John Jumper

Who he is

Affiliation (CHANGED 2026-06-19): Joining Anthropic — announced on X on 2026-06-19 after nearly nine years at Google DeepMind (Wikipedia, fetched 2026-09-21: "served nearly 9 years (until June 2026)"), with a break to recharge first; no title, team or start date has been disclosed by him or by Anthropic as of 2026-09-21 (TechCrunch, 2026-06-20; CIOL, 2026-06-20: "He did not disclose the position he will hold at the company"). Previously: Google DeepMind, c. 2017–2026 (start inferred from the PhD year and "nearly 9 years") — led the AlphaFold team from roughly six months after his PhD (his own words, X post as quoted by TechCrunch); title trail on primaries: "Director" (DeepMind, 2024-10-09; EMBL, 2024-10-09) → "Distinguished Scientist, Google DeepMind" (Royal Society, elected FRS 2025). The v1 title "Senior Research Director" appears on none of them — see Delta log. 2024 Nobel Prize in Chemistry: "one half of the prize has been awarded to David Baker 'for computational protein design' and the other half jointly to Demis Hassabis and John M. Jumper 'for protein structure prediction'" (EMBL, quoting the Academy). Born 1985, Little Rock; Vanderbilt BS physics/maths 2007; Cambridge MPhil 2008 (Marshall Scholar); University of Chicago PhD in theoretical chemistry 2017 under Tobin Sosnick and Karl Freed (Wikipedia, fetched 2026-09-21 — his widely reported pre-PhD industry stint in molecular-dynamics simulation was on no page fetched this cycle and is left out). Honors on the same page: Nature's 10 (2021), BBVA Frontiers (2022), Breakthrough Prize, Gairdner and Lasker (all 2023), FRS (2025, confirmed on the Society's own page), National Academy of Engineering (2026 — Wikipedia only; primary not fetched).
One-line position: AI can solve fundamental scientific problems that resisted decades of traditional research — AlphaFold proved it for protein structure, and the approach generalizes. His own 2024 framing: "It is a key demonstration that AI will make science faster and ultimately help to understand disease and develop therapeutics" (DeepMind, 2024-10-09). The 2025–26 addendum is the part v1 didn't have: he now says the next instrument is the language model — "I'll be shocked if we don't see more and more LLM impact on science" (MIT Technology Review, 2025-11-24) — and seven months later he moved to the LLM lab.

Discipline & technical bet

Trained as a physicist and theoretical chemist (protein-folding simulation), retooled into a machine-learning architect. The Royal Society's citation is the cleanest statement of the discipline: he "led the development of the AlphaFold2 and AlphaFold3 systems at Google DeepMind, where he developed entirely new machine learning architectures to dramatically increase the ability of AI systems to learn from biological and chemical data" (Royal Society, 2025). That is the old bet: purpose-built networks that bake physical constraints into the architecture and learn from small, expensive datasets — "You have a small amount of data to use, and you have to use it really well" (Chicago Maroon, 2025-01-23).

The new bet, stated then acted on: structure prediction becomes a component, not the product. "We're trying to figure out how to make structure prediction an even bigger part of the problem, because we have a nice big hammer to hit it with," and, on language models: "We have machines that can read science. They can do some scientific reasoning... How do you get these two technologies to work together?" (MIT Technology Review, 2025-11-24). The wager is that a general model that reads and reasons about science, wired to specialist instruments like AlphaFold, beats either alone. Revealed behavior matches the words: in June 2026 he left the lab that owns the instrument for the lab whose product is the reader (TechCrunch, 2026-06-20). What he will build there is undisclosed; the file treats the bet as live and unscored.

Key claims (Says)

Empirical: AlphaFold solved protein structure prediction — predicted 3D structures of ~200M proteins at experimental-level accuracy (AlphaFold 2, 2020–21; AlphaFold DB, 2022). Held — CASP14: "our latest AlphaFold system achieves a median score of 92.4 GDT overall across all targets," where "a score of around 90 GDT is informally considered to be competitive with results obtained from experimental methods" (DeepMind, 2020-11-30); the Nature paper (Jumper first author, Hassabis last; 2021-07-15) claims "accuracy competitive with experimental structures in a majority of cases"; the database reached "over 214 million predicted protein structures" (Nucleic Acids Research, 2023-11-02). Two asterisks he now states himself: "It's a database of predictions, and it comes with all the caveats of predictions" (MIT TR, 2025-11), and the model is "largely unable to predict when a point mutation causes defective protein folding" (Buel & Walters, Nat. Struct. Mol. Biol., 2022-01). "Solved" means the static fold of a typical chain, not everything a structure can be asked.
Empirical: AI as a scientific instrument is real, not theoretical — a deployed result, not a forecast; structures that took a year now take minutes. Held — "over 4,501,953 total users accessing the AFDB website, more than 18,000 proteome archives downloaded" (Nucleic Acids Research, 2025-11-22); DeepMind's Nobel-day count was 2 million scientists in 190 countries (2024-10-09). His own timing claim: structures "in about five or 10 minutes instead of in a year" (Chemistry World, 2024-11-22).
Empirical: The approach generalizes — learning physical constraints from data extends to drug discovery, materials, molecular dynamics and other domains. Drifted — held on scope, unproven on payoff. Scope: AlphaFold 3 (2024-05-08) extended the method to "proteins, DNA and RNA, as well as small molecules, also known as ligands," claiming "at least a 50% improvement compared with existing prediction methods" for those interactions (Google/DeepMind/Isomorphic, 2024-05-08 — self-reported). Independent checks since: co-folding tools (which place protein and drug molecule together) "largely memorise ligand poses from their training data, hindering their use for de novo drug design" (Škrinjar et al., Basel/SIB, v3 2025-08-04); "popular methods such as AlphaFold 3 are still challenged by prediction targets with novel binding poses" (PoseBench, v7 2026-03-20); at CASP16, "AlphaFold3 performs slightly better than AlphaFold2 for proteins" but "when massive sampling is applied to AlphaFold2, the difference disappears," and RNA accuracy "is relatively low" (Elofsson, Proteins, 2025-08-25); "AlphaFold 3-based models did not significantly outperform older methods for predicting protein-ligand interactions" in CASP16's preliminary results (Wikipedia, Isomorphic Labs, fetched 2026-09-21). Materials: the sister claim (GNoME) drew "scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility" (Cheetham & Seshadri, Chem. Mater., 2024-04-08). Drug discovery: no AlphaFold-lineage compound is in a human as of the last fetched source — see Predictions.
Empirical / methodological: Scientific AI requires domain expertise — AlphaFold needed deep biophysics, not just ML engineering. Unfalsifiable as written — rewritten. Observable attached: does a team without structural-biology depth produce a method that wins a CASP category, or does a general LLM agent do so, by CASP17 (the biennial cycle puts results in late 2026) or CASP18 (2028)? Early data cuts both ways: MIT's Boltz-1 "matches the performance of Chai-1 and, therefore, AlphaFold3" (Boltz, 2024-11-16) — but it was built by an ML group from AlphaFold 3's published recipe, so it tests reproducibility, not origination. His own 2025 line sits on the fence: "AI systems aren't like chatbots that get to learn from the entire internet" (Chicago Maroon, 2025-01) — and then he joined the chatbot lab.
Normative: Open access amplifies impact — AlphaFold DB was released freely and the open-science bet accelerated impact enormously. Drifted — mechanism Held, behavior retreated. Mechanism: the free database is the 4.5M-user number above. Behavior: AlphaFold 3 shipped 2024-05-08 as a paper plus a rate-limited server and no code; Nature's own editorial defended publishing it that way (Nature editorial, 2024-05-22); "more than 1,000 scientists" signed a protest letter, DeepMind's Pushmeet Kohli promised release "for academic use" within six months, and Jumper "reportedly told the press that DeepMind and Isomorphic Labs had consulted more than 50 experts in biosecurity, bioethics, and AI safety and concluded that AlphaFold 3's marginal biosecurity risks were far outweighed by the system's potential benefits" (GEN, 2024-06-13). Code arrived 2024-11 under Apache 2.0, but the weights are gated — "You may only use AlphaFold 3 model parameters if received directly from Google" — and "for commercial use, AlphaFold 3 is available via Gemini Enterprise Agent Platform on Google Cloud" (GitHub README, fetched 2026-09-21); weights became publicly downloadable in 2025-02 "still retaining the non-commercial restriction" (Wikipedia, AlphaFold). The open-science evangelist of 2021 co-authored the gated release of 2024, with Alphabet's drug company as co-developer. Not Wrong — he never claimed AF3 would be open — but the file can no longer carry "open science" as an unqualified trait.

Notable predictions — with falsifiable checks

(2020–21) AlphaFold could predict protein structures at experimental accuracy. Held — CASP14 median 92.4 GDT (DeepMind, 2020-11-30); independent benchmarks since have contested the edges (mutations, dynamics), not the core. Check going forward: does AlphaFold-lineage accuracy on novel folds hold in CASP17 (results expected late 2026) against physics-plus-sampling baselines?
(2022) Free release of 200M+ structures would accelerate global research. Held — 4.5M total users and 18,000 proteome downloads (NAR, 2025-11-22); the database keeps growing (1.7M high-confidence homodimer predictions added, EMBL, 2026-03-16). Check: NAR's annual database paper (each November) shows total users still rising through 2027; a plateau or a paywall on the DB would Drift it.
(Implicit, 2021–24) AI-for-science will expand to other domains — materials, chemistry, drug design. Drifted — expanded in scope (AF3 to ligands and nucleic acids; Boltz/Chai/OpenFold3 replications; Isomorphic's internal oncology and immunology programmes, Fortune, 2025-07-06) but not in demonstrated payoff: Isomorphic "expects to have its first clinical trials by the end of 2026" after Hassabis "had said last year the company would have AI-designed drugs in clinical trials by the end of 2025" (Reuters, 2026-01-20); the 2026-05-12 Series B release does not mention a trial (Isomorphic, 2026-05-12); the latest trade piece fetched says the company "is set to begin human clinical trials" (ChemDiv, 2026-04-27 — headline says "launches," body says plans). Materials: Cheetham & Seshadri (2024). Check: (a) a named Isomorphic compound doses a first patient by 2026-12-31 — Hassabis's own date; (b) a Phase 1 readout by 2028-12-31. Two more misses on (a) and this grades Wrong on the date, direction logged separately.
(2025-11-24) "I'll be shocked if we don't see more and more LLM impact on science." Live — new this cycle, not graded. Check: by 2027-12-31, a peer-reviewed result in a top-tier journal where an LLM agent is credited with the hypothesis, design or analysis step (not just literature search), or a Jumper-bylined Anthropic paper showing an LLM plus a structure model beating the structure model alone on a held-out task. Absence through 2027 is a Drift he would own, since he moved jobs on it.
(2025-01-23) "It does seem likely that we will make generally intelligent AI systems at some point." Unfalsifiable — no date, no bar; recorded as his only found AGI statement (Chicago Maroon). Rewrite or retire: the file will grade it only if he attaches a horizon.
(2024-10-09) AI "will make science faster and ultimately help to understand disease and develop therapeutics." Live — "faster" is Held on the database numbers; "develop therapeutics" has the Isomorphic check above plus a longer one: an AlphaFold-lineage-designed drug approved anywhere by 2030-12-31.

Revealed behavior (Does)

Left the instrument lab for the reader lab (2026-06-19). Nearly nine years at DeepMind, Nobel in hand, and he chose Anthropic — a company whose science vehicle is Claude, not a structure model. Sequence at the destination: Claude for Life Sciences (2025-10-20, "Increasing the rate of scientific progress is a core part of Anthropic's public benefit mission"); Allen Institute and HHMI/Janelia agent collaborations (2026-02-02); Coefficient Bio bought for "a $400 million stock deal" — "around 10 people" led by two founders from Genentech's Prescient Design, into "Anthropic's health and life science team" (TechCrunch, 2026-04-03); Karpathy to pre-training (2026-05-19); then Jumper. He is the capstone hire of a dated, deliberate build-out.
Left during a DeepMind talent drain, not before it. Same week: Noam Shazeer to OpenAI (TechCrunch, 2026-06-20). About seven weeks later Hassabis stepped down as DeepMind CEO (2026-08-05, per the Hassabis bench file, reviewed 2026-08-28, which also logs Jeff Dean's exit). Jumper's post was gracious — Hassabis "took a real chance letting me lead the AlphaFold team just six months after finishing my PhD"; "GDM is a special place and I'll still be excited to hear about what amazing things they discover next" (X, 2026-06-19, as quoted by TechCrunch and Digital Applied, 2026-06-27). No reason given beyond the move itself.
Chose gated release for AlphaFold 3 and defended it publicly (2024-05). The 50-expert biosecurity consultation is the one concrete safety-governance act on his record, and it cut toward restriction (GEN, 2024-06-13; Google blog, 2024-05-08: "engaged with more than 50 domain experts... across biosecurity, research and industry"). v1's "his views on AI governance, safety... are less developed" missed that he had already made a governance call on his own artifact.
Kept publishing through the Nobel year — co-senior author on AlphaFold 3 (GEN: "one of the senior authors of the AlphaFold 3 article and lead author on AlphaFold 2"), which was co-developed with Isomorphic Labs, Alphabet's commercial drug arm. Byline position matters (chris-olah rule): AF2 is his paper (first author); AF3 is his programme (co-supervisor with Bapst, Kohli, Jaderberg, Hassabis).
Time went to talks, not policy. 2025 public output found: Chicago Maroon interview (2025-01-23), YC AI Startup School talk (2025-06-16, per the episode listing), a Nobel Prize podcast episode (2025; nobelprize.org returned 403 this run, date not verified), MIT Technology Review (2025-11-24). Zero policy testimony, zero open letters signed, zero board seats found on any fetched page. Since the 2026-06-19 post: silence.

Feels

Wants AlphaFold to be remembered as "one of the first AI systems that deeply mattered [for] biology but just one of many, many that did" (Chemistry World, 2024-11-22) — the fear underneath is being the man of one result. Allergic to overclaiming on his own behalf: "It's not like we were one protein structure away from curing any diseases" (MIT TR, 2025-11).

Hears

Structural and experimental biologists first ("to inform the incredible work of experimental biologists," 2024-10-09); the CASP community that scores him; the DeepMind science bench he just left; and now Anthropic's model and agents teams. Little sign of the policy or x-risk discourse in anything he says.

Sees

What a purpose-built scientific model can and cannot do from the inside — including the failure modes the benchmarks found later — plus, from mid-2026, what a frontier LLM lab's internal models do on scientific reasoning before anyone outside sees them. Nobody else on the bench has held both vantage points.

Incentive map

Re-derived from scratch this cycle; the position moved. Through 2026-06 he was paid by Alphabet, and his signature 2024 artifact (AF3) was co-owned with Alphabet's drug company — which explains the gated release better than biosecurity alone (Nature's editorial notes usage restrictions on drug development; the commercial route is Google Cloud per the README). From 2026-06 he is an Anthropic hire on undisclosed terms; standard practice implies equity in a company whose life-sciences line is a product. What he cannot easily say now: "LLMs won't matter much for wet-lab science," "purpose-built models beat general ones," or "Anthropic's biology push is early." Note the alignment of thesis and paycheck: his 2025-11 "more and more LLM impact on science" line predates the move by seven months, so the belief came first — but the next restatement will be from inside. The Nobel is a platform, not a conflict (a shared prize purse is not a business). No fund, no board, no advisory seats found; absence search run on fetched pages only. Rule-25 discount: low on AlphaFold facts (peer-reviewed, replicated three times over); moderate and rising on any claim about LLMs-for-science, Anthropic's science programme, or Isomorphic-vs-Anthropic comparisons; the "AI cures disease" timeline claims now carry his employer's stake as well as his old one's.

Theories aligned with

AI-for-science — the definitive proof case; no theory file on the bench yet (would sit under techno-optimism)
Open science / public-good AI deployment — qualified since 2024: holds for the database, not for AlphaFold 3's weights
Adjacent to techno-optimism — but grounded in demonstrated results, not forecasts
Country of geniuses in a datacenter / compressed century of biology (Amodei) — AlphaFold is that thesis's single strongest data point, and its author now works for the thesis's author

What he's reacting against

The frame that AI is only about chatbots and language models — AlphaFold demonstrates AI as a scientific instrument, not a consumer product. 2026 note: he now argues the two must combine, not that one replaces the other
Decades of slow progress on protein folding — AlphaFold solved it in a way the structural biology community did not expect (CASP co-founder John Moult: "We have been stuck on this one problem... for nearly 50 years," DeepMind, 2020-11-30)
Skepticism that AI can make fundamental scientific contributions — the Nobel Prize is the strongest possible validation
The separation of AI from the natural sciences — argues AI and scientific disciplines must converge
Overclaiming in his own name — "AlphaFold will only tell us some [things] about biology and there is so much more – about the cell, about interactions, about the mesostructure, about regulation – that isn't directly even askable with the AlphaFold framework" (Chemistry World, 2024-11-22)
Copy-the-internet ML as a template for science — "AI systems aren't like chatbots that get to learn from the entire internet" (Chicago Maroon, 2025-01-23)

Where he overlaps / splits (with Rich)

Overlaps with Hassabis on AI as a scientific tool (co-authors: Hassabis last author on AF2, co-senior on AF3); splits on scope — Hassabis pursues AGI broadly, Jumper focused on structural biology. Drifted — the split is now institutional as well as intellectual: Jumper left Hassabis's lab (2026-06-19) and Hassabis then left the CEO seat (2026-08-05), keeping Isomorphic; the AlphaFold pair are on opposite sides of the DeepMind–Anthropic line for the first time
Overlaps with Koller (insitro; no bench file) on AI for biology; splits on application — Jumper did fundamental science, Koller does drug-discovery commercialization. Drifted — AF3 was co-developed with Isomorphic (commercial) and his new employer bought a drug-discovery startup two months before hiring him; the fundamental/commercial line no longer runs between them
Overlaps with Amodei's "compressed century of biology" — provides evidence for it; AlphaFold is the single strongest data point. Drifted into an incentive-map entry — since 2026-06 Amodei is his employer; the overlap is real and the independence is gone (moonshots rule: mutual confirmation between personas who work together is not two data points)
Splits with Marcus / Mitchell on what AI can do — AlphaFold is a case where AI genuinely solved a problem humans couldn't, not just pattern-matching. Held on the fold; partial concession on the drug — Škrinjar et al. (2025) found co-folding tools "largely memorise ligand poses," which is the pattern-matching critique landing on AF3's newest capability; neither bench file mentions AlphaFold, so the point is unclaimed on their side
Overlaps with Karpathy as the other 2026 Anthropic hire on the bench (31 days apart) — different teams, same signal about where senior research talent went this year
Less engaged with the bench's labor, governance and x-risk personas (Bengio, Hinton) — his only found governance act is the AF3 release decision; his only found AGI line is undated

Track record

Nobel Prize in Chemistry (2024-10-09) — jointly with Hassabis "for protein structure prediction"; the highest possible scientific validation. Held
AlphaFold DB used by millions of researchers worldwide — impact is measured, not speculated. Held — 4,501,953 total users (NAR, 2025-11-22)
CASP14 victory (2020) — one of the most dramatic scientific achievements of the decade; established AI's credibility as a scientific tool. Held — 92.4 GDT median (DeepMind, 2020-11-30)
The science is replicable and the results are open — unusually verifiable for AI claims. Drifted — replicable: yes, three independent lineages re-implemented AF3 from the paper (Chai-1; Boltz-1 2024-11-16 and Boltz-2 2025-06; OpenFold3 — release date not on any page fetched). Open: AF2 yes, AF3 weights non-commercial and Google-gated (GitHub README). Grade the two halves separately from now on
AlphaFold 3 (2024-05-08) — co-senior author; extended the method to nucleic acids, ligands and ions; self-reported "at least a 50% improvement" on interactions; independent benchmarks since show memorisation on novel ligands and parity with sampled AF2 on proteins (Elofsson, 2025; Škrinjar, 2025; PoseBench, 2026)
Honors since v1: FRS (2025, Royal Society page), NAE (2026, Wikipedia only). Zero public misses found — but also zero dated forecasts of his own to miss; his record is artifacts, not predictions, which is why this file attaches checks to the few he has made
Pattern: exceptional on the science; cautious and accurate in his own claims about limits; institutional choices (gated AF3, the 2026 move) track his employer's commercial interest each time — not damning, but the file should stop treating him as commercially weightless

Empirical vs normative

Almost entirely empirical: protein structure predictions, accuracy benchmarks, database usage statistics, downstream research impact — and, new, a bet that LLMs will matter for science
Normative content is minimal: science should be open (now qualified by his own AF3 decision); AI should serve scientific discovery; biosecurity risk-benefit should be weighed by expert consultation — low-controversy positions
Source quality: highest tier on AlphaFold 2 (peer-reviewed, CASP-scored, replicated); high but self-reported on AlphaFold 3's headline numbers, with independent benchmarks now the better source; thin on everything since 2026-06 (one X post, no interviews)
Conflict note (CONVENTIONS rule 25): Alphabet employee through 2026-06 with a commercial co-developer on his last major artifact; Anthropic employee on undisclosed terms since — moderate weight on LLM-for-science and drug-timeline claims, low weight on structure-prediction facts

Weak spots / open questions

AlphaFold is a singular achievement — unclear how well the approach generalizes to less structured scientific problems. Held — the co-folding memorisation result (Škrinjar, 2025), PoseBench's novel-pose finding (2026), CASP16's ligand parity, and the GNoME critique (Cheetham & Seshadri, 2024) all point the same way: the method generalizes to problems that look like the training set
Protein structure prediction ≠ protein function prediction — the shape doesn't tell you what the protein does in context. Held — he says so himself ("not one protein structure away from curing any diseases," MIT TR, 2025-11); mutation effects remain out of reach (Buel & Walters, 2022)
Drug discovery applications are still early — structure is necessary, not sufficient. Held — no dosed patient on any fetched source; target slipped from end-2025 to end-2026 (Reuters, 2026-01-20). This is the check that will move first
Jumper is primarily a scientist, not a public commentator — his views on AI governance, safety and societal impact are less developed. Drifted — the absence search found one governance act (AF3's 50-expert biosecurity review, 2024), one AGI line (2025-01), and a move to the lab that brands itself on safety (2026-06) with no stated reason. Still thin, but no longer blank; the next validation should search his Anthropic-era statements specifically
The Nobel Prize creates a halo — AlphaFold's success doesn't validate all AI-for-science claims but is used that way. Held — the GNoME episode is the halo in action; Anthropic's and Isomorphic's science marketing both lean on the AlphaFold name
New: his Anthropic role is undisclosed. Any statement that he "leads biology" or "leads AI for science" at Anthropic is invention until a primary says so; the file will carry the gap until it closes
New: is the 2025-11 LLM bet a scientific judgment or a recruiting narrative? The order of events (statement first, move seven months later) favours judgment; the next restatement from inside Anthropic will be the test
New: his stated titles disagree across primaries (Director 2024; Distinguished Scientist 2025; "Director and senior research leader" in 2026 press; "VP and Engineering Fellow" on one blog, unverified). Minor, but a reminder that the bench's affiliation lines need a source per title

Rich's take

(your synthesis here)

Delta log

2026-09-21 — v1→v2 migration + validation (bulk run)

Grades: 12 Held / 8 Drifted / 1 Wrong (plus 2 unfalsifiable items marked and rewritten, and 2 new live predictions attached with checks, not graded). Review window: 2026-06-16 → 2026-09-21, treated as never-reviewed per the May/June-pass rule; effectively everything after 2024 was re-checked.
Wrong: affiliation — v1 said "Senior Research Director, DeepMind" on a review stamped 2026-06-16; he announced his departure for Anthropic three days later (X, 2026-06-19; TechCrunch, 2026-06-20), and the title itself appears on no fetched primary (Director per DeepMind/EMBL 2024-10-09; Distinguished Scientist per the Royal Society, 2025). Another May/June-2026 v1 file confirmed defective on the employer check.
Drifted: "the approach generalizes" (scope yes, payoff not shown — Škrinjar 2025-08-04, PoseBench 2026-03-20, Elofsson 2025-08-25, CASP16 ligand parity, Cheetham & Seshadri 2024-04-08); "open access amplifies impact" (mechanism Held on 4.5M users, behavior retreated on AF3 — Nature editorial 2024-05-22, GEN 2024-06-13, GitHub README); implicit expansion-to-drug-design prediction (Isomorphic target slipped 2025→2026, Reuters 2026-01-20; no dosing on the 2026-05-12 Series B release); "replicable and open" track-record line (split into halves); "governance views less developed" weak spot (one governance act found); Hassabis, Koller and Amodei overlap/split lines (author-list and employer checks moved all three).
Held: AlphaFold solved structure prediction (CASP14 92.4 GDT; NAR 2023-11-02: 214M); AI-as-instrument is real (NAR 2025-11-22: 4,501,953 users); experimental-accuracy prediction; free-release-accelerates prediction; Nobel, DB-usage and CASP14 track-record lines; four of five weak spots; the Marcus/Mitchell split on the fold itself.
Unfalsifiable, rewritten: "scientific AI requires domain expertise" (observable attached: CASP17/18 winner without domain depth, or an LLM agent); "generally intelligent AI systems at some point" (no horizon — his only found AGI line, Chicago Maroon 2025-01-23).
Most surprising delta: the bench's proof-case for purpose-built scientific AI moved to the LLM lab — seven months after saying "I'll be shocked if we don't see more and more LLM impact on science" (MIT Technology Review, 2025-11-24), and as the capstone of a dated Anthropic build-out (Claude for Life Sciences 2025-10-20; Allen/HHMI 2026-02-02; Coefficient Bio $400M 2026-04-03; Karpathy 2026-05-19). Says and Does agree; the incentive map is what changed.
Employer/role/funding check (mandatory): run 2026-09-21 on dated news — job change found (above); no role, title, team or start date disclosed by 2026-09-21; no post-2026-06-19 public statement from him found. Adjacent funding: Isomorphic $600M (TechCrunch, 2025-03-31) and $2.1B Series B led by Thrive with Alphabet, GV, MGX, Temasek, CapitalG and the UK Sovereign AI Fund (Isomorphic, 2026-05-12) — he is not at Isomorphic, but AF3 was co-developed there and its trial timeline is the payoff check on his generalization claim.
Correction caught in-run: a search summary attributed an FDA clearance for "ISM8969" to Isomorphic; the ISM prefix is Insilico Medicine's and the claim did not enter this file. The Hassabis bench file's "Isomorphic launched human trials 2026" line is unsupported by every primary fetched here through 2026-05 — flagged for that file, not edited (one-persona rule).
Incentive map: re-derived — Alphabet/Isomorphic co-ownership of AF3 explains the gated release; Anthropic employment (terms undisclosed) now sits under every LLM-for-science claim. Discount raised from "minimal" to moderate on forward claims; unchanged (low) on AlphaFold facts.
Tier: quarterly — an undisclosed role at a new employer, a first-dosing deadline he doesn't control but his claim depends on (2026-12-31), and CASP17 results all land within two quarters. Wake triggers cover role disclosure, first Anthropic byline, Isomorphic dosing or miss, rival structure models, CASP17, any AGI/safety/open-release statement, any new seat.
Confidence: medium, unchanged — the AlphaFold record is the best-sourced on the bench, but everything since 2026-06 rests on one X post and the absence of anything else.

Sources

migrated v1→v2 2026-09-21 · backup: _archives/john-jumper.html.bak-20260921 · converts-from: people/john-jumper.md · AI & Society domain