diff --git a/compendium/00-what-this-is.md b/compendium/00-what-this-is.md index 370b236..a8bdf97 100644 --- a/compendium/00-what-this-is.md +++ b/compendium/00-what-this-is.md @@ -35,4 +35,10 @@ The compendium tracks those four layers separately so a new model release does n 1. `01-maps-of-intelligence.md` — the main human models 2. `02-measurement.md` — how intelligence is scored 3. `03-frontier-as-of-2026-09.md` — snapshot of the artificial frontier -4. Changelog at the bottom of each file +4. `04-socher-eureka-machine.md` — unverified 2026-10-04 lead on Richard Socher and *The Eureka Machine* +5. Changelog at the bottom of each file + +## Changelog + +- 2026-09-13: v0.1 introductory map. +- 2026-10-04: reading order gains an unverified Socher lead. The index stays 0.1. The lead does not revise the maps. diff --git a/compendium/01-maps-of-intelligence.md b/compendium/01-maps-of-intelligence.md index 32f37a1..9e4da9b 100644 --- a/compendium/01-maps-of-intelligence.md +++ b/compendium/01-maps-of-intelligence.md @@ -80,6 +80,9 @@ For routing work to models, Intelligence Radar uses a **capability stack**, not A task is scored by which layers it actually needs. Benchmarks are chosen to hit those layers. Theories from sections 1–7 explain *why* a layer is hard; they do not replace the score. +A 2026-10-04 research lead, [Richard Socher and *The Eureka Machine*](04-socher-eureka-machine.md), describes scientific discovery as a loop of records, physical models, simulation, and laboratory measurement. It is an engineering recipe for collective intelligence. It is not a revision of the seven layers above. Two secondary accounts of Socher's own dimension list disagree with each other. Neither account replaces this stack. + ## Changelog - 2026-09-13: v0.1 initial map. +- 2026-10-04: pointer to the unverified Socher lead. The seven-layer stack is unchanged. diff --git a/compendium/04-socher-eureka-machine.md b/compendium/04-socher-eureka-machine.md new file mode 100644 index 0000000..b918f91 --- /dev/null +++ b/compendium/04-socher-eureka-machine.md @@ -0,0 +1,105 @@ +> Research lead, 2026-10-04. The book has not been captured. Nothing in this file is a reviewed fact, a benchmark result, or a change to advisor policy. See [integration review](../docs/integration-review.md). + +# Richard Socher and The Eureka Machine + +Version 0.1 · 4 October 2026 + +The operator asked for a digest of Richard Socher's work, starting from notes at a Gemini share that uses the German spelling *Heureka*. The published title is *The Eureka Machine: Why AI Is the Key to Unlocking a New Era of Scientific Discoveries* (PublicAffairs, on sale 22 September 2026, hardcover ISBN 978-1-5417-0570-8, 288 pages). + +This file records that request. It separates three layers: + +1. What publishers and Socher's own organizations say he does. +2. What a disclosed-conflict review by someone who read an advance copy reports about the book's argument. +3. What the operator's Gemini notes add, including a ten-part list that a second secondary account does not match. + +Durable claims still need the book, the cited papers, or another primary. A thinker being relevant is a reason to watch him. Source catalog policy already says that is not a reason to accept his claims. + +## 1. The work this facility can name without the book + +PublicAffairs' author page describes Socher as CEO and cofounder of Recursive, CEO and founder of You.com, and founder and managing partner of AIX Ventures. It places him earlier as a Stanford computer-science PhD and adjunct professor, and as chief scientist and EVP at Salesforce after the acquisition of MetaMind. The same page credits "advances in word vectors and inventing prompt engineering." The word-vector lineage is the part other public accounts repeat. The "inventing prompt engineering" sentence is the publisher's attribution. This repository has not checked it against the priority dispute, and it does not adopt it. + +Rajiv Pant's January 2026 career essay, written by a disclosed You.com shareholder and informal adviser, attaches two more specific research names: GloVe, and DecaNLP (2018) as a precursor to stating tasks in natural language. Pant also quotes Socher on the gap between a prototype and accuracy at scale, and on a management claim Pant dates to late 2025: the future of work is people becoming managers of AI, which means delegating, specifying requirements, and building trust. Pant's citation counts and quotations are Pant's. They are leads toward the papers and the talks. + +Recursive's public description, repeated on the book site, is a company "building self-improving superintelligence to automate knowledge discovery." That sentence is a mission. It is not an execution profile, a benchmark, or an outcome. + +## 2. The book's argument, as a review with the book in hand reports it + +Pant published a review on 6 September 2026 from an advance digital copy. He discloses friendship, Young Global Leaders overlap, a You.com shareholding, and a past formal advisory role. Several biomedical examples in the book, he writes, come from companies Socher discloses investing in. Treat the review as a situated reading with unusual access, and follow its bibliography back to the papers. + +The argument Pant extracts: + +- Scientific AI has to learn across language, equations, spectra, molecules, and measurements. A model trained on writing about oxygen has not absorbed the measurements. Pant cites this to p. 48. +- The bottleneck is larger than paper volume. Specialization separates a result in one field from a problem in another, and institutions decide what receives time, money, and publication. Socher's line, as Pant quotes it from p. 74: "Collectively, we know far more than we understand." +- Socher qualifies the "science is slowing" premise. Progress may now consist of connecting knowledge about complex systems, rather than only the isolated breakthrough associated with earlier eras (p. 58 in Pant's citation). Pant points at Park, Leahey, and Funk, *Nature* 2023, on a falling *share* of disruptive papers alongside a stable absolute count of highly disruptive papers and patents, and at a methodological critique (arXiv:2306.01949). This file has not re-read those studies. +- Search can be aimed at connections outside the paths researchers already walk. Pant cites Sourati and Evans on what their paper calls "alien" hypotheses (arXiv:2306.01495) and says Socher marks that work as a proof of concept (pp. 68–71). +- Chapter 8 states the machine as four connected parts (pp. 184–195): a knowledge base spanning literature, code, images, and measurements; a model grounded in physical data; simulations of interventions; and laboratories that make and measure things and return the disagreement. Agents coordinate literature search, experiment proposals, and result examination. Pant quotes the constraint Socher puts on the laboratory: "There is no substitute for reality" (p. 188). The parts, Pant says, exist in use or development and do not yet form the combined system. +- Biological examples are the clearest contact with matter, and each has its own endpoint. Pant's account of AlphaFold separates coverage of the human proteome from confidence on residues. His account of ProGen, work Socher participated in, separates a million generated sequences from a selected set that was assayed: functional in a lysozyme assay is not a medicine. His account of an AI co-scientist separates cell-model tests from clinical success. A virtual cell is an abstraction problem: which simplifications still predict the intervention being asked about. Pant reports Socher forecasting a tenfold increase in the pace of biological research and drug discovery over the next decade (p. 113) and says the book does not say which quantity increases tenfold. +- Economics, neuroscience, and astronomy are there to show that one score cannot cover them. A simulated tax policy is a result inside a simulation. A brain metaphor is not a theory. An astronomical reconstruction mixes measurements with assumptions. Pant reports a thought experiment about a model that predicts well inside a mistaken cosmology and still fails to surface the bad assumption (pp. 175–177). +- Failed experiments belong in the record that the next experiment can use (p. 191). Self-improvement examples Pant names, the Darwin Gödel Machine and AlphaEvolve, improve workflows or candidate programs against an evaluation that a person supplied. Pant's distinction: name the object that improved. A workflow, a frozen foundation model, a trained model, and a whole scientific process are different claims. +- The closing chapter proposes knowledge and agency as principal components, then capabilities that include perception, memory, reasoning, physical action, social coordination, creativity, and self-assessment (pp. 208–227). Pant writes that Socher treats the categories as overlapping and the formalization as unfinished. Human performance is not assumed to be the ceiling. Who chooses the objective, and who can use the result, stays a human and institutional question. Pant cites Socher's question about whether cell-modeling efforts remain open once they are valuable (pp. 111–112, 199–200). + +Kirkus (review posted online 17 August 2026) calls the book an optimistic insider account and paraphrases a late speculative passage about an advanced intelligence seeking control of its own computational substrate. That paraphrase is a reviewer's impression of one passage. It is not the architecture above. + +## 3. The operator's Gemini notes + +Share: . The page says the conversation was created with Gemini 3.8 Flash on 4 October 2026 at 12:03 PM and published at 01:27 PM. The rendered text was read the same day. Gemini's own page says its statements about people can be wrong. The notes are the operator's initial cut. They are not a substitute for the chapters Pant cites. + +The notes organize the book as five ideas: + +1. **Fragmentation sclerosis.** Paper counts are high, breakthrough velocity is down, and the cause is specialization so narrow that a person cannot synthesize across fields. AI is the connective tissue over multimodal scientific corpora. +2. **A three-phase loop.** Ideation (hypotheses outside the researcher's orthodoxy), implementation (robotics, closed-loop simulation, code), and validation (predictive checks, Bayesian updates, targeted real-world tests). +3. **Biology before physics.** The notes call this Socher's contrarian bet: fundamental physics is limited by scarce instruments and slow cycles; biology is a large system of known parts with nonlinear interactions, so learning from data fits it. Virtual cells and organoids turn drug search into parallel candidate pipelines. +4. **Grounded recursive improvement.** Self-improving agents that write code, run simulations, and test against ground truth, with reward hacking named as the hard problem. Software can iterate faster than chip fabrication, labs, and energy, so the notes describe a long physical integration rather than an overnight singularity. +5. **Concrete governance.** Human control of lethal autonomous weapons and critical infrastructure, and open verifiable science so a few sovereign or corporate actors do not monopolize the record. + +A second question in the share asks for ten dimensions of intelligence. The rendered answer drops several headings, so the list below is reconstructed from the prose that survived rendering. It is Gemini's list, attributed by Gemini to Socher: + +1. Perception across modalities +2. Language, including pragmatics +3. Working and long-term memory +4. Formal and multi-step reasoning +5. Planning under constraints +6. Cross-domain creativity and hypothesis generation +7. A world model of cause, permanence, and space +8. Social and emotional intelligence, including theory of mind +9. Sample-efficient learning without catastrophic forgetting +10. Physical or digital agency that closes the perception–action loop + +Pant's account of the final chapter does not match this list. He reports two principal components, knowledge and agency, and a capability set of perception, memory, reasoning, physical action, social coordination, creativity, and self-assessment. Language, planning, an explicit world model, and sample-efficient learning appear in the Gemini list and not in that sentence of Pant's. Self-assessment appears in Pant's sentence and not in the Gemini list. Both documents are secondary. This facility keeps both until chapter 8 and the final chapter are captured. It does not install either list as the capability stack. + +The three-phase loop and the four-part machine can sit together without being the same diagram. The phases are a cycle of propose, intervene, and check. The four parts are the stores and tools the cycle uses: records, a physical model, a simulation, and a measurement that can disagree. Agents are the coordination layer on top of both accounts. + +## 4. Where this already meets the intent + +INTENT asks two questions: what there is reason to believe about intelligence, and which execution profile to use for a task today. Socher's public project is aimed at a third object this compendium already tracks, collective intelligence, and at one engineering recipe for it: close the scientific loop so a result in one field can change an experiment in another. + +The agreements that matter here: + +- **Intelligence stays a stack.** INTENT's non-goal is a universal score. The working synthesis in `01-maps-of-intelligence.md` already routes work through seven layers. Socher's final chapter, in both secondary accounts, also refuses a single scalar. The compendium's measurement file already says scientific novelty is not what current exams measure. +- **A benchmark is an instrument with a domain.** Pant's point that a protein assay, a simulated economy, and a sparse astronomical image are different tests is P3 and P5. Five correlated leaderboards are not five observations. A scientific task needs instruments that can fail when the intervention changes, which is a stronger demand than a saturated knowledge exam. +- **Outcomes outrank the leaderboard.** "There is no substitute for reality" is P9 in laboratory language. The v0.2 advisor already withholds Frontier, Efficient, and Budget on the source-backed example because the evidence is incomplete. That withholding is the behavior this lead supports. +- **Price is not token price, and access is part of the choice.** A generated protein that still needs a lab, or a cell model that stays inside one company, is a deployment question. Available Frontier already means the profile the requester can actually use. +- **Decomposition is an optimization, and the objective stays human.** Specialist agents for search, simulation, and criticism are P12. Pant's reading of the economics chapter leaves welfare weights with people and their representatives. The advisor input contract already takes acceptance criteria and risk from the requester. +- **Disagreement stays on the record.** Failed experiments, a simulation that the lab contradicts, and two secondary lists of "dimensions" are all cases of P10. The compendium's job is to keep them. + +The manager-of-AI claim from Pant's earlier essay maps to a narrower operational point. Directing a system, specifying the task, and checking the result are the advisor's inputs and the outcome ledger. They do not require adopting Pant's "synthesis engineering" vocabulary. + +## 5. Nudges + +The ideas this lead suggests for later exploration live in [nudges/](../nudges/README.md). They are notes. They are not policy edits, catalog edits, or task-status changes. The index lists nine notes from 2026-10-04, each pointing back here. + +## 6. Sources read for this lead + +- Operator notes, Gemini share, 4 October 2026: +- Book site, read 4 October 2026: +- PublicAffairs title page: +- Kirkus review: +- Rajiv Pant, "The Eureka Machine: What It Takes for AI to Do Science," 6 September 2026: +- Rajiv Pant, career essay, 25 January 2026: + +Papers named above are named because those secondary sources name them. They have not been ingested. + +## Changelog + +- 2026-10-04: v0.1 unverified lead, from the operator's Gemini notes and public secondary sources. No advisor policy change. +- 2026-10-04: the exploration ideas moved to `nudges/`. This file keeps the digest. diff --git a/nudges/2026-10-04-capture-eureka-chapters.md b/nudges/2026-10-04-capture-eureka-chapters.md new file mode 100644 index 0000000..8670fd1 --- /dev/null +++ b/nudges/2026-10-04-capture-eureka-chapters.md @@ -0,0 +1,9 @@ +# Capture the chapters and the named papers + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Keep this lead in the compendium until the book is a capture." + +The compendium lead rests on the operator's Gemini share and on Rajiv Pant's review of an advance copy. A later reading would capture chapter 8 and the final chapter from a source the operator holds, then the papers that review already names, and append a correction to the lead. The 13 September maps stay as they are. + +The Gemini share is an operator note: . Pant's review is . diff --git a/nudges/2026-10-04-decomposition-stays-unranked.md b/nudges/2026-10-04-decomposition-stays-unranked.md new file mode 100644 index 0000000..dc8e85e --- /dev/null +++ b/nudges/2026-10-04-decomposition-stays-unranked.md @@ -0,0 +1,9 @@ +# Leave decomposed science unranked + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Keep decomposed science unranked until it is measured against one profile." + +Four agents around a literature search, a plausibility check, a simulation, and a lab are a hypothesis that collaboration beats one Frontier profile on quality, cost, or time. INTENT principle P12 already treats decomposition as an optimization problem. Pant writes that the parts exist in use or development and do not yet form the combined system. + +[IR-WP-0007-T04](../workplans/IR-WP-0007-evidence-and-publication.md) already waits on a measured comparison with monolithic execution. A chapter describing the arrangement is not that comparison. This note does not edit that workplan. diff --git a/nudges/2026-10-04-instrument-slowing-science.md b/nudges/2026-10-04-instrument-slowing-science.md new file mode 100644 index 0000000..a54dc74 --- /dev/null +++ b/nudges/2026-10-04-instrument-slowing-science.md @@ -0,0 +1,9 @@ +# Give the slowing-science claim an instrument + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Give the slowing-science claim an instrument before it becomes a premise." + +Pant's review says Socher treats slowing progress as a broad premise and then qualifies it: progress may now be connection and synthesis. Pant points at Park, Leahey, and Funk, *Nature* 2023, for a falling share of disruptive work beside a stable absolute count, and at a methodological critique (arXiv:2306.01949). This repository has not re-read those studies. + +Two forecasts in the public material have no declared numerator: "a century of breakthroughs in a decade," and a tenfold rise in biological research and drug discovery. Hypotheses, completed experiments, and validated treatments are different quantities. Until one is declared, the forecast stays a disputed claim with no sufficiency threshold. diff --git a/nudges/2026-10-04-label-evidence-layer.md b/nudges/2026-10-04-label-evidence-layer.md new file mode 100644 index 0000000..d02f482 --- /dev/null +++ b/nudges/2026-10-04-label-evidence-layer.md @@ -0,0 +1,9 @@ +# Label the evidence layer on a recommendation + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Label the evidence layer on a recommendation." + +Pant's account of chapter 8 describes four parts: records, a model of physical regularities, a simulation, and a measurement that can disagree with the model that proposed the action. For a scientific task, a recommendation could say which of those parts its evidence occupies. + +Literature-only evidence supports a knowledge-layer report. It does not support a Sufficient or Budget winner. The v0.2 source-backed example already withholds those winners. The idea is to make the missing layer visible in the report. INTENT principle P9 is the policy this would illustrate. It is not a change to `policies/advisor-v0.2.json`. diff --git a/nudges/2026-10-04-openness-as-availability.md b/nudges/2026-10-04-openness-as-availability.md new file mode 100644 index 0000000..48e7f12 --- /dev/null +++ b/nudges/2026-10-04-openness-as-availability.md @@ -0,0 +1,9 @@ +# Treat openness as availability + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Treat openness as availability." + +Pant cites Socher asking whether cell-modeling efforts stay open once they produce valuable discoveries. A profile whose weights, data, or lab protocol the requester cannot inspect or use sits outside Available Frontier for that requester, whatever its paper reports. The deployment definition in INTENT already includes access, privacy, and product constraints. + +What remains is a watch: when a named cell model, dataset, or protocol changes who can use it, that change is a deployment observation. This note does not add a ranking feature. diff --git a/nudges/2026-10-04-physical-iteration-cost.md b/nudges/2026-10-04-physical-iteration-cost.md new file mode 100644 index 0000000..43bb616 --- /dev/null +++ b/nudges/2026-10-04-physical-iteration-cost.md @@ -0,0 +1,9 @@ +# Price physical iteration as its own scenario + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Add physical iteration to cost only as a declared scenario." + +The operator's notes contrast biology's combinatorial search with physics' scarce instruments and slow cycles. Token scenarios do not price a wet-lab cycle, a telescope slot, or a chip run. INTENT principle P8 already requires retries, tools, and verification inside expected task cost. + +A scientific workload could carry an explicit cycle-time and instrument-access bucket, marked as an assumption until a dated observation supports it. A missing lab-time quantile stays missing. This note does not change the advisor cost schema. diff --git a/nudges/2026-10-04-reward-hacking-in-taxonomy.md b/nudges/2026-10-04-reward-hacking-in-taxonomy.md new file mode 100644 index 0000000..faa194b --- /dev/null +++ b/nudges/2026-10-04-reward-hacking-in-taxonomy.md @@ -0,0 +1,9 @@ +# Name reward hacking when a taxonomy is written + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Name reward hacking in the field-outcome taxonomy." + +The operator's Gemini notes call reward hacking the hard problem in autonomous science: an objective harness that measures discovery rather than a model exploit. Pant's review makes the same demand in laboratory language. The checker has to be able to disagree with the system that produced the result. + +Outcome records already require `independent_verification`. [IR-WP-0007-T02](../workplans/IR-WP-0007-evidence-and-publication.md) already owes a failure taxonomy before any paid sample is called representative. When that taxonomy is written, reward hacking and a verifier that shares the generator's errors are worth a place in it. This note does not edit that workplan. diff --git a/nudges/2026-10-04-stack-gaps.md b/nudges/2026-10-04-stack-gaps.md new file mode 100644 index 0000000..4bffa8b --- /dev/null +++ b/nudges/2026-10-04-stack-gaps.md @@ -0,0 +1,9 @@ +# Look at three stack gaps + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Examine three stack gaps." + +The seven-layer stack in [compendium/01-maps-of-intelligence.md](../compendium/01-maps-of-intelligence.md) stays as written. Both secondary accounts of Socher's dimensions press on a hole [compendium/02-measurement.md](../compendium/02-measurement.md) already names: scientific novelty and hypothesis generation. Pant's account of the final chapter adds self-assessment. The operator's Gemini notes add sample-efficient learning. The two lists do not match, and neither replaces the stack. + +A later look would ask whether each of those three constructs has an instrument with a version, a construct, and a known weakness. SciCode and Humanity's Last Exam do not become discovery tests by being relabeled. The Gemini list of ten dimensions stays in the compendium lead as an unverified reconstruction. diff --git a/nudges/2026-10-04-watch-socher-primaries.md b/nudges/2026-10-04-watch-socher-primaries.md new file mode 100644 index 0000000..a4431f8 --- /dev/null +++ b/nudges/2026-10-04-watch-socher-primaries.md @@ -0,0 +1,9 @@ +# Watch the primaries + +Nudge. Noted 2026-10-04. + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md) §5, "Watch the primaries, and watch Socher as a discovery route." + +Candidates for a later source-registry review: the book site (), the PublicAffairs title page, and Pant's review as a secondary index. Papers worth their own observations, each kept at the endpoint the paper actually reports: ProGen's assay, the AlphaFold human-proteome study, the co-scientist paper, Sourati and Evans, the Darwin Gödel Machine, and AlphaEvolve. + +Recursive and You.com are already discovery lines on [watchlist/SOURCES.md](../watchlist/SOURCES.md). A mission statement about self-improving superintelligence becomes an observation only after it names the object that changed and the test that accepted the change. diff --git a/nudges/README.md b/nudges/README.md new file mode 100644 index 0000000..b06f72f --- /dev/null +++ b/nudges/README.md @@ -0,0 +1,19 @@ +# Nudges + +Raw ideas to maybe explore later. A nudge is a note. Work starts only when someone deliberately writes an intake or a workplan. Nothing in this directory is scheduled, assigned, or synced to State Hub. + +Each note names the context it came from. When an idea is taken up, leave the note here and add one line saying which work record received it. + +## 2026-10-04 — *The Eureka Machine* + +Derived from [compendium/04-socher-eureka-machine.md](../compendium/04-socher-eureka-machine.md). That lead digests the operator's Gemini notes and public secondary sources. The book has not been captured. + +- [Capture the chapters and the named papers](2026-10-04-capture-eureka-chapters.md) +- [Watch the primaries](2026-10-04-watch-socher-primaries.md) +- [Give the slowing-science claim an instrument](2026-10-04-instrument-slowing-science.md) +- [Label the evidence layer on a recommendation](2026-10-04-label-evidence-layer.md) +- [Look at three stack gaps](2026-10-04-stack-gaps.md) +- [Name reward hacking when a taxonomy is written](2026-10-04-reward-hacking-in-taxonomy.md) +- [Leave decomposed science unranked](2026-10-04-decomposition-stays-unranked.md) +- [Price physical iteration as its own scenario](2026-10-04-physical-iteration-cost.md) +- [Treat openness as availability](2026-10-04-openness-as-availability.md) diff --git a/watchlist/SOURCES.md b/watchlist/SOURCES.md index 1b8ec52..849fdb5 100644 --- a/watchlist/SOURCES.md +++ b/watchlist/SOURCES.md @@ -16,6 +16,7 @@ Daily monitoring prefers **primary outlets**. Commentary is used to find primari - Mistral, Cohere - DeepSeek, Alibaba Qwen, Moonshot Kimi, Z.ai / Zhipu GLM, MiniMax, Tencent Hunyuan, Thinking Machines - Cognition, Adept-class agent firms, World Labs, Pathway (architecture outliers) +- Recursive and You.com — added 2026-10-04 as discovery routes for Richard Socher's scientific-AI claims. A company mission is not a measurement. See `compendium/04-socher-eureka-machine.md`. ## B. Measurement houses (primary for the ledger) @@ -57,7 +58,7 @@ Daily monitoring prefers **primary outlets**. Commentary is used to find primari Keep the list short and rotate. Current standing set: -**Builders / research leads:** Dario Amodei, Demis Hassabis, Shane Legg, Ilya Sutskever, Noam Brown, Jared Kaplan, Jack Clark, Sepp Hochreiter, Yann LeCun, Geoffrey Hinton, Yoshua Bengio, Fei-Fei Li, Oriol Vinyals, Chelsea Finn, Jim Fan, John Schulman-class post-training leads, Moonshot / DeepSeek / Qwen chief scientists when they publish. +**Builders / research leads:** Dario Amodei, Demis Hassabis, Shane Legg, Ilya Sutskever, Noam Brown, Jared Kaplan, Jack Clark, Sepp Hochreiter, Yann LeCun, Geoffrey Hinton, Yoshua Bengio, Fei-Fei Li, Oriol Vinyals, Chelsea Finn, Jim Fan, John Schulman-class post-training leads, Moonshot / DeepSeek / Qwen chief scientists when they publish. Richard Socher, added 2026-10-04: read the papers and *The Eureka Machine* in the original. Secondary summaries, including the operator's Gemini notes, are discovery aids. **Measurement / forecasting:** Epoch AI authors, Artificial Analysis, ARC Prize (Chollet and successors), METR / eval orgs, Ajeya Cotra–class compute forecasts, Stanford HAI index leads.