What we mean
- Prediction
- A model's estimate of what may happen or what a measurement may show.
- Experimental confirmation
- Evidence from an experiment that tests a particular prediction or explanation.
- Demonstrated benefit
- An improvement observed in the setting where the result is intended to be used.
Values we bring to the question
- Expand access to scientific understanding.
- Connect promising findings to careful testing and practical benefit.
Reasoning and the proposed connection
- AI can make large spaces of possible explanations or candidates easier to investigate.
- Laboratories, domain knowledge, and independent review help determine which candidates withstand testing.
- In our proposed learning system, an unresolved shadow could be revisited when a new sensor, concept, or method becomes available. A relationship from another field could suggest a prediction that competing explanations do not share.
- Following data, interpretation, experiment, and outcome across layers could help independent groups decide what to investigate or build. The choices of questions and beneficiaries also involve values; distinguishing evidence from values does not remove them from research.
Where the reasoning stops
The examples do not establish that the whole genome is understood, that aging is solved or that human research work is unnecessary. Research announcements and selected validations have narrower evidential scope.
The strongest objection
More candidates may overwhelm validation capacity. Attractive analogies can rely on similar words while transferring no useful mechanism. A compelling announcement or summary can conceal a weak interpretation, selection effects, missing conditions, or substantial human work.
Read the evidence
Each source has a specific role in the stated claim. Its findings, review date, and access limits are recorded below. Our proposed architecture and experiments require their own tests; an editorial revision does not mean the source was reviewed again.
R-SCI-01 · Supports part of this claim (or shows a bounded mechanism)
AlphaGenome Atlas: Molecular predictions for 9 Billion human DNA variants
Google DeepMind and research collaborators · Research announcement
Read: Author summary · Primary source reviewed
Published: 2026-09-08 · Reviewed: 2026-09-30
A molecular prediction atlas offers selected experimental examples; the reviewed source is a research announcement.
- What it reports
- The atlas provides large-scale predictions of the molecular effects of single-letter changes in human DNA, with selected examples of experimental use.
- Limits
- Billions of predictions are not billions of laboratory validations. Selected validation does not establish clinical validity for the whole atlas or a solution to aging.
Review scope and version
Read: Author summary.
Primary research announcement
This illustrates AI-assisted scientific hypothesis generation. Predictions, laboratory findings and demonstrated clinical benefits remain different stages.
Link to this source noteR-SCI-02 · Supports part of this claim (or shows a bounded mechanism)
Claude discovers a novel enzyme system with CRISPR-like repeats
Anthropic and experimental research collaborators · Research announcement
Read: Author summary · Primary source reviewed
Published: 2026-09-23 · Reviewed: 2026-09-30
Agent-assisted research identifies a candidate system for laboratory investigation while its biological function remains open.
- What it reports
- Agents working from a human-provided research direction helped identify a system involving reverse-transcriptase-related repeats and accessory components, which was taken forward for laboratory investigation.
- Limits
- The system's biological function remains open. A company announcement is not independent replication or clinical validation; human direction and laboratory work remain part of the process.
Review scope and version
Read: Author summary.
Primary research announcement
The work illustrates agent participation in discovery, rather than establishing that the whole scientific process has become autonomous.
Link to this source noteR-AG-09 · Supports part of this claim (or shows a bounded mechanism)
AlphaEvolve: A coding agent for scientific and algorithmic discovery
Alexander Novikov and co-authors · Research white paper
Read: Selected sections · Primary source reviewed; no independent reproduction
Published: 2025-06-16 · Reviewed: 2026-10-04
The paper reports algorithmic discoveries under automated evaluation; scientific interpretation and wider consequences require further judgment.
- What it reports
- An evolutionary pipeline combines LLM-generated code with supplied evaluators, including evolving search algorithms. Reported results include a 48-multiplication algorithm for two 4-by-4 complex matrices and optimizations to Google's computational infrastructure.
- Limits
- The main boundary is availability of automated evaluators. The paper describes self-improvement feedback as modest and occurring over months. It does not demonstrate unrestricted recursive improvement, ethical goal selection or Leviathan's architecture. Results and deployments were not independently reproduced here.
Review scope and version
Read: Selected sections.
arXiv:2506.13131v1; white paper submitted 16 June 2025
Read v1 task definition, code search, evaluation, selected results and discussion. Reported code and deployments were not reproduced; supplied evaluators constrain the demonstrated discovery process.
Related source links
R-AG-13 · Supports part of this claim (or shows a bounded mechanism)
GitSwarm: Decentralized Compounding Inference
Meta Superintelligence Labs and University of Texas at Austin researchers (as indexed by DAIR.AI) · Preprint
Read: Indexed technical digest (DAIR.AI Papers of the Week 2026) · Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced
Published: 2026-10 · Reviewed: 2026-10-21
In GitSwarm’s decentralized repository workflow, 70.1% of failed research commits are cited by later accepted work (reviewed via indexed digest). This supports preserving failed attempts with explicit lineage in an archival repository.
- What it reports
- Asynchronous agents collaborate through a shared Git repository without a central planner, reading prior commits and tests before publishing new commits with explicit lineage. In the reported ML research benchmark, 70.1% of failed research commits are cited by later accepted work.
- Limits
- Reviewed via the curated DAIR.AI technical summary; experiments were not rerun. Coordination over a shared Git repository with benchmark evaluators does not solve value disagreement, power asymmetries, or governance among independent groups.
Review scope and version
Read: Indexed technical digest (DAIR.AI Papers of the Week 2026).
arXiv:2610.04862; indexed in DAIR.AI Papers of the Week (Oct 5–Oct 11, 2026)
Provides quantitative support within a coding and ML research setting for preserving failed attempts with explicit lineage rather than discarding negative results.
Related source links
R-LRN-07 · Limits the inference
Scaling Discovery Requires Forgetting
UC Berkeley, Stanford University, Bespoke Labs, and University of Texas at Austin researchers (as indexed by DAIR.AI) · Preprint
Read: Indexed technical digest (DAIR.AI Papers of the Week 2026) · Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced
Published: 2026-10 · Reviewed: 2026-10-21
Keeping full attempt history inside an agent’s active working context can cause anchoring and plateauing during search, whereas periodically resetting working context to the best candidate improved discovery by 15%–39% (reviewed via indexed digest). Archival retention and active prompt context serve different roles.
- What it reports
- In iterative test-time scientific and algorithmic discovery, accumulating full interaction and attempt history in context causes agents to anchor on minor edits and plateau. Periodically discarding session history and restarting from only the best discovered candidate improves performance across discovery tasks by 15% to 39%.
- Limits
- Reviewed via the curated DAIR.AI technical summary; experiments were not rerun. The benefit of clearing working-context history during search does not mean archival provenance or negative results should be deleted from a shared research library.
Review scope and version
Read: Indexed technical digest (DAIR.AI Papers of the Week 2026).
arXiv:2610.20563; indexed in DAIR.AI Papers of the Week (Oct 19–Oct 25, 2026)
Clarifies the distinction between an archival record of failed attempts and an agent’s active working context during search.
Related source links
What could change our view?
Independent experiments that fail to confirm useful predictions would narrow the claims. Reproducible findings and demonstrated benefits in intended settings would strengthen them.
The next question
For a promising connection, what new prediction does it yield, and which observation could distinguish it from a simpler explanation? Track the actual links from data to independent experiment and practical use. For Leviathan, compare whether preserving unresolved questions and their context helps another group find a useful test when new methods become available.
Revision record
Added two examples of AI-assisted science, with separate labels for prediction, laboratory investigation and demonstrated benefit.
Linked scientific candidate generation to proposed shadows, new sensors, cross-field hypotheses, and independent experiments. Clarified the role of values in choosing research questions. Existing scientific examples, evidential stages, and source reviews remain unchanged.
Reviewed selected primary-source sections proposed in the external assessment and added scoped connections, access limits and research questions. Interlat retains its existing source ID; its reviewed preprint is distinguished from the final conference text. No experiment was reproduced, claim status promoted or governance rule adopted.
Added GitSwarm (R-AG-13) on citation of failed research commits and Scaling Discovery Requires Forgetting (R-LRN-07) to distinguish archival lineage from active working-context resets during search. Both sources were reviewed via the DAIR.AI 2026 technical digest without independent reproduction.
This note records an editorial position. Independent people and groups can bring another interpretation, a useful method, or an objection to the framing. Explore it with your own assistant if helpful and choose what to share. Use the exact revision below when contributing; the dated history explains why our account changed but does not uniquely identify a revision.
Cite this revision
This snapshot preserves C07 and the source cards shown with it. Copy its link into a question or objection so a later reader can recover the version you read.
Read or save the exact revision (JSON)C07 · SHA-256: 6242e829a06f61edb6dab72e4f82ca43ae3fd9a08cf9d3203e30d42c7853a28b
The digest covers the complete downloaded file, including its final newline. It identifies this record; it does not establish that its claims are true. External papers and linked pages may change. Earlier dated history entries do not imply that their full snapshots were captured.
Bring a question about this revision