The claim
When a scientific problem has enough data and a clear evaluator, large-scale learning and search can become a reusable discovery engine.
What has to hold for this route to work.
The real bottleneck. The binding constraint is not model capability but whether a scientific problem comes with a cheap, trustworthy evaluator. From blind CASP assessment to Lean and ECMWF's operational scoring, DeepMind repeatedly picks domains that already have a referee. Pushing Co-Scientist into physical experiments in 2026 is an attempt to buy a referee where none exists, using the experiment itself as the evaluator.
What is established. Where a referee exists, large-scale prediction, generation and search are established, and already packaged as infrastructure other people query. AlphaGenome Atlas precomputed and published 9 billion variant effects; WeatherNext Cyclones was benchmarked in Nature against mature operational systems with roughly a day of extra lead time; AlphaProof Nexus produced machine-checked proofs for nine open Erdős problems.
What is still a claim. End-to-end autonomous research is still a claim: physical execution runs through university partners, and the automated evaluator has itself been shown to be gameable. Co-Scientist's experiments were run at Duke, Columbia and Texas A&M, which says nothing about DeepMind's own lab; in the September swarm study one agent found a way past the evaluator and the exploit spread through shared libraries. Automated research is only as reliable as its referee.
The chosen product abstraction. The product abstraction is a precomputed public predictive layer, not a lab as a service, and not an agent subscription. AlphaFold DB, AlphaGenome Atlas and Weather Lab share one shape: turn one expensive inference run into a resource anyone can query repeatedly, distribute it through Google products, and let vertical companies such as Isomorphic capture the commercial value.
The DeepMind AI4S directions that can currently be confirmed are materials, closed-loop execution that connects research agents to real instruments, quantum error correction and genomic variant prediction, plus thinner lines in fusion, weather, mathematics, cellular mechanisms and world models. Drug discovery sits under the separate company Isomorphic Labs.
Three things are worth saying before the detail:
- DeepMind has connected its models and agents to real equipment, but almost every lab doing the work belongs to someone else: Duke's CVD reactor and microscopes, Columbia's wet lab, Omar Yaghi's lab at Berkeley, synchrotron beamlines in Japan.
- Its own lab is being built. The name and purpose are now the company's own words (the Materials Discovery Laboratory, a "closed-loop discovery system") rather than this site's inference from job titles, but nothing public gives an opening date, an address, a named lead or a result.
- The computed line, GNoME, is finished as research and continues as a dataset, and outside audits keep landing on the same problem: many entries are known compounds in an artificially ordered form. Four of the six authors now publish from elsewhere.
One theme runs through the rest: most of the new 2026 evidence is preprints and official announcements rather than peer-reviewed results, and the details are more conservative than the headlines. AlphaQubit 2's real-time configuration carries a measurable accuracy trade-off; Co-Scientist's materials experiments still need humans to load and unload samples; WeatherNext was announced as planned National Laboratory access rather than a delivered capability; and the claim that Agent_H beats six frontier models on HealthBench holds only after length adjustment.
Scope note: this is a tracker of newly available evidence and active directions in 2026, not an exhaustive catalogue of DeepMind's historical AI4S output. Mature programs such as AlphaFold, GNoME, and GraphCast appear only where they affect the current-direction assessment.
1. Materials: a lab under construction and a dataset under audit
Materials is DeepMind's most visibly resourced AI4S line and its most tangled one: the computed output is large but being corrected from outside, every physical experiment sits with a partner, and the in-house lab has a plan and job postings. Three mostly different sets of people do those three things.
1.1 The lab: the company's own words, and what they still do not say
| What can be confirmed | Details | Source wording and source | Confidence and boundary |
|---|---|---|---|
| DeepMind's own posting names the lab and states what it is for | "Google DeepMind is establishing the Materials Discovery Laboratory, a state-of-the-art automated facility designed to bridge the gap between AI predictions and physical reality. This lab will serve as the engine of our materials research program, combining high-throughput synthesis and characterization with advanced robotics to create a closed-loop discovery system." | Senior Laboratory Automation Engineer, London, on DeepMind's own Greenhouse board (req 7578225); read by this site's author on 2026-10-29 and supplied verbatim | High for the wording: it is the company's own. Two limits stay. It says what the facility is designed to do, not that it does it. And this session could not re-read the posting (the Greenhouse board answers automated readers with 404, and web.archive.org was unreachable), so the quote rests on the author's reading. |
| The lab was announced for 2026, with superconductors at the head of the target list | "Google DeepMind's first automated laboratory in the UK in 2026", with robotics to "synthesize and characterize hundreds of materials per day"; targets are superconductors at ambient temperature and pressure, "advanced batteries, next-generation solar cells and more efficient computer chips". The UK release puts superconductors first, but the memorandum both sides signed commits only to "support the launch of automated R&D and lab facilities in the UK", runs 36 months, is "voluntary, not legally binding" and carries no money | DeepMind blog, 2025-12-10; gov.uk release, 2025-12-11; the MoU (all full text) | High: primary documents. Neither announcement names a site, lead, budget or headcount, and the materials focus, the superconductor emphasis and "Gemini-integrated" are not in the signed text. "Hundreds of materials per day" is a target, not a demonstrated throughput. |
| In May 2026 the lab was still being built, and its stated job is validation | Pushmeet Kohli: "We have a wet lab, and we recently announced we are building a new autonomous materials lab. The purpose of those labs is validation." Results feed back into the model as an active-learning loop | Decoding Science interview, 2026-05-01 (the passages on materials and the wet lab read) | High: a named DeepMind executive. Our searches of DeepMind's blog (100 posts, Feb-Oct 2026), its sitemap, gov.uk and the press found no announcement that the lab has opened, no address beyond "London" and no result. |
| Hiring continues, across two separate systems | On 2026-10-29 four of five lab roles were live: Senior Laboratory Automation Engineer (London, Greenhouse req 7578225); Materials Laboratory Scientist, Science (London, Google req 139730949402174150); Research Scientist, Material Intelligence (London, Greenhouse req 7353237, closing 6 November); and Materials Laboratory Scientist, Thin Films / Solid State Synthesis. The Physical Properties role returns "Job not found". A week earlier this site had recorded the September cluster as gone from the board | The author's reading of the live postings, 2026-10-29; the earlier state from this site's own check of all 127 DeepMind listings on 2026-10-22 | High that the roles are open; the week-to-week swing is not evidence about the lab. A board where these roles have opened and closed more than once tells you about reposting and about two separate ATS systems (DeepMind's Greenhouse and Google's own), not about whether a facility is running. The lab-scientist role asks for "novel materials predicted by our generative AI models", and no DeepMind generative materials model has been published since GenMS (2024). |
| Who leads the build | Artem Mishchenko's University of Manchester profile said that "since November 2025, he has led the creation of the first experimental laboratory at Google DeepMind dedicated to closing the loop in materials discovery". In September he recruited for the lab directly: a postdoc in powder metallurgy and a senior scientist to lead thin-film growth, neither role on the official board | Manchester profile (read in September; on 2026-10-22 it returns a server error); his recruiting posts, read from a screenshot the site's author supplied on 2026-09-22 | Medium. The profile could not be re-read, and his June 2026 arXiv paper (2606.15321) prints only Manchester. The recruiting posts are his own words but were not read at source by this site. |
1.2 How GNoME holds up from outside
The 2023 Nature paper predicted 381,000 "stable" crystals; the data reaches the Materials Project in batches (117,000 by June 2026) under a non-commercial licence. The outside tests since then all circle one question: how many of those entries are actually new.
| Test | What it found | Source |
|---|---|---|
| Cheetham and Seshadri (Chem. Mater., 2024-04) | Of 10 randomly chosen "stable" entries, all 10 were already in the ICSD crystal database, usually at higher symmetry; the cause is artificial ordering of similar cations, which 0 K DFT favours. Over 18,000 entries contain Pm, Ac or Pa. They still found "the underlying approach is sound" | doi:10.1021/acs.chemmater.4c00643 (full text) |
| Jakob, Walsh, Reuter, Margraf (Adv. Mater., 2025-10) and Liverpool with UCL (2026-10) | A classifier trained on the ICSD estimates that 80-84% of GNoME compositions are likely to show substitutional disorder, against 39-47% for the Materials Project; Liverpool and UCL separately found that most GNoME structures sit within 0.05 Å of a known crystal's geometry. The first group adds that this does not make ordered predictions useless | doi:10.1002/adma.202514226 (full text); arXiv 2610.13885 |
| Outsiders who made a GNoME prediction | No count exists. Two cases found: MnFeCo₄Si₂ (Fukuoka Institute of Technology, arXiv 2603.22748), made single-phase in the predicted rhombohedral structure, a soft ferromagnet with a Curie temperature of 1039 K (one compound, lab X-ray diffraction only); and Nd₃BTeO₉ (TUM and Altrove, arXiv 2511.01790), which a 2023 paper had already made | arXiv abstracts and the relevant sections |
A related correction: the 2023 A-Lab paper, with two DeepMind co-authors, was retitled on 2026-01-19 from "novel" to "inorganic" materials, with "new" meaning "new to the prediction platform, not necessarily new to science"; 36 of 40 reported successes were confirmed on re-analysis (Nature, full text). DeepMind's public answer to the critiques remains its April 2024 statement: "We stand by all claims made in Google DeepMind's GNoME paper" (The Register). Crossref lists no correction to the GNoME paper.
The people have changed too. Ekin Dogus Cubuk, who led materials science and chemistry and a group of 22, left in 2025 to co-found Periodic Labs (Physics Today, 2026-03-20); Muratahan Aykol and Gowoon Cheon are on Periodic's founding team; Samuel Schoenholz prints OpenAI on papers since December 2023; Sherry Yang prints NYU in 2026; Simon Batzner announced on 2026-10-07 that he has joined Anthropic (his post). Of GNoME's six authors, only Amil Merchant still printed Google DeepMind on 2025 work. The 2026 lab effort is a different set of people, and the 2026 physical experiments came from the Co-Scientist group.
2. Closed-loop execution: Co-Scientist connected to real instruments
This line pushes Co-Scientist from "generating ideas" toward experiments that touch real equipment. The August 27, 2026 preprint (arXiv 2608.26701) reports validation across three domains, and it is the clearest public evidence of how DeepMind reaches real experiments today.
| What can be confirmed | Details | Source wording and source | Confidence and boundary |
|---|---|---|---|
| Materials: connected to a CVD reactor, growing monolayer MoS₂, MoSe₂ and WS₂ | DeepMind handled hypothesis generation, recipe generation and the agent system; the Duke team designed and built the automated CVD reactor and ran synthesis and most characterization (partly at Duke SMIF); Texas A&M ran TEM and HAADF-STEM | "integrating fast inference for direct reactor hardware control"; "single-attempt synthesis … with flakes exceeding 50 micrometers."; Author Contributions and Acknowledgments | High: preprint evidence, but not end-to-end automation. Humans still loaded samples and ran characterization; only the growth stage was program-controlled. All three materials already have established CVD literature, and the difficulty is that recipes do not transfer cleanly between furnaces ("their successful synthesis remains highly system-dependent"), so this is a recipe transfer, not the discovery of a new material. |
| Materials: Co-Scientist proposed a safer precursor route | C₂Cl₆ in a bottom-up CVD route toward an MXene-like material | "Co-Scientist proposed a safe precursor (C₂Cl₆) for bottom-up CVD MXene synthesis." | High: preprint evidence, but the path was difficult. A target peak appeared only after 25 design iterations; reproduction rose from 3 of 26 runs to 17 of 25 after equipment troubleshooting; the team used the second-ranked proposal and modified it by hand. The atomic structure is still unresolved: "post-growth oxidation and low product yield prevent definitive atomic-scale phase assignment without cross-sectional atomic STEM." |
| Biology: predicting E. coli swarming morphology | Tal Danino's lab at Columbia (Associate Professor of Biomedical Engineering, head of the Synthetic Biological Systems Laboratory) handled engineered strains, wet-lab assays and ground-truth measurements, with Marian Shaw and colleagues doing the bench work; on the DeepMind side Samuel Schmidgall, Tao Tu and others designed Co-Scientist and the phenotypic vision-language prediction pipeline | Same paper, Author Contributions; public Columbia and Google researcher profiles | High: preprint evidence. The structure is again DeepMind's prediction and agent layer with a partner's wet lab, not DeepMind running the E. coli experiments itself. |
| Computer science: autonomous search for an inference-time scaling architecture (Agent_H) | On HealthBench Hard and Professional, Agent_H outperforms six frontier models after length adjustment | Table 1 states that it "outperforms six frontier models on HealthBench Hard and Professional (length adjusted)," while the abstract and introduction omit the qualifier | High: preprint evidence, but the qualifier matters. On the unadjusted Professional score, Claude Opus 5 scores 0.697, GPT-5.6 Sol 0.664 and Agent_H 0.645, placing it third. The physician blind review used three physicians over 106 questions with one score per query, against unscaffolded Gemini 3.1 Pro; of nine dimensions only likelihood of harm reached significance (p = 0.0486), and the authors write that the automated advantage "did not translate into a perceived difference". |
| Autonomy varies sharply across the subprojects | In the mathematics work AI explored while humans completed and verified the proofs; in materials humans iterated alongside the system; only the architecture search is described as autonomous | Paper descriptions | High: paper wording. These should not be collapsed into "AI autonomously does science." |
So DeepMind has two different layers of experimental capability, and they are best kept apart. The first is external collaboration that is already running: DeepMind supplies the model, the agent and the experimental plan, and partner labs at Duke, Columbia and Texas A&M do the physical execution. The second is capability being built in-house, and for now that layer consists of the company's own description (the posting quoted in 1.1) and the Manchester profile saying the build has been underway since November 2025. Nothing public establishes whether the London lab is in routine operation, and the Duke and Columbia experiments cannot be counted as evidence that it is.
DeepMind has connected its scientific-intelligence / orchestration layer to real experimental environments, but physical execution still relies heavily on external collaborators; at the same time, it is progressively internalizing that capability in London.
3. Quantum error correction: AlphaQubit 2
A boundary case: AI applied to quantum information science, readable also as computing infrastructure. The paper reports two configurations, a full model optimized for accuracy and a smaller real-time version optimized for speed, and they should not be collapsed into one claim.
| What can be confirmed | Details | Source wording and source | Confidence and boundary |
|---|---|---|---|
| Real-time throughput reaches the target regime | Below 1 μs per cycle for surface code distance 11 and color code distance 9, running on Trillium TPUs without custom ASICs or FPGAs | "We demonstrate real-time decoding faster than 1μs per cycle on commercial accelerators." arXiv 2512.07737, v1 2025-12-08, v2 2026-03-11 | High: preprint evidence. This is throughput, not end-to-end latency. |
| The real-time version gives up some accuracy | At surface-code distance 11, the logical error rate is 3.73×10⁻⁶ for AQ2-RT versus 1.89×10⁻⁶ for the full model; the drop is larger for color codes | "the speed increase of AQ2-RT entails a modest trade-off in accuracy" | High: paper wording. Speed and accuracy come from two different configurations. |
| Real-hardware data only goes to distance 7, and this is not online correction | The model was evaluated on recorded data from the 105-qubit Willow chip, which implements surface codes at distances 3, 5 and 7; the distance-11 result is under simulated noise | Same paper; listed as future work: "We have not yet addressed reducing the latency between the final measurement and the decoding result" | High: paper wording. There is no evidence of end-to-end online operation; the result covers the decoder stage. |
4. Genomics: AlphaGenome Atlas
!AlphaGenome Atlas coverage overview
The project turns predictions for the molecular effects of every possible single-nucleotide change in the human genome into a queryable resource.
| What can be confirmed | Details | Source wording and source | Confidence and boundary |
|---|---|---|---|
| Predictions for 9 billion variants are precomputed and released, with a ranking score, AVI | Every possible single-letter change in the human genome; AVI combines AlphaGenome and AlphaMissense predictions across coding and non-coding regions | "Predictions for the effects of 9 billion single-nucleotide variants — every single-letter change possible — in the human genome." DeepMind announcement, 2026-09-08 | High: official disclosure. |
| Three named collaborations already use it | Rare disease: Laura Covill and Anne O'Donnell-Luria of the GREGoR Consortium and the Broad used AVI to reprioritize overlooked variants and found a DNM1-related variant creating an aberrant splice site. Common traits: Gareth Hawkes at Exeter used whole-genome data from 54,000 UK Biobank participants and increased non-coding associations by 22%. Regulatory grammar: Julia Zeitlinger and Melanie Weilert at the Stowers Institute separated transcription factors that only alter DNA accessibility from those that switch genes on or off | Same announcement | High: the announcement names the researchers and gives specific results. The last two are not disease-variant work: one is common complex traits, the other basic regulatory mechanism. |
| Not for clinical use | DeepMind draws this boundary explicitly | "AlphaGenome has not been validated for, and is not approved for, any clinical use." | High: official statement. A research tool, not a validated diagnostic system. |
5. The other directions, and how far each one goes
Each of these has public material, and none produced new evidence this year that would carry a section of its own.
| Direction | What can be confirmed | Boundary |
|---|---|---|
| Fusion | Plasma simulation and real-time control with Commonwealth Fusion Systems, through the open-source TORAX simulator (arXiv 2406.06718, revised 2026-07-09); CFS's Devon Battaglia confirms by name that SPARC uses it | A paper revision only shows the line is maintained. No new 2026 experimental result or delivery was found. Hassabis has mentioned possible materials design in this collaboration, with no research material behind it. |
| Cellular and cancer mechanisms | Perturb-ME with Genentech: genome-wide CRISPR screening plus multimodal single-cell measurement of MHC-I expression in A375 melanoma cells, fitting 221 regulators onto 1,998 responsive genes (bioRxiv, 2026-08-18, with a code repository) | A preprint, not peer reviewed, with no independent replication seen. Co-Scientist is the interpretation layer, not the screen or the model. Aviv Regev's affiliation on this paper is Genentech (head of gRED), not the Broad. The mRNA vaccines and CAR-T mentioned in a LinkedIn post are examples of potential application, not a pipeline. |
| Weather and Earth science | WeatherNext was listed for planned access to U.S. National Laboratories (Genesis announcement, 2025-12-18) | Only Co-Scientist was described as available "starting today". Announced access and delivered access are two different things; no independent evidence of rollout was found, and no new 2026 research result. |
| Algorithms and mathematics | Co-Scientist worked with mathematicians from Clayton State University and elsewhere on Chowla sets, with a formal paper (arXiv 2607.24847); AlphaEvolve has been opened to National Labs, and BASF used it to optimize supply chains | The paper is explicit about the division of labor: humans formulated the problem and independently verified and completed every argument, while AI explored examples and possible proof strategies. That supports "AI did not complete the proof here", not "AI could not". The BASF statement supports supply-chain optimization only. |
| World models for scientific simulation | Genie and Veo learn dynamics from raw data, and the team builds physics benchmarks in game engines to test them | Only an interview statement from a late-2025 podcast, with no paper identified. Hassabis says the fidelity is insufficient: visually convincing at a glance, but not accurate enough for uses such as robotics. Scientific use is not established. |
| Drug discovery (Isomorphic Labs) | A separate company, so not DeepMind's own pipeline. The IsoDDE design platform (February 10), the multi-target Johnson & Johnson collaboration (January 20) and a $2.1B round led by Thrive Capital (May 12) are public; in April, President Max Jaderberg said at WIRED Health: "We're gearing up to go into the clinic" | Performance claims come largely from the company's own reporting, with no independent validation found. Novartis, Lilly and Johnson & Johnson are strategic partners, not investors in that round; the same release lists them separately and the two are easily merged. First patient dosed, programme count and stages: no formal confirmation found. |
What remains unverified
- How much compute and headcount are allocated to science: never publicly disclosed. Any resource-allocation inference in this tracker is inference, not a reported fact.
- The operating status of the UK lab: announced for 2026 and "building" in May 2026; no opening, address, scale or result found as of 2026-10-29.
- Whether a superconductor project is running: superconductors lead the UK lab's announced targets and Hassabis calls them a "dream breakthrough", but no project, paper or result has been published. The same holds for specific battery chemistries and customers.
- The "generative AI models" the lab job names: no DeepMind generative materials model has been published since GenMS (2024).
- A California experimental lab: the only lead was a now-deleted post by Simon Batzner, who has since left for Anthropic.
- A virtual-cell program: vision-level statements, with no corresponding public team or research output identified here.
- Named 2026 Genesis projects and deliverables: not verified in this review.
How to read this tracker
Scope: only work where AI is being applied to scientific research. General model capability and consumer products are excluded. World-model work is included only where there is an explicit scientific-simulation angle. Quantum error correction is a boundary case and is included with that caveat.
Confidence is assigned claim by claim, not to an entire research direction. Within one topic, "the team is hiring" and "the technology is already working" can have very different evidence strength. High means a paper, an official disclosure or a named independent partner source; Medium means a researcher's own public statement or media reporting not independently re-verified here; Needs verification means a lead exists but the evidence is insufficient or the original source is no longer accessible. The arXiv and bioRxiv papers cited here are preprints, and "there is a paper" means the authors reported the result.
Two operating rules: revised papers are dated by version and are not treated as entirely new projects; evidence that a project exists is not the same as evidence that the technology has succeeded.
The materials section was re-read in full on October 22, 2026 (job board, the UK announcements and memorandum, paper bylines, outside audits); the lab subsection was updated on October 29, 2026 with DeepMind's own description of the Materials Discovery Laboratory, supplied by this site's author from the live posting, and the whole tracker was cut back the same day. LinkedIn quotations rely on the post records available for this research; the original pages were not re-opened, since LinkedIn does not serve automated readers, and the September 2026 recruiting posts were read from screenshots the site's author supplied.
Primary sources
Papers
- arXiv 2608.26701, 2026-08-27: Co-Scientist's three-domain execution validation and autonomy study (preprint)
- Nature 624, 80, 2023-11-29: GNoME; the A-Lab authors' correction, 2026-01-19
- Chem. Mater., doi:10.1021/acs.chemmater.4c00643, 2024-04-08; Adv. Mater., doi:10.1002/adma.202514226, 2025-10-23; arXiv 2610.13885, 2026-10: outside tests of GNoME's novelty and disorder
- bioRxiv 10.64898/2026.08.17.745330, 2026-08-18: Perturb-ME, the MHC-I regulation study with Genentech (preprint)
- arXiv 2607.24847, 2026-07-25: the human-AI collaboration on Chowla sets
- arXiv 2406.06718, revised 2026-07-09: the TORAX plasma transport simulator
- arXiv 2512.07737, revised 2026-03-11: the AlphaQubit 2 real-time decoder
Official disclosures
- AlphaGenome Atlas (2026-09-08); Gemini for Science (I/O 2026)
- The fusion collaboration (2025-10-16); the UK government partnership (2025-12-10); Genesis (2025-12-18)
- Isomorphic Labs: the J&J collaboration; IsoDDE; the financing
- Job postings: the Materials Discovery Laboratory passage comes from the Senior Laboratory Automation Engineer posting on DeepMind's own Greenhouse board (London, req 7578225), read by this site's author on 2026-10-29; that board answers automated readers with 404 and this session could not reach web.archive.org, so it was not re-read here. The others: Materials Laboratory Scientist, Science (Google Careers, read 2026-10-22); Research Scientist, Material Intelligence; Senior Laboratory Automation Engineer (both Internet Archive snapshots, where every line quoted here can be checked)
Interviews and staff moves
- The DeepMind podcast; Stanford GSB (published 2026-06-18); a long-form interview
- Cybernews (2026-04-26): the WIRED Health remarks
- LinkedIn: Tao Tu, Vivek Natarajan, Artem Mishchenko, Simon Batzner (deleted)
Evidence timeline
| Date | Event | New evidence | What it does not prove | Sources |
|---|---|---|---|---|
| AlphaGenome Atlas released | Precomputed molecular-effect predictions for 9 billion single-nucleotide variants across the human genome were released. | The resource is explicitly not validated or approved for clinical use. | Official release ↗ | |
| Co-Scientist enters physical experimental workflows | A preprint reports cross-institution validation including materials CVD and bacterial swarming experiments. | Physical work was primarily executed at Duke, Columbia and Texas A&M; it does not show that DeepMind's internal lab is operational. | Preprint ↗ | |
| From specialist models to agents and algorithm search | Co-Scientist, AlphaEvolve and WeatherNext expand the boundary of automatically evaluable scientific tasks. | Can success in structure and prediction extend to causal discovery and general scientific agents? | — | |
| AlphaFold 3 released | Structure prediction expands to proteins, nucleic acids, small molecules and ion complexes. | Can success in structure and prediction extend to causal discovery and general scientific agents? | — | |
| GNoME and GraphCast | The approach expands from biology into materials and Earth systems. | Can success in structure and prediction extend to causal discovery and general scientific agents? | — |
On evidence. A CEO claim is not a benchmark; a benchmark is not a physical experiment; a physical experiment is not manufacturing scale; a third-party test is not qualification; qualification is not deployment. Every line keeps its source and says what it supports and what it cannot show.
Sources
- Google DeepMind Science · Organization
- AlphaFold · Project
- GNoME, Nature · Paper
- Co-Scientist · Technical release