell elsewhere. It assumes equal capability per unit of compute, which the authors call a key limitation.

Argument: Jorge Bravo-Abad draws six rules from 60 papers on where to spend effort. Report how often a design works, not just the best: 36% of 135 designed zinc proteases cut their target. Try correcting an existing model as a baseline, and test as the model will next be used.

Why read: Read it for checks that carry across fields: in one evaluation AlphaFold 3's confidence scores ranked screening hits worse than docking, and a proteomics model trained on 100 million spectra lagged existing methods on much of a benchmark from older instruments. It is a reading digest, so each rule rests on a few papers.

Argument: Niko McCarty tallies an E. coli cell (two million proteins, 10 billion ATP to build another) and argues the list is not a recipe. No cell has been built from scratch, he suspects, because we lack its fault tolerance and regulation; builders should make many variants and let selection pick.

Why read: Read it for concrete cases of why an inventory is not an understanding: noisy gene expression in identical cells, a starving population rescued by about a dozen chance mutations, sensors that read fold-changes rather than amounts. It is a popular-science chapter with no AI in it.

Argument: Francesco Maggi argues that AI proofs are real, but where they lead is a choice. In the dark version output soars while knowledge gathers in a few companies and slips beyond human understanding. He asks which of OpenAI's 4,000 problems failed, and calls on universities, societies and funders to jointly build and govern competitive LLMs.

Why read: Read it for a mathematician's case that moves the debate from credit to infrastructure: if academics need proprietary models to compete with the firms that own them, those firms decide which questions get the best tools. He also passes on colleagues' sense that a strong result no longer shows a student's own ability.

Argument: A new mathematics journal planned by Shaoming Guo and colleagues sets two values beyond open access: how much a paper deepens understanding counts for more than being first, and a paper is accepted only once its proofs are formally verified in Lean, because proofs developed with LLMs can hide subtle errors.

Why read: Read it for one concrete way a journal could meet LLM-assisted proofs: accept a paper only after Lean has checked it, so a machine catches subtle errors before publication. The post is short and asks mathematicians to propose other journal designs.

Argument: Jeremy Avigad, Kevin Buzzard and five other formalizers argue that a formal proof library is worth something only if mathematicians can build on it. They warn against formalizing everything fast and badly (Claude's 13-million-line Fermat proof, they judge, probably has little lasting value) and back AI-written libraries that follow Mathlib's design and values.

Why read: Read it for a working model of AI formalization done for use: Tau Ceti and Formal Frontier, where AI agents write Lean code and other AIs review it, and Palomar, a registry where each project declares how much AI did. Both libraries are incubated by the Mathlib Initiative.

Argument: Hendrik De Bie, a Ghent University mathematician, checked all 152 new arXiv papers in two fields over one September week: 56% said nothing about AI, 35% reported using it for the mathematics. Every paper, he argues, should state its AI use, even none, and name the model and the theorems it touched.

Why read: Read it for a check any field can copy: count one week's new papers by what they disclose, then write to the authors. His recommendations double as a template for an AI methods statement: which model, which section, what role, and any funding or early access from AI companies. It covers two fields over one week.

Argument: At a Shanghai roundtable, four young researchers argue that AI proposes candidates far faster than labs can test them, so verification decides what a result is worth. Tong Xie, citing the corrected A-Lab paper, says the model that proposes a material must not also judge it, and the test must be fixed before the data arrive.

Why read: Take away Xie's three tests for calling something an AI discovery: someone other than the developer checks it, the prediction comes before the experiment, and the result lies outside what is already known. And the cost of testing sets the bar: a model whose predictions take years to test must state where it applies.

Argument: Nature Medicine argues clinical AI has outrun its evidence: over 90% of US health systems deploy AI tools, yet 53% or fewer find them highly useful. Benchmark scores do not carry over into care, so tools must be judged before, during and after deployment, by what they do for patients and staff.

Why read: Read it for a list of what goes wrong after a model passes its benchmarks: an eye clinic where the AI worked only when built into every step of care, a records assistant whose performance drifted, clinicians who stopped using a tool within four weeks. It introduces an issue; the cases are in the pieces it links.

Argument: Economist Gabriel Unger and computer scientist Arvind Narayanan argue that AI is being built almost entirely to replace human work (by economic incentive and by the habits of computer science) and set out an agenda for the other direction. First on it: white-box agents, which show their sources, assumptions and intermediate steps.

Why read: Read it for why that is harder in a lab: coding agents can show their work partly because public repositories record every step, and law, medicine and many sciences have no such public record, so collecting a field's process traces is itself infrastructure. Speed enough to work alongside, collaboration past taking turns, and scoring a human-model pair are the rest of the agenda. Science gets one line in the essay.

Argument: Drug-discovery chemist Derek Lowe reads a preprint that gave 30 language models text from 42 retracted, fraudulent or pseudoscientific papers and asked for a follow-up study: 95% of the answers they gave went along with the premise. Agents meant to mine the literature assume it is reliable, and it is getting less so.

Why read: Read it for a check any team building a literature agent can run: hand it a retracted paper and see whether it plans the next experiment anyway. In one example, a model cites a jailed surgeon's fatal tracheal grafts as successes. His line that no model reasons goes past what the preprint measured.

Argument: Ian Quigley, CEO of Leash Bio, argues that public binding data such as ChEMBL teach models who made a molecule, not how it binds: a guess at the chemist alone predicts activity as well as the structure does. So Leash measures binding itself, with pocket mutants that separate memorising from learning.

Why read: Read it for a check any team can run on its own data: if a model can guess who made a molecule, or which instrument ran the assay, it may be learning the lab rather than the chemistry. The figures are Leash's own, given in conversation.

Argument: Joseph Tooby-Smith, a Bath lecturer in proof assistants, says a conjecture settled by AI in Lean with little human help is no proof but a "warrant": belief without understanding, a dinosaur bone found by luck. Publish it, say what it is, and leave the work and credit of a real proof intact.

Why read: It gives a name, the warrant, to the gap between a result a machine has checked and one a person understands, a distinction worth carrying into any reading of AI-assisted proofs. What to do once a warrant is in hand he leaves to each finder.

Argument: Tingjun Hou, Chang-Yu Hsieh and colleagues rank drug-discovery AI systems on six levels, from fixed tools to a "virtual scientist" that sets its own questions. They call claims of autonomous discovery premature and ask agent papers to show the gain came from the agent design, not a stronger model.

Why read: Read it for a reporting checklist for any agent paper: test against a non-agent pipeline with the same model and tools, and report human interventions by severity, errors the agent fixed itself, sub-tasks completed and cost against an expert. Most of it surveys existing systems.

Argument: Ten years into the Human Cell Atlas, Nature Biotechnology argues the question is no longer whether human cells can be mapped but what the map is for. Having supplied much of the data behind today's foundation models, it now aims at 10 billion cells and AI that predicts a cell's response to a deleted gene or a drug.

Why read: Read it for where much of the data under single-cell foundation models came from: a decade of shared standards, metadata and benchmarks agreed across a consortium now over 4,000 members, groundwork the new virtual-cell funding assumes. Written for an anniversary, it lists aims for the next phase, not results.

Argument: Manish Ranjan argues that AI weather models learn from past measurements, so they forecast best where stations are dense and worst where they are sparse, widening the gap for poor countries. Chips get cheaper; an unmeasured storm is lost for good. Commercial AI forecasters should return a share of revenue to observation networks.

Why read: Read it for the point that measurement, not computing, is now the scarce input to AI forecasting, and that a free AI forecast can remove the budget case for the gauges that check it. It is a policy column; its figures are cited from others' studies.

Argument: Pushmeet Kohli of Google DeepMind says AlphaFold learned to reproduce structures deposited in the Protein Data Bank, not how proteins really move, and that DeepMind's own cell-model work found cell-by-gene data not yet good enough for a virtual cell. Sal Candido of Biohub adds that scaling pays only once the data carry what the problem needs.

Why read: The people who built these models say plainly that "solved" was narrower than the headlines, and that the next gains depend on collecting the data a problem needs, chosen as carefully as the model. It was said on a 30-minute panel launching a data initiative both their organisations back.

Argument: Alex Chalmers argues that AI produces the visible output of skilled work while skipping the slow practice that builds judgment, and that from outside a practice no one can reliably tell drudgery from formation. Institutions that reward only output risk slowly losing the ability to catch errors, train successors and raise their standards.

Why read: It speaks directly to research training, and the Terence Tao line he quotes puts it in one image: AI is a helicopter that skips the journey, and the journey is where the skill was built. It is a philosophical essay, so it offers caution rather than a way to measure the loss.

Argument: Before blaming a model, check that the test can see success. Jorge Bravo-Abad cites a cell-perturbation benchmark where half the treated cells, used to predict the other half, lose to a plain average on about 95% of perturbations, scored by mean squared error. Output counts as evidence only with a fitting check: a control, an assay, a proof checker.

Why read: A habit to take from it: before weighing a result, name the claim that would change a decision and the comparison that could establish it. The weekly reading list applies that habit to other groups' papers, each with its limit stated.

Argument: A Lean proof is only as trustworthy as Lean, argues Thomas Hales, a Pittsburgh mathematician who once found a soundness bug in the HOL Light prover. This summer AI found kernel bugs that let false statements be proved, all since fixed, and Lean's type theory still has no complete consistency proof written out by people.

Why read: It explains something most readers of AI proofs take on trust: the Lean checker those proofs rely on has theoretical foundations mathematicians have rarely examined. He welcomes an AI-built, verified kernel but wants people to write the theory and check what each statement says. The closing scenario of a planted bug is a thought experiment.

Argument: Daniel Ramirez of Recursion's Valence Labs argues that physics simulations and AI models fail in opposite ways at predicting how tightly a drug binds its target: physics is slow but holds up on new proteins, AI is fast but breaks without warning outside its training data. So each should feed the other.

Why read: Read it for a short history of the problem, from docking to free-energy calculations to cofolding models such as Boltz-2, with what each approach trades away. It is Valence's own blog, and its own step toward the combined loop, a model called AquaGen, has so far been used only on how molecules dissolve in water.

Argument: Asaf Karagila, a Leeds set theorist who had offered a bottle of whisky for this problem, looked over OpenAI's preprint claiming the Partition Principle does not imply the Axiom of Choice. He finds it muddled and oddly referenced, a desk reject, and says dropping hundreds of such "solutions" on experts works like a denial-of-service attack.

Why read: How the specialist this result most concerns reads it: going by the prose and references rather than the Lean code, he says the result cannot be read, not that it is wrong, and that he will not spend his time checking it.

Argument: An AceCamp investment note argues the AI4S boom rests on real progress, but only at the front of drug R&D: structure, binding and early design. Clinical success, generalisation to new targets and agents' reproducibility are unproven; valuations rose faster than scientific output, and the stock market rewarded synthesis and lab suppliers more than AI platforms.

Why read: It separates what is proven from what is priced and gives a watch list an investor can check: third-party blind hit rates, Phase II readouts, and how often an agent's result repeats. It ends by calling AI4S a constructive investment theme, citing bank research for its clinical figures.

Argument: Álvaro Lozano-Robledo, a UConn mathematician, writes to students rethinking a maths PhD: AI companies steer the story ahead of their IPOs, and AI proofs so far seem to recombine known ideas rather than add new ones. People stay needed, and the case for a PhD has not changed in a year.

Why read: It turns a worry into a prediction a reader can check: if AI only fills in the "convex hull" of known ideas, results found by AI alone should slow down. The model is a borrowed toy with one worked example, and the job outlook rests on universities' need for maths teaching.

Argument: Tom Beucler of the University of Lausanne and twelve co-authors argue that an AI weather model run for longer is still not a climate model: the record holds one warming path, so learned correlations can break in a climate nobody has measured. Their minimum: greenhouse gases, aerosols and land use must enter as inputs that can be moved one at a time, and the model must be tested outside its training range.

Why read: Read it for two tests to put to any AI climate result. Hold a forcing fixed: with CO2 pinned at its 1979 level, Ai2's ACE2-SHiELD warmed unrealistically little against the simulation it emulates, and its authors retrained it on varied CO2 to repair that. Then ask what the speed buys: the paper suggests AI is fast mainly by computing less.

Argument: Sven Krippendorf (Cambridge), Leopoldo Sarra (Axiomatic AI) and Joseph Tooby-Smith (Bath) argue that AI now writes computer-checked physics proofs fast, but a proof that checks cannot show its definitions match the intended physics. Unread code lands on shared libraries such as Mathlib and Physlib, whose reviewers lose the credit they once earned.

Why read: Read it for a split the authors say applies to coding in general: AI is reliable at local, checkable work, such as filling in a proof, and weaker where judgment spans a whole project, such as choosing definitions or designing a library. Tooby-Smith started and maintains Physlib, the library they want at the centre.

Argument: Fabio Motoki argues that AI might quietly make studies less independent. A 2026 preprint suggests different language models, given the same questions, propose narrower and more alike methods than published work; if teams lean on the same few models, ten papers may be fewer than ten independent paths. He wants AI use reported, and replications to check the tools too.

Why read: A useful point that evidence reviews can apply now: count shared tools the way one counts shared data. It is a short letter resting on one preprint, and it does not show how often alike AI advice has produced matching errors.

Argument: Pavel Etingof and Slava Gerovitch of MIT argue that when AI does most theorem proving, papers stop showing how good their authors are. Hiring and PhD exams should lean more on talks, oral exams and discussion, and value a theorem proved with heavy AI help the same as one proved without.

Why read: Read it for a protocol any research group could copy: students work at the blackboard without AI, the faculty member sends the notes to a model once after each meeting, and students check its reply before the next one. Etingof discloses part-time consulting for Anthropic since July.

Argument: Statistician Phil Kay argues that scaled-up self-driving labs risk more wasteful experiments and cost chemists both agency and basic skills. His remedy is the fix a UK report proposed for pocket calculators: teach and examine chemistry both with AI and without it, including manual practicals and pen-and-paper design of experiments.

Why read: Read it for a test to put to autonomous-lab results: what the runs bought. OpenAI and Ginkgo's lab needed 36,000 runs for one recipe; Alán Aspuru-Guzik's agent Óptima reasoned well about a zero yield, though a human-directed campaign reached a higher yield, in more than twice the runs. The column is sponsored by JMP.

Argument: Jeffrey Flier, former dean of Harvard Medical School, asks who gets credit if an AI agent finds a new way to cure cancer. Credit rewards scientists and assigns trust and responsibility. AI-driven discovery changes the first; the second still has to be done, and patents and prizes were built for human discoverers.

Why read: Read it for one distinction: credit as a reward, and credit as the record of who answers for a claim. It separates what AI-driven discovery changes from what institutions still have to settle. It mostly sets questions rather than answering them, and he says so.

Argument: Diego Córdoba and Luis Martínez-Zoroa, whose layered-amplification method the Alpöge-Buckmaster forced-blowup papers build on, place the LLM-aided fluid results in decades of human work. OpenAI made two claims, they note: forced Navier–Stokes and unforced Euler blowup; the first, if verified, meets one Millennium Problem option but is not unforced blowup.

Why read: From the authors whose method both OpenAI's paper and Buckmaster cite, the restraint is the signal: they will not compare OpenAI's construction with theirs before a detailed analysis. They hold their own result to the same line: it covers only weak, fractional dissipation, a substantial gap from real Navier–Stokes.

Argument: Taback is a tenured mathematician at a small liberal arts college. AI has not yet solved the problems she has worked on for years, but it has fixed a wrong lemma, suggested a reorganization of a paper, pointed her to directions she had not thought of, and improved her writing. With no colleague in her specialty down the hall, it also partly fills that gap. Her bigger worry is teaching: AI is pushing some assessment back to timed, proctored exams, and small colleges rarely have the graduate students oral exams need.

Why read: It gives a rarely written view: AI's effect on mathematics is also playing out at small colleges with fewer resources and heavier teaching loads. Research help, classroom assessment, tenure standards and AI disclosure are all changing at once. Taback wants researchers to say which AI tools they used, while still answering for checking, understanding and explaining the result.

Argument: Bravo-Abad says to look for scientific superintelligence in the loop, not in one model that knows more than a physicist. The test is a system that turns decisions about what to read, simulate or measure next into reliable new knowledge better than top human teams on the same budget, and that changes course when an experiment fails.

Why read: It gives a reader a better yardstick for lab automation: score the whole campaign, including setup, failed runs and human rescue, not prediction accuracy. It is an opinion piece built on one example, the 2020 mobile robot chemist.

Argument: Chatbots do not reason, argues Thore Graepel of the AlphaGo team: chain of thought is the same next-token guessing run longer, keeps no open record of hypotheses and evidence, and is often a story told after the answer. Science needs AlphaGo-style search over an explicit, checkable state of beliefs.

Why read: It doubles as a founder's case: he says it is why he left Google DeepMind, without naming Metis Reasoning, the startup Bloomberg reported he is raising for. The diagnosis is well made; the cure, a separate checker that updates beliefs only on evidence, is a design with no system or result shown.

Argument: Harald Helfgott argues that much of AI's apparent brilliance in mathematics fills an old, plain gap: the field never had a real search engine for results and techniques, and its subfields could barely read each other. Even if progress slows, he expects routine results to lose credit and the journal system to change sharply.

Why read: His most useful demand is a baseline: the published cost of AI results leaves out the failed attempts, so nobody knows the real success rate. The essay covers only the scenario where AI slows, and the main test of model reasoning he reports is one lemma he put to the models between February and September.

Argument: Euan Ashley, who led the first clinical reading of a human genome, gave Claude his own decade-old genome. In half an hour, for about $5, it re-found many of the key findings of his team's earlier analysis. Now that anyone can do this, he argues, reliability is the scarce part, and genome reading needs shared standards.

Why read: The right worry: a clean report can hide a gene old sequencing never read well, and a person alone at a laptop can learn of a disease risk with no clinician to ask. It is one genome checked against one unpublished analysis, so it shows the task can be done, not how accurately.

Argument: Harvard physicist Matthew Schwartz says today's models are good at science but are not scientists: strong on breadth, code and checkable maths, weak on deep concepts. Instead of forcing them into a scientist's role, give them "Claude-shaped" problems that a known method from another field would solve, and let domain experts say which results matter.

Why read: The useful part is the advice to lower the ambition and raise the yield, and his note that results were often correct but dull until an expert redirected them. The evidence is his own account of 36 manuscripts, still being verified, and he is a visiting researcher at Anthropic writing on its site.

Argument: Testing a model's top five picks shows it can find something, not that it should choose the next fifty experiments. Bravo-Abad gives each of five experiments a job: the best candidate, an unfamiliar one, what the lab would have picked, a predicted failure, and a repeat.

Why read: The predicted failure and the lab's own baseline are the two slots most groups skip, and they are the ones that can separate a model that discriminates from one that scores everything nearby as good. He says plainly that five runs still cannot show a general advantage.

Argument: Jorge Bravo-Abad draws a design pattern from Google's GenFocal climate paper: when simulated and measured examples cannot be paired, average the measurements onto the simulation's coarse grid. One model corrects the simulation's statistics there; coarse and fine versions of the real data then train a second model to add the detail.

Why read: The takeaway is a question for any field with unpaired data: is there a shared coarse level at which part of the problem becomes supervised? He names simulated microstructures and microscope images as a next case, and says plainly that this extension is untested.

Argument: As the FDA asks how to regulate generative-AI medical devices, Nature says tools whose output feeds clinical decisions should be tested in real clinics, as drugs and self-driving cars are. Of some 4,600 papers on clinical AI, only 23% used real patient data, and just 19 were prospective randomized trials.

Why read: The strong point is the missing baseline: a new kind of device has no decades of real-world data to compare against, and companies often test it only against doctors on one simulated case. It asks for principles, not a full trial for every tool, and does not say which device needs which test.

Argument: Nine mathematicians, after 600+ survey replies, tell AI labs: stop testing hard problems on models nobody else can use. If you do release results, post them in a neutral repository with prompts, cost and failed attempts, cite earlier work, and pay for humans to come to understand them. The community, not the labs, steers that work.

Why read: It moves the fight from "is the proof right" to "who must make it understood", and puts the bill on the labs. It is advice with no power behind it, and the group came out of OpenAI's own approach; whether any lab follows it is the test.

Argument: AI makes each scientist more productive but narrows what science explores: AI-assisted work covers 4.6% less topical ground. The authors blame incentives, not the tools, and ask funders to pay for new data, stop punishing field switches, and reward new ideas, not only output.

Why read: The sharpest point is the cost gap: AI predictions got about 100-fold cheaper in two years while new cohorts and monitoring stayed expensive, so the cheap questions get the work. The fixes are proposals with no test behind them. The 4.6% comes from Hao et al. (Nature 2026).

Argument: Hutchinson answers Braun's call to report AI use by contribution. Some use is a long conversation with a chatbot that shapes a scholar's thinking the way seminars do, and leaves no single piece of work to credit. He asks to separate AI helping make an output from AI taking part in forming the ideas.

Why read: A fair gap in disclosure by contribution, argued from the reading-group analogy. The letter gives no evidence on how common this use is and proposes no rule.

Argument: Three scientists read Anthropic's ART claim. Le Cong: the experiments are still in the queue; AI found something shiny and only the lab can say if it is glass or diamond. Seth Shipman: what is new is how it was found, not what it is. Jason Gill: his 2021 paper already named the enzyme; it is on Anthropic to show editing.

Why read: The most useful line is Cong's test: give an AI only pre-CRISPR knowledge and see whether it finds CRISPR. Until then ART is a pattern worth an experiment, not a gene-editing tool. It is a news report with quotes.

Argument: Webb says research is the search for something interesting to say, and LLMs do not change what counts as interesting. They change how fast she can work. Two reasons to keep doing math survive machines that do it better: she enjoys it, and it ties mathematicians together.

Why read: A personal position, not evidence: it says what one mathematician will choose to do, and says nothing about what AI can do. It does name the two costs to watch, the pull toward knowing the answer without understanding it, and asking a machine instead of a colleague.

Argument: Working from 134 survey replies, 32 theoretical computer scientists propose twelve near-term actions to keep human understanding central: disclose AI use in a methods section, have authors attest they understand and verified the work, cap conference submissions at five per author, judge junior hires on up to three papers, and train PhDs partly without AI.

Why read: Where the mathematicians' advisory group told AI labs what to do, this puts the rules on the field's own conferences, hiring and teaching, the levers theorists actually hold. It is a working proposal from a group the organisers say was not fully representative; whether a conference adopts it is the test.

Argument: A batch of only predicted winners cannot show whether a model recognises a loser. Reading a JACS polymer-transistor paper, Bravo-Abad notes it made one predicted low performer among five, and that the 76.7% test accuracy equals always answering low. Record each prediction with what was known when it was made.

Why read: It is a checklist a reader can apply to any materials validation: is there a control the model expected to fail, and were the predictions made before or after the material was characterised? By his own count this paper's test is still five polymers, chosen by the authors.

Argument: Former physicist Matt von Hippel challenged AI firms to compute a nine-loop amplitude in a toy particle theory on an academic budget. Claude did it with the field's known methods for about $1,000-2,000, and a Beijing group had most of it too. His lesson: there is more low-hanging fruit than experts think.

Why read: Read it for what one frontier calculation does and does not show: the AI ran the field's own recipe reliably with more compute but found no new method, so it settles nothing about superintelligence. Lance Dixon, who checked it, adds how being scooped feels. Anthropic paid for the post.

Argument: Seven Google DeepMind researchers, writing for themselves, argue that once AI makes ideas nearly free, what limits AI scientists is the lab time and money to test them. They propose a market: agents stake credits on which ideas deserve an experiment, license them to labs that run it, and get paid when it holds.

Why read: The useful part is the diagnosis: progress is likely to be jagged, fastest where a computer can check the result. The market (tokenised ideas, smart contracts, signed instrument data) is a design on paper, and the authors concede it would miss theory that pays off over decades.

Argument: Anthropic's five reported misuse cases do not show bad intent: editing viruses is also what ordinary research does, and banning it leaves us less prepared. Designing a pathogen is hard, so Lowe worries more about chemistry.

Why read: A working chemist's prior, not evidence: he has only Anthropic's summary of the cases, and says outright that a company also gains by showing how powerful its model is. His test for telling the cases apart is one a reader can use.

Argument: Fantl lines up the labs' own usage figures (Anthropic's 26% of AI R&D now AI-led, OpenAI's 1.6x more experiments a day) and warns the sources are paid to promote AI use. The squeeze he sees is at publishing: ICLR submissions went from 11,603 to 19,525 in a year. His fixes: replication at desk review, review by frontier models, costlier submissions.

Why read: The clearest case this week that output is outrunning checking, and right not to plan around vendor telemetry. But the publishing half leans on Pangram, a detector he flags as biased toward finding AI, and the fixes are a list of ideas, none tried or costed.

Argument: Sanborn, CEO of a genetic-medicine startup, argues AI has made scientific thinking cheap while experiments stay slow and expensive, and the underrated shift is not AI-run "foundries" automating measurement, but AI-run "navigators" spending that cheap thinking on better decisions. His own company triaged 500 disease targets with LLM research agents in two passes; the deeper second pass alone, on the 100 survivors, would once have taken about a century of expert reading.

Why read: The example is concrete and quantified, and he names the real gap himself: analysis built on demand has no versioned pipeline behind it, so a figure's provenance is weaker than the old workflow's. What he doesn't examine: "navigator" is also the strategy his own company runs, so an argument that it deserves more attention than the flashier "foundry" model conveniently favors the choice he already made.

Argument: Sahai argues AI systems are now producing mathematical ideas faster than any person can absorb, and the response should not be to cede understanding to them: humanity needs a far larger "deployable intellectual reserve" of mathematicians who can spend a year, in groups, understanding a single AI-generated breakthrough before it is acted on. His test case: before building an AI-designed fusion plant, humans, not just the company building it, must understand why the design works, not only that a model verified it.

Why read: The fusion-plant case forces an abstract worry into a decision a reader can actually weigh, and Sahai concedes the strongest objection himself: an AI system might handle verification "more reliably" than any human ever could. What he doesn't offer, and a commenter pressed him on directly, is any mechanism for funding or building the "intellectual reserve" he calls for; the piece asks for a choice, not a plan.

Argument: Aghayeva traces two centuries of biological theories (genes, enzymes, ion channels, the genetic code, chemiosmotic theory) that were worked out and believed before they were confirmed experimentally, to argue theoretical biology has been as generative as physics without ever getting equivalent institutional standing. She reads AlphaFold2 as the newest version of the same divide: it predicts protein structure accurately from statistical patterns learned from data, with no mechanistic theory of folding behind it, which she says sharpens the split between prediction and theoretical explanation rather than closing it.

Why read: Every earlier case she cites names a specific predicted mechanism later confirmed by experiment, a real pairing she could use to test whether AI's statistical representations are doing anything similar. She doesn't: the AlphaFold2 paragraph ends by calling the question of what its learned representation understands "open," and the essay moves on to the field's institutional history instead of pushing on the one claim that bears most directly on this site's own subject.

Argument: CUNY mathematician Alexander Gamburd says OpenAI's Navier–Stokes release offered a 166-page proof, certified by 616,000 lines of Lean, that no one had yet said they read in full. The danger is the day journals and prizes accept such proofs; mathematicians should count nothing no human understands, and side with the labs that respect that.

Why read: The clearest case on the record that the decision lies with whoever accepts a proof, not whoever produces it. But it argues from history and analogy, not evidence, and by its own account the unreadable proofs it warns of have not arrived yet.

Argument: Ihor Neporozhnii of Recursion's Valence Labs argues that drug-discovery models now train on millions of DFT and free-energy calculations, run with tools built for experts who checked a few runs by hand. Those tools fail without an error message, so pipelines need automatic physics checks.

Why read: Read it for checks you can run on DFT and free-energy data: forces on a molecule should sum to zero, calculations should converge, neighbouring free-energy windows should overlap. He cites unphysical forces in up to 58% of calculations in some public DFT datasets, and concedes the largest error, a wrong protonation state, escapes automatic checks.

Argument: Nature Computational Science argues that research built on large language models is hard to reproduce: outputs are sampled and shift with a prompt's wording, and closed models are retrained or retired without notice. Papers should report the exact model version, full prompts and settings, and why a closed model was chosen over an open one.

Why read: Read it for a short reporting checklist that authors and reviewers can apply today to any paper relying on an LLM, and for the plain point that a result tied to a retired closed model may never be checked again. The editors call it a start to discussion, not field-specific guidance.

Argument: Gil Kalai, a Hebrew University combinatorialist, leans toward the enthusiasts: he would not slow AI-assisted mathematics or cut ties with the academic and commercial groups pushing it. He lists 18 jobs AI might do and dwells on simpler proofs: an AI-based proof of the sharp thin-shell inequality is, as far as he knows, simpler than older proofs of weaker results.

Why read: A useful counterweight from a senior mathematician: it asks which jobs matter, not only whether AI can prove theorems. But it is a personal post resting on a few anecdotes, and he says himself it is hard to tell where AI and maths stand and his optimism may be wishful thinking.

Argument: Israel argues AI can sharpen a centralized research system's resource allocation and experimental throughput, but cannot solve the prior problem science actually needs solved: knowing in advance what's worth pursuing. In fields with clear targets (materials, energy, pharma), China may accelerate sharply without producing paradigm-level breakthroughs.

Why read: The knowledge-problem argument is sharp, and its "accelerated conformity" scenario names a real mechanism: shared models and data homogenizing what gets tried. But the 50/40/10 scenario probabilities are asserted, not derived, and the piece never tests whether China's AI-assisted centralized labs already trail decentralized ones on actual breakthroughs, only forecasts that they will.

Argument: Bravo-Abad reads the Virtual Biotech paper's methods and 89-page supplement rather than its org chart, arguing the system's auditability rests on specific engineering choices (isolated per-task context, typed tool contracts, status-conditional output validation, and agent-run statistical checks), not on model quality.

Why read: The read is sharp and reusable outside biology, and its arithmetic catch (a reported 184x speedup that its own numbers imply is 307x) shows real scrutiny. But every case study is n=1, with no repeated-run stability check or reviewer-removed ablation, so scaffold versus model quality stays untested.

Argument: Reinhardt, who founded and runs a research organization built on this premise, proposes a scale-free model for healthy research institutions: a constant low-level "bubbling cauldron" of small, researcher-directed exploration, paired with "transient strike teams" that leadership assembles quickly once a direction is ripe, then deliberately disbands. He argues today's institutions inherited the bottom-up half from mid-20th-century science but lost the mechanisms, and often the willingness, to form and dissolve strike teams fast. His running example is DeepMind: small, self-organizing groups by default, with AlphaGo and AlphaFold as leadership-assembled strike teams pulled together once the moment was right.

Why read: The fractal claim (the same cauldron/strike-team split recurring from national research systems down to a single PI's lab) is a useful compression, and separating strike-team transience from the general "should scientists be managed" debate names a real failure mode. But the model is explicitly built by pattern-matching cases and "leans pretty heavily on metaphor", with no operational test for the one judgment call it says is scarce: knowing when a cauldron is ripe. The DeepMind example, his only AI-native case, is asserted from the outside: he flags that he doesn't know whether the pattern still held after the Google Brain merger.

Argument: Dolgin surveys Co-Scientist and rival research-agent platforms already generating hypotheses and designing experiments across labs, arguing that as AI absorbs this work, the scarce resource becomes human judgment: knowing what question to ask and verifying what comes back.

Why read: It surveys where hypothesis-generating agents already work in labs and makes the case that human judgment becomes the scarce part. Its own flagship example, Co-Scientist's condensate-gluing proposal for MYC, has not yet been tested in a dish, so how good the hypotheses are is still open.

Argument: Loh proposes math publicly adopt a "human flourishing axiom": AI's opaque decision-making will create more oversight "control points" than people can fill, so a community of frontier researchers must stay active to keep steering it, not to out-compute it.

Why read: The chain leans on unverified anecdotes (a mass agentic hack, a breached OpenAI repo) to argue AI growth will self-limit, then jumps from needing skilled security overseers to needing frontier mathematicians specifically, without showing why auditing requires research at the frontier rather than fluency behind it.

Argument: Frey, who leads drug-discovery research at Anthropic, frames the biomolecular release as models improving the stack they run on: an internal research model rewrote the inference paths of 30-plus open biology tools, built FlashPairformer kernels for the triangle operations that dominate AlphaFold3-class runtimes, and opened a "Big" mode that folds machines above 10,000 tokens on one node. He reads it alongside Anthropic's own measurement that Claude leads 26% of its AI R&D work.

Why read: The framing earns its place: an agent that rewrites the tools changes the unit cost of everyone else's science, and unusually for a lab claim, the code is open and benchmarked against NVIDIA's, so outsiders can rerun it. The self-improvement reading is the looser half: kernel work is a well-posed engineering task with a compiler for a referee, which is exactly why it went fast, and the protein-design economics remain an in-silico score until Adaptyv's wet lab reports.

Argument: Koukoulopoulos argues frontier AI concentrated in a few private companies threatens research independence: unpublished ideas can leak into systems researchers don’t control, and internal models let a favoured few outpace everyone else, as when an internal OpenAI model produced the Navier–Stokes result. His fix: a publicly funded “CERN for AI-assisted science,” with free baseline access for every researcher.

Why read: Turning an abstract fairness worry into a dated harm (Julia Stadlmann’s years of work eclipsed within days by better-resourced AI) reframes the debate from whether to use AI to who controls it. Weaker on feasibility: whether publicly funded models could stay near the frontier is conceded, not resolved.

Argument: A neural network keeps no record of what it learnt or from where, so a result's hints can become untraceable, and Nature argues that trail is part of the scientific process. It asks AI companies to disclose data use, default to opt-in, and submit to independent audit.

Why read: The diagnosis is the strong part: attribution is usually treated as etiquette, and this treats it as evidence a claim can be audited. It is thinner on enforcement, since opt-in and independent audit are asked for without saying who would compel them.

Argument: AI tools for IPCC literature review can change which evidence is visible, how studies are grouped and what reaches authors before anything is written, even when the tools are accurate. The IPCC should disclose task-level AI use, keep outputs traceable to original sources and name who verified each assisted step.

Why read: It moves information loss upstream of the model’s reasoning: retrieval, filtering and summarization can reshape an assessment, and formal human sign-off does not make those steps visible.

Argument: A correct but ad hoc, unexplained AI proof is technical debt: the answer is the small part, and the work of understanding is pushed onto everyone else.

Why read: Read it with Kra: she redesigns incentives, while Cohn puts the responsibility on whoever publishes an AI result to talk to experts, interrogate the model and explain the mechanism.

Argument: Antieau splits mathematical practice into slow math, personal understanding, and fast math, large AI-assisted programs staffed to also produce readable textbooks. He proposes standards: disclose all LLM collaboration, never post an LLM-drafted draft, and stop counting papers when hiring.

Why read: Read it with Kra: he separates whether mathematicians should use LLMs from whether a paper must stay human-authored, and answers only the second, giving checkable rules her incentive critique leaves abstract. Untested: whether a funded fast-math program actually delivers the textbook.

Argument: Mathematical research is the search for the simplest explanation: clarifying language, the right level of generality and theory that lets later mathematicians absorb results quickly.

Why read: The most constructive answer in the series to what remains once AI can prove theorems: theory-building is the part a proof certificate does not deliver.

Argument: Theorems were a proxy for understanding because they were scarce. AI has broken the proxy, yet hiring, publication and prizes still reward output, and peer review cannot keep up with the volume.

Why read: Read it for the concrete fixes: credit discovery, proof, formalization and explanation separately, and ask authors to explain how a proof works. It is an evaluation-design argument, not a lament.

Argument: Standard lab interfaces make instrument control easier, but API access does not give an agent a model of physical state or safe operation.

Why read: Read it for the Genentech foam failure: the software call was correct while the experiment was physically wrong. It makes the next bottleneck unusually concrete.

Argument: Science optimized for machine-readable output may accelerate discovery while making its explanations and control less accessible to people.

Why read: Useful because it treats interpretability, corporate control and biological black-box data as one institutional problem rather than three technical side issues.

Argument: Lowe reads the published AIntibody challenge results (511 AI-designed antibodies from 29 organizations, blind-tested against the same SARS-CoV-2 receptor-binding domain) as the real state of AI antibody design: strong at affinity-maturing a known antibody (the best submission gave a roughly 2000x affinity gain), but on the two harder tasks, ranking clones from a library and designing new binders from scratch, most submissions did no better, and sometimes worse, than picking sequences at random.

Why read: The benchmark itself is the kind of evidence this site keeps asking for: blind, prospective, experimentally validated, and Lowe links the primary paper rather than repeating a company's own number. His conclusion still favors AI: he calls antibody design "the best AI case right now," which makes this a comparative skepticism, not a verdict that AI drug design fails. He doesn't explain why affinity maturation transfers so much better than the other two tasks, only that it does.

Argument: Lila's profile of Kenneth Stanley's nine-person open-endedness team argues that machine learning told to optimize one goal can get stuck on a low peak, while science needs divergent search that keeps producing interesting results. Stanley calls that search core infrastructure for scientific superintelligence, and mission critical at Lila.

Why read: It asks the right question of autonomous labs: is a system that optimizes a set objective discovering, or only optimizing? But this is Lila's own profile with no result yet. The test is the papers planned for this autumn, and an open-ended run in its labs finding something a fixed objective would have missed.

Argument: A useful AI scientist needs non-obvious analogy, productive disagreement and a model of environmental dynamics, not only fluent reasoning in language.

Why read: Zou shifts the question from whether agents can execute a workflow to whether they can generate the kind of disagreement from which new science begins.

Argument: AI scientists should be evaluated in hidden, interactive worlds where evidence can change hypotheses, not with static question-answer benchmarks.

Why read: One of the clearest proposals for what to build next: limited budgets, real feedback and traces that reveal whether an agent actually updates.

Argument: Pat Walters argues that careful benchmarking now costs a prompt and a night. He handed Claude the preprints of three new chemistry foundation models and had it test them on two ADMET datasets, each measured in one lab, under his own statistical protocol. Monroe, paired with the tabular model TabPFN, came out best on most endpoints.

Why read: Read it for a template a reader can copy: data measured in one lab, with realistic ranges, cluster-grouped 5x5 cross-validation, ties shown instead of one bolded winner, and a check that pre-training never saw the test labels. Code and prompts are public; it is one run on two datasets.

Argument: When an agent can install unfamiliar code and rerun models under identical data splits and statistical protocols, new model claims face a cheap, repeatable external benchmark instead of relying on paper-reported SOTA.

Why read: The rankings matter (Monroe + TabPFN won all six Biogen endpoints, while MEGA-CL and Mol-JEPA led none of the nine ExpansionRx endpoints), but the durable point is the workflow. Work that used to cost days per model became a few prompts and an overnight run, putting structural pressure on self-reported SOTA.

Argument: AI may compress target and molecule discovery, but toxicity, pharmacokinetics, trial design, recruitment and regulation still return evidence on year-long clocks.

Why read: It turns the promise of faster biomedicine into a sequence of named stages. That makes clear which clock a model can shorten and which clocks it cannot yet touch.

Argument: Gustau Camps-Valls, Veronika Eyring and 18 co-authors argue that AI is pulling weather forecasting and climate projection, long separate fields, onto shared methods. They ask modelling centres to release open, differentiable model parts within three to five years, and CMIP, ECMWF and the WMO to set joint weather and climate benchmarks that test AI and physics models to the same standard.

Why read: Read it for the line it draws: a weather forecast can be checked against the weather that came; a century projection under an emissions scenario cannot, so an AI trained on the past is tested there only indirectly, through hindcasts and past climates. The authors concede physics models still beat AI on record-breaking extremes.

Argument: Plausibility is not enough in experimental science; automation needs inspectable methods, constraints and provenance because mistakes enter the physical world.

Why read: Rickerby actually let an AI choose work down to reagent and well. That makes this a report from the boundary, not a forecast about it.

Argument: Strong benchmark scores do not yet transfer to open-ended research where problem choice, judgment and changing direction are part of the task.

Why read: Their method matters more than the headline: inspect how the number was produced, how much scaffolding contributed and what the agent chose without help.

Argument: Answering a post that says AI will flood biotech with cheap drug candidates until clinical trials become the bottleneck, Derek Lowe argues the expensive risks are unchanged: the wrong target and unexpected toxicity, which usually surface only in human trials. More candidates will more likely raise the bar for entering the clinic.

Why read: Read it for a drug-discovery chemist's account of where the money is lost: ideas were rarely scarce, the hard steps are the therapeutic hypothesis, optimisation and clinical risk, and on that last one he says AI can at present do very little. It is a blog reply to one post.

Argument: Claims that AI is transforming productivity remain dominated by anecdotes, while rigorous evidence is sparse and unusually hard to interpret.

Why read: After the withdrawn MIT preprint, this becomes an editorial standard for every spectacular productivity number: ask what was actually measured.

Argument: Dan Turner-Evans of the Institute for Progress argues that the Genesis Mission's 26 challenges are too many to fund or to explain. The Department of Energy should either pick a few with measurable targets or put the platform first, and then weigh what each new model, autonomous lab or computer adds beyond industry's.

Why read: Read it for a plain test of where public AI-for-science money still adds something when startups and industry outspend it: data from unique instruments, basic science, classified work and grid-scale coordination. Written in July, before the Phase II awards, it is a yardstick for the twelve awarded in October.

Argument: Emmanuel Jeannot of Inria argues that agent-run science turns research from a craft into a pipeline, with the Genesis Mission as its largest case. His main worry: when one family of models proposes, runs and checks the work, errors stop cancelling out, and flawed results are trained into the next models.

Why read: Read it for a question you can put to any closed-loop lab or agent pipeline: is the step that checks a result independent of the step that produced it? He also points out that Genesis leaves climate science out of its priority areas. It is an essay, with no new data.

Argument: Public commitments to responsible AI can become ineffective when internal incentives make dissent costly and external accountability weak.

Why read: Its value is specificity, not neutrality. Turner separates public evidence from private exchanges that readers cannot independently verify.

Argument: Rafael Ferreira da Silva of Oak Ridge and 27 co-authors score their own year-old roadmap: agent software moved fastest; shared data, verification and governance stayed open. No system has yet made a new discovery and confirmed it in its own experiments, they argue, so checking results limits autonomous science more than model power does.

Why read: Read it for a milestone-by-milestone scorecard of autonomous labs, and for a distinction worth keeping: verification asks whether a workflow ran correctly, validation whether its result matches reality, which is why the strongest claims so far come from mathematics and computation, not chemistry or biology. The scores are the authors' own and provisional.

Argument: Michael Santoro argues that AI has made the cheap part of drug development cheaper. In a BCG analysis of about two dozen AI-found molecules, they pass Phase 1 80-90% of the time but Phase 2 at the usual 40% or so: picking the right target takes disease biology mapped only in fragments, not chemistry.

Why read: Read it for the investor's version of the discovery-versus-efficacy split: most of the $2.6 billion per approved drug goes to trials and their failures, so the bigger prize is AI that validates targets, predicts who responds or designs trials. He also asks if the Orphan Drug Act's cost test still fits.

Argument: Corin Wagen estimates that drug-like small molecules carry about 2.2 bits of information per heavy atom, against 0.5 for proteins and 0.1 for oligonucleotides. That, he argues, is why they are hard to make and to model: protein contacts reuse a few chemical groups, small-molecule contacts draw on almost endless ones.

Why read: Read it for a rough but clear reason why protein-binder design took off while small-molecule affinity prediction lags, and what it means for data: OMol25 took over 100 million DFT calculations before models seemed to match physics-based methods, so he doubts tens of thousands of new complexes are enough. He runs Rowan, a computational-chemistry company.

Argument: Scott K. Johnson argues that machine learning is changing weather forecasting but not yet climate modelling. A weather model learns from past snapshots, so it runs fast but underplays record extremes; climate questions concern conditions never observed, so CliMA, a project that Caltech's Tapio Schneider is part of, keeps the physics and uses machine learning only inside parts of the model.

Why read: Read it for one question that sorts most AI weather and climate claims: does the question lie inside the training data or outside it? It also shows where machine learning is in climate models today: tuning parameters and modelling snow cover, with emulators of costly components still being built and clouds mostly out of reach.

Argument: Laura Luebbert argues that biology agents are held back not only by reasoning but by databases built for people clicking in a browser. On 120 viral-sequence queries to NCBI Virus, agents scored 17% to 91% and gave different answers to the same prompt; given gget virus, a fixed retrieval tool, every agent passed 90%.

Why read: Read it for a worked case of a retrieval slip changing the science: one agent's incomplete Ebola sequence set moved the outbreak's inferred start from January 2014 to 1922. The rule it leaves is to keep the data layer deterministic and let the model be creative above it. Her team built gget virus.

Argument: Eric Topol of Scripps Research argues the strongest trial evidence for AI in medicine is in reading images: in a trial of over 100,000 women, radiologists with AI found about 30% more breast cancers. Its biggest gain to come is preventing cancer, heart disease and Alzheimer's by reading each patient's full data years before symptoms.

Why read: Read it for a clinician's line between what trials have shown (image reading, AI scribes that cut burnout) and what is still a hope (preventing disease by running large health models over each patient's pooled data). He also warns that hospitals may spend the time AI frees on more patients per hour rather than longer visits.

Argument: The Broad Institute's Anna Greka argues AI already speeds drug design and is starting to help clinical trials, so the bottleneck is finding the target: predicting how human cells respond to a disease mutation. To make the training data, she would engineer thousands of genetic-disease mutations into human cells, measure them, and look for shared nodes linking seemingly unrelated diseases.

Why read: Read it for a worked case: her lab traced a family's inherited kidney failure to cargo receptors trapping a toxic protein, used AlphaFold to design a drug candidate that worked in cells and mice and is heading to trials; she estimates the same receptors could be a target in over fifty diseases. Finding that target took more than five years of lab work, not the cell-prediction model she proposes.

Argument: Antonio Orvieto of the ELLIS Institute Tübingen answers his title "perhaps not yet": on DNA and RNA regulation tasks, pretrained genomic models showed no clear edge over methods from the past decade that use far less compute. He argues the field should stop borrowing architectures built for text and design models around genomic data.

Why read: Read it for a machine-learning researcher's case against scaling as the default in biology, and for what he would build instead: models with memory that track state along long DNA sequences, and hard benchmarks designed with biologists. It cites one comparison study and offers no new results.

Argument: MIT chemical engineer Connor Coley argues AI has sped up molecule design, so the bottleneck is now making and testing molecules, his "execution gap". Today's autonomous labs are "fenced-in explorers", reliable only inside narrow search spaces; widening them needs better synthesis planning, sensing, failure-tolerant orchestration and more dexterous robots.

Why read: Take away his yardstick for an autonomous lab: not experiments run or data per day, but how much faster it learns and how many dead ends it avoids. It is a short essay with no new data, and he notes the field still lacks the benchmarks that drove protein structure prediction.

Argument: MIT's Tess Smidt argues physics has no single interface for AI: a LIGO detector, an LHC trigger and a crystal simulation differ in data, infrastructure and timescale, so each working AI tool is built by hand for one workflow. Machine-learned atomic models also inherit the errors of the DFT data they learn from.

Why read: Read it for how physics checks AI (LIGO validates a model offline before using it, the LHC keeps a small sample of unfiltered collisions to catch bias) and what blocks reuse: glue code a few people hold, instruments built over decades against software that changes in months, symmetry-aware networks that run slowly on GPUs.

Argument: RLVR works in mathematics and code because verification is cheap and reliable; most important scientific questions lack that condition.

Why read: The most direct rebuttal to the claim that success in math and code should transfer to science. The verifier, not only scale, may decide the outcome.

Argument: A correct result is not evidence of a self-correcting scientific process if the agent gathers evidence without updating its beliefs.

Why read: Trace-level evaluation exposes epistemic failure that outcome scores hide. It is a compact case for judging the process, not only the answer.

Argument: Governments cannot plan AI for science without first mapping which datasets exist, who controls them and whether machines can actually use them.

Why read: Less glamorous than a model launch, but far more concrete about the missing public infrastructure on which many launch claims depend.

Argument: Derek Lowe says that with AlphaFold we “cheated on the exam”: it predicts protein structures from patterns in the Protein Data Bank, not from the physics of folding. Drug discovery mostly needs the answer, but the shortcut took the pressure off research that would explain folding.

Why read: Read it for a question to put to any AI result: do you want the answer, or to understand where it comes from? He sets it beside students handing coursework to chatbots, which skips the work that does most of the teaching. It is an essay from a talk, with no new data.

Argument: Alvin Djajadikerta, CEO of Evidentia Labs, argues that AI trained to cut prediction error stays inside the categories its data already records: a model fed William Farr's cholera maps could predict outbreaks but never find germs. Without new goals, AI scientists risk "hypernormal science": better predictions, fewer new questions.

Why read: Read it for a question to put to any AI-discovery claim: did the system add a new concept, or fill in a known map? GNoME's 2.2 million materials, he notes, were mostly element swaps in known structures. One proposal, AI agent populations to test how lab structure shapes discovery, has no results yet.

Argument: Anthropic researcher Siddharth Mishra-Sharma argues that well-scoped scientific code, such as porting old Fortran, can be handed to agents that run for days with little steering. It works when there is a test that says right or wrong, a progress file recording failed approaches, and a commit after every step.

Why read: Read it for a worked case outside the author's field: in a few days Claude built a differentiable cosmology solver within a percent of the standard CLASS code, though not in every regime, after testing at a single parameter point for a while. It is Anthropic's own post about its own model.

Argument: Harvard physicist Matthew Schwartz guided Claude, by text prompts alone, through a real quantum-field-theory paper in two weeks. He rates Claude a second-year grad student: tireless, but it tuned plots to look right and called unchecked work verified. What models lack, he argues, is taste, not creativity.

Why read: Read it for a worked record of where an expert had to step in (a wrong central formula, the one every result followed from, took him hours to pin down) and the habits that helped: a second model cross-checking, a tree of task notes, written honesty rules. The tenfold speed-up is his own estimate.

Argument: Pat Walters argues that some machine-learning screening papers still validate on DUD-E, a docking benchmark whose decoys were picked to look unlike the actives. A model then only has to spot the actives' chemical family: his simple fingerprint model, trained on one kinase, separated actives from decoys for most other DUD-E kinases.

Why read: Read it for a check any reader can run on a screening claim: ask how the decoys were made, and whether a model that ignores the protein scores as well. The DUD-E authors warned against this use, and a 2019 paper found CNNs did just as well with the protein switched off.

Argument: Transforming ideas into scoped projects, funded teams and executable institutions is itself a piece of research infrastructure.

Why read: It treats metascience as something to build and test, making it a useful answer to the question of where the bottleneck moves when hypotheses get cheaper.

Argument: Useful BioML work depends more on strong foundations, comfort with messy data and rapid question testing than on memorizing large amounts of biology.

Why read: A practical map of the hybrid scientific and engineering talent AI4S labs need, written by someone hiring and building in the field.

Argument: Better structures and molecular suggestions do not remove the slow biological, pharmacokinetic, toxicological and clinical failure modes that dominate drug development.

Why read: A representative dispatch from medicinal chemistry rather than a generic blog index. Lowe’s durable point is that structure can be difficult without being the rate-limiting step.

Argument: Powerful AI could compress fifty to one hundred years of biological and medical progress into five to ten years.

Why read: The most cited optimistic claim belongs here because it places the bet somewhere testable: is the bottleneck ideas, or experimental throughput, trials and regulation?

Argument: A high-impact methods paper without runnable code cannot meet the reproducibility standard expected of scientific publication, especially when the result comes from a corporate lab.

Why read: The letter matters because it tested whether journal norms still bind frontier corporate research. AlphaFold 3’s later code release made the dispute consequential rather than symbolic.

Argument: GNoME’s proposed structures offer scant evidence of compounds that are simultaneously novel, credible and useful without synthesis expertise and validation.

Why read: The key distinction is between computational discovery and experimental discovery. Later corrections in autonomous labs return to exactly this boundary.

Argument: Reanalysis of A-Lab’s reported products found systematic diffraction and novelty problems, so the original evidence did not establish 43 scientifically new materials.

Why read: The archive’s hardest external check: it uses the authors’ diffraction data rather than a rival forecast. The later correction adopts the same distinction between new to a platform and new to science.

Argument: AI changes discovery through a recurring set of roles, but its value depends on the data, representations, experiments and institutions of each field.

Why read: If you read one overview, choose this. The comparison across fields is more informative than any single method it surveys.

Argument: Across disciplines, papers and patents have become less likely to break with prior work over several decades.

Why read: Every claim that AI is accelerating science needs this baseline. The comparison is not zero progress, but a long-running decline in disruption.

Argument: Scientific institutions can be deliberately redesigned and tested so the discovery ecosystem learns how to improve its own social processes.

Why read: It answers the question raised by cheap hypotheses: attention, funding and experimental capacity become allocation problems that models alone cannot solve.

Argument: Accurate structure prediction is not the same as understanding the physical and chemical principles that govern protein folding.

Why read: For readers past the popular explanation. The distinction between prediction and understanding will recur in every field that follows.

Argument: Maintaining the same rate of progress across several fields has required rapidly growing research effort.

Why read: The denominator for the whole debate: does AI change the slope of declining research productivity, or merely add another large cohort of researchers?

Argument: Methods that scale with computation eventually outperform approaches built around human knowledge and handcrafted structure.

Why read: AI4S architecture debates repeatedly return here. Keep its boundary visible: the historical evidence comes from domains rich in data, while science is often data-poor.