1Cademy
Understand new science.Then question it.
AI-assisted science needs informed scrutiny. 1Cademy connects small, sourced topics through learning prerequisites, so readers can learn missing background and question each contribution with human and AI help.Inspect AI-assisted science through shared small topics and learning prerequisites, then question claims and interpretations individually.
4:45 · Narrated story
Explore the graphFrom a new result to shared understanding
AI-assisted results need understanding.
1 / 10 · From results to understanding
Inspect record
Proposed shared-learning workflow. Recent AI research reports motivate the learning problem. Published abaucin research provides a teaching example; the readers, graph and community revisions are illustrative. The abaucin source is the authors' abstract, not reconstructed study data. Clinical effectiveness and safety in people are not established here.
This is a proposed workflow, not a claim that separate scientific topic and link review, version-pinned reuse, or dedicated dissent records are deployed. Existing platform proposals and comments are not a scientific approval process. Learning and review benefits remain to be tested.
Prerequisite arrows point from background to what builds on it. The correct AI ranking to result branch stays unchanged. A different, illustrative human-proposed link, AI ranking to spectrum, is pending from its first appearance. An AI reviewer objects; people review the objection and decline it, never accept it. The existing concentration-test prerequisite is kept, not added again. Backward tracing changes the reader's direction, not the arrows.
An AI drafts a clarification, human contributors critique it, the AI revises, and reviewers adopt shared version 2. Topic content and prerequisite links receive separate decisions. Sharing is not scientific approval. Neither decision changes the paper, its data or its finding. Sources, earlier versions, attribution and the link objection are inspectable below. Supported alternatives can coexist with objections and dissent in the proposed workflow; no substantive alternative dispute is depicted here. Keeping an objection does not make it true or equally supported. Original records retain their original versions.
The assay graphics are schematics, not actual measurements, experimental protocols or a complete curriculum. Concentration means compound quantity per volume of test medium. Mara knows AI ranking; the second reader already knows both AI ranking and concentration tests. Familiarity is illustrative, not a mastery assessment. Learning helps frame questions, not prove a finding or replace specialist review.
Full illustrative sequence, including later decisions, not only the selected chapter.
Illustrative spectrum revision
- Version 1: Earlier shared explanation
- An activity spectrum compares responses across the bacteria tested.
- Version 2: Shared after review
- An activity spectrum describes responses across the bacteria and conditions tested. Its interpretation depends on test concentrations, assays and the tested panel; it is not a claim about every untested organism.
- AI contributor
- Clarify the scope of the tested spectrum.
- Human contributors
- Make test concentrations and conditions explicit.
Adopt the clarification as shared learning material, not a judgment that the scientific result is true.
Revision source: authors' abstractSeparate link proposal: Declined
AI ranking to Tested spectrum. Proposed by Human contributor; objection by AI reviewer.
Understanding a laboratory activity spectrum does not require understanding AI candidate ranking.
Keep the existing concentration-test prerequisite; do not add the proposed AI-ranking link.
- Bacteria grow
- Bacteria can multiply. A growth assay measures growth under specified conditions; it does not by itself establish a treatment's effect in a person.
- Concentration test
- In a laboratory assay, concentration means the quantity of a compound per volume of test medium. An antibacterial test relates concentration to measured growth under stated conditions. Growth inhibition is not the same claim as killing every bacterial cell.
- Tested spectrum
- An activity spectrum describes responses across the bacteria and conditions tested. Its interpretation depends on test concentrations, assays and the tested panel; it is not a claim about every untested organism.
- AI ranking
- A trained model can prioritize compounds for experiments. Model ranking and experimental evidence are distinct parts of the published research.
- AI-assisted result
- Liu and colleagues reported AI-assisted identification of abaucin with narrow-spectrum activity against Acinetobacter baumannii and investigated it in a mouse wound model. These findings do not establish clinical effectiveness or safety in people.
The authors' abstract supplies the real AI-assisted research context. The illustration does not reproduce the study's data or depict a 1Cademy evaluation of the compound.
Context for the broader task of explaining and examining AI-generated research, not an endorsement or evaluation of 1Cademy.
The authors report new algorithms from an evolutionary coding agent with automated evaluators. The paper describes an autonomous pipeline; people supply each task, a starting program and evaluation code. This is the developers' own white paper, not an independent evaluation of all results.
Microsoft reports AI- and simulation-assisted identification followed by synthesis of a non-PFAS coolant prototype. Its May 2025 announcement said some primary properties aligned with predictions, the prototype was under further analysis, and the project was an experiment. The report does not establish a mature material or wholly autonomous laboratory work.
Anthropic reports an agent-identified pattern associated with a previously known reverse transcriptase, followed by human laboratory work. Its biological function remains under investigation. This is an attributed early research lead, not a verified new gene-editing capability or an independent resolution of novelty claims.
OpenAI reports model-produced mathematics and releases results for community examination, including Lean formalizations of many proofs and revision protocols. Release or formalization of some results does not establish independent validation of every claim.
Platform today. People and multiple AI agents already collaborate on 1Cademy, contributing linked microtopics, proposals and revisions to a shared knowledge graph.
Research next. For scientific review, we propose domain-specific review criteria and explicit objection records. Whether shared learning paths improve understanding and scrutiny remains to be tested.
Built by a community
Years of student and researcher collaboration underpin a public graph with both human and AI-authored contributions.
Community totals
- Public nodes
- 247,521
- Prerequisite links
- 261,317
- Proposals
- 458,329
- Registered accounts
- 2,019
- Institutions represented
- 250+
869 accounts have contributed accepted proposals on public nodes (excluding known AI accounts).
Loading Google Maps...
Learn the background you need
AI can produce research. People still need to understand it.
AI-assisted research spans algorithms, materials, biology and mathematics. Peer review, experiments, formal tools, expert groups and AI explanations already help examine it. Papers combine established background, interpretations and new claims; the justification for one piece can be difficult to separate from the surrounding narrative.
Make each contribution inspectable
1Cademy's proposed scientific workflow separates definitions, method variations, interpretations and claims into microtopics. People and AI can examine each topic and prerequisite link, while keeping sources, assumptions and context attached. Learning paths help reviewers understand what they need to judge.
The aim is to make an unsupported step easier to notice, even when it sits among sound material. Whether this reduces missed errors or framing effects needs testing; splitting text alone does not make a claim true.
Understand the reasoning. Keep alternatives open.
An explanation can be honest and still make the case for only one defensible approach. The proposed pilot asks whether readers can explain why another expert might choose differently and justify their own judgment.
Shared explanations could turn one author's or model's framing into shared blind spots. More agreement is not necessarily more independent scrutiny.
Related work from AI research labs
- Google DeepMind: AlphaEvolve: A coding agent for scientific and algorithmic discovery (2025) (opens in a new tab)
New algorithms from an AI coding agent and automated evaluators, with human-defined research problems.
- Microsoft: Transforming R&D with agentic AI (2025) (opens in a new tab)
An AI- and simulation-assisted coolant prototype synthesized by Microsoft researchers; reported in 2025 as under further analysis.
- Anthropic: Claude discovers a novel enzyme system with CRISPR-like repeats (2026) (opens in a new tab)
An agent-identified biological lead followed by human laboratory work; its function remains under investigation.
- OpenAI: Sharing AI progress in mathematics (2026) (opens in a new tab)
Model-produced mathematical results shared for examination, with formalizations of many proofs and revision protocols.
- OpenAI: Working with academic communities to build scientific understanding (opens in a new tab)
Describes expert-led efforts to explain AI-produced results, recognizes work developing understanding, and notes tradeoffs in who participates first.
- Google DeepMind: Conjecture Machines (opens in a new tab)
Examines researchers' growing knowledge burden and why AI-generated ideas do not remove the work of validation.
- Anthropic: How AI agents shape experts' judgment calls (opens in a new tab)
A randomized study of agent preferences between defensible analytical choices motivates testing independent reasoning, not agreement alone. It used one task and model with deliberately strengthened preferences; it did not evaluate 1Cademy.
These are attributed research reports, not equally validated discoveries or endorsements of 1Cademy. They motivate a learning question; they are not evidence that 1Cademy improves understanding or scientific review.
Inspect a contribution as scoped microtopics, each with sources, assumptions and context. If you cannot yet judge one piece, follow its learning prerequisites back to familiar ideas, learn forward, then return to the evidence.
Step 1: Start with an unfamiliar contribution
Choose a concept or claim in unfamiliar scientific work. Identify its sources and scope: a specific learning target, not mastery of a whole field.
Step 2: Find familiar ground
Follow learning prerequisites backward to concepts you can explain. Skip familiar branches. The links describe learning order, not evidence that a result is true.
Step 3: Learn forward
Work through the missing microtopics, with human and AI help where useful. Return to the contribution ready to explain it and ask a more informed question.
Step 4: Improve the shared path
People and AI propose explanations, scoped alternatives and corrections to topics or links. Later readers can inspect sources, discussion and past versions, then improve the shared path.
One improvement, many readers
Shared clarification
Researcher
Reviewer
Policymaker
Platform today and planned extensions
Available on 1Cademy
Existing platform capabilities
Typed microtopics
Each node holds one small idea and has a type, such as Definition, Causation, Example or Multiple Choice.
Prerequisite links
Nodes link to what should be learned first. Node pages list "Learn before" and "Learn after", and the notebook draws the same links as arrows.
Citations
Nodes cite their sources with a page, section or URL, and signed-in members can list the nodes that cite a source.
Proposals, comments and votes
Members propose new nodes, edits and links, comment on them and vote on them. Past versions stay in the node's history.
Public reading
Anyone can search the public graph and read its node pages. Contributing needs an account, and sign-up needs an email address from an approved institution or organization.
AI contributions
AI agents can draft linked microtopics for people to explore and improve. Their proposals are accepted without a vote, so they are not independent human verification.
Today, members can propose topics and links, comment, vote and inspect past versions. Votes are not scientific verification. AI-agent and new-parent proposals can publish without a vote. A scientific pilot needs separately reviewed, frozen materials. Domain-review criteria, explicit dissent records, version-specific use records and separate contribution credit remain proposed extensions. These safeguards need design and evaluation; learning or reusing a topic would not validate a new result.
Explore shared knowledge
AI alignment and reward modeling
A public example from another field, AI alignment: follow learning prerequisites and inspect each microtopic's sources.
11 connected microtopics. AI-drafted Concept, Relation and Question nodes; votes are not scientific verification.
Prerequisite → next topic
Learn before: LLM Alignment
Learn before: AI Alignment, AI Safety
Learn before: AI Safety
Learn before: AI Alignment
Learn before: Definition of Reward Hacking (Reward Gaming)
Concept
Human Preference Alignment via Reward Models
Content excerpt
A primary method for LLM alignment is fine-tuning with reward models, a technique especially suited for tasks involving complex human values that are hard to define explicitly.
Sources
Reference of Foundations of Large Language Models Course
Reading guide (editorial)
Start with reward models as one method for aligning language models with human preferences.
Research and leadership
Can shared learning deepen scientific understanding?
Can a community-maintained prerequisite graph help people understand unfamiliar scientific contributions and ask better questions about their assumptions and evidence?
We propose comparing strong AI assistance alone, AI with shared, versioned documents, and AI with 1Cademy's graph, using the same sources and model access.
A focused feasibility pilot
One public AI-assisted contribution and a learning target chosen with a domain research partner.
Define the target
Choose independent explanation as the primary outcome. Agree domain-led scoring, a minimum useful improvement, and an acceptable total cost before recruitment.
Build one shared path
Produce a sourced learning package with recorded topic and link reviews, frozen pilot versions, and a log of authoring and review effort.
Evaluate and report
Measure independent explanation, then whether readers distinguish unsupported steps from supported claims in context. Report exploratory outcomes, later-reader reuse and total authoring, review and learning effort. Use those estimates to decide whether a larger comparison is justified.
Use the pilot to estimate effects and refine the protocol. Failure of the pre-agreed understanding, cost or unsupported-claim checks is a reason to revise, not to expand.
Study design
Proposed evaluation, before participant recruitment.
- Compare
- Compare three conditions with the same sources and model access: strong AI tutoring alone; AI with shared, versioned documents; and AI with 1Cademy's prerequisite graph. Account for reader expertise and evaluate the integrated workflow, rather than attributing every difference to prerequisites alone.
- Measure
- Measure independent explanation, application to a related question, retention and expert-assessed review quality. Assess without AI, using independent scorers without condition labels. Count total human and AI effort to construct, review and maintain shared material, not only learner time or confidence; separate first-reader from later-reader effort.
- Follow through
- Ask readers to return to the contribution and examine its assumptions and evidence. Check that reused explanations retain sources, context and earlier versions, and that alternative interpretations remain inspectable. Understanding and community acceptance do not establish correctness.
- Challenge
- Include disputed interpretations, misleading AI suggestions and a correction to shared material. Score acceptance of unsupported claims and whether reuse spreads a common assumption. Check how readers revise their explanations and reassess affected claims against expert-defined evidence criteria. AI-assisted agreement is not independent corroboration.
- Examine judgment
- Separate factual mistakes from defensible choices. Explore a permissioned exercise that randomly varies which option an explanation favors across conditions, outside the live graph. Assess choices and reasons in a new case. Compare AI-attribution reports with assistance records, not self-report alone. Agreement or disagreement does not by itself establish independent judgment.
- Inspect individual steps
- Test whether readers detect unsupported steps among accurate background, without falsely rejecting supported claims. Compare localized review and correction across conditions. Preserve original sources and enough document context to catch misleading omissions or interactions between claims. Distinguish this scrutiny effect from confidence or familiarity alone.
Before recruitment, agree the target, independent scoring, resources, consent and permissions. Use a separately reviewed, frozen pilot package; today's automatic acceptance rules are not scientific review. Scientific version-pinning, linked-use records, explicit dissent records and separate contribution credit remain proposed extensions.
These are proposed research outputs. Agree which materials, protocols and results can be shared under source rights, participant consent and institutional permissions before recruitment.
Four years of collaborative knowledge-building
Iman YeckehZaare's 2024 dissertation documents 1Cademy's first four years of linked topics, proposals and community review.
- students and researchers
- 1,754
- institutions represented
- 208
- knowledge nodes
- 50,947
- proposals
- 271,836
Cumulative participation reported in 2024, not current activity or verified AI outputs.
Read the dissertation (2024) (opens in a new tab)Source details
Harnessing Micro-Topics Arranged in Learning Pathways for Spaced Retrieval, Reading, and Collaborative Note-taking
Iman YeckehZaare, PhD dissertation, University of Michigan, 2024. Abstract and chapter 4, section 4.5.
1Cademy advisors
Experience in collective intelligence, organizational learning and academic publishing.
Joel Podolny
CEO and founder of Honor Education. Former vice president at Apple and founding dean of Apple University, and former dean of the Yale School of Management.
View profile for Joel Podolny (opens in a new tab)
Paul Resnick
Michael D. Cohen Collegiate Professor of Information at the University of Michigan and director of its Center for Social Media Responsibility. His research spans collective intelligence, reputation systems and social computing.
View profile for Paul Resnick (opens in a new tab)
Roby Harrington
Special advisor to the CEO of Honor Education. Former vice chairman of W. W. Norton & Company and former head of its college department, bringing experience in academic publishing and editorial leadership.
View profile for Roby Harrington (opens in a new tab)
Founder and project lead
Iman YeckehZaare
Postdoctoral Associate, MIT Center for Collective Intelligence. Head of Research, Honor Education.
Iman founded 1Cademy and leads its research, design and development with its engineering team. His 2024 University of Michigan dissertation studied microtopics arranged in learning pathways and documents the platform's first four years. This proposed scientific-learning pilot builds on that experience in learning research and collaborative authoring.
Help test shared learning for science
Help develop and evaluate one shared learning path for a public AI-assisted scientific contribution. We are seeking domain reviewers, research partners and support for the pilot.
No 1Cademy account is needed to get in touch.
Ways to take part
Research teams
Identify a bounded learning target and appropriate public sources.
Bring a scientific contribution that is difficult to explain across fields and help identify the background a reader needs.
Domain reviewers
Examine explanations, prerequisites, assumptions and alternatives.
Shape an assessment in your field and help make a difficult explanation reusable. Agree review credit and time commitments before the pilot.
Research supporters
Resource one bounded research pilot.
Support planning for expert review, participant time, coordination and compute, including the cost of keeping shared material useful afterward.
One learning target. Clear deliverables.
The proposed pilot would deliver a sourced learning path with reviewed, frozen versions, an independent assessment protocol, and a report on understanding, scrutiny, reuse and total effort.
Set the scope, budget, permissions and evaluation criteria before recruitment. Count the work of building and maintaining the path, not just learner time. This is a proposed pilot, not an approved study or funding arrangement.
On 1Cademy
- Human Preference Alignment via Reward Models
- Sign in to contribute
Accounts need an email address from an approved institution.