Map the argument.
Explore source-linked graphs, notation, and reading guides. 5 sample papers have experimental maps; the open-source skill also works with your own mathematical papers.
Super-intelligence is more
than a correct answer.
Its value also lies in what people can understand, question, learn, and build from it. Our ambition is super-intelligence for human intelligence. We’re starting with mathematics.
Adapt the explanation.
Set your preferred style, confirm what you know, and choose scoped notation aliases. Keep the complete original statement alongside an adapted explanation.
Ask about a step.
Anchor a conversation to a passage and follow its source links. AI explanations and style suggestions require a configured provider; maps, preferences, notes, and context export work without one.
Mathematics is our first community.
We’re beginning with mathematicians reading AI-generated papers: following arguments, questioning difficult steps, and discussing which ideas are worth pursuing. Their needs shape this first reading room.
Over time, we want to welcome anyone who wants to understand and engage with AI’s work, across subjects and forms of expression. The purpose stays the same: help people develop their own understanding and judgment as AI becomes more capable.
What we measure—and what we don’t.
Reading activity records visible text regions during active, foreground sessions, only when you enable it. An engaged read means at least 60 seconds of observed activity. This convention is not evidence of comprehension.
“I understand” is a reader’s self-report. Reaching the final page is not completing a proof. An explicit stopping reason is distinct from closing a tab. Feedback on value, writing, and method familiarity remains separate.
Faithful papers, clearly identified feedback.
The reader renders the original PDFs. Highlight text or select an area around an equation. These selections identify places to discuss; they are not a certified decomposition of a proof. The sidebar’s text excerpts are extracted approximations; consult the rendered PDF for exact notation.
Every note is attached to a manuscript version and a selection or block. The repository revision and PDF checksums are retained. These are AI-generated manuscripts at different stages of verification; inclusion in this collection is not an endorsement of their claims.
Joining the first reading room.
The mathematics pilot starts with academic email sign-in. Verification will confirm access to an institutional mailbox, not mathematical expertise. Institutions outside the United States can use their own email domains. An institution review process can extend the pilot list.
In the local demo, contributions and reading progress stay in this browser. In a configured community deployment, comments and assessments are public. Your saved marks are private, while their aggregate counts contribute to paper signals. Legacy public reactions remain public. Following lists are private. Community leaderboards count public comments, not expertise or mathematical quality. Stopping reasons and website feedback are visible to you and the service administrators. Reading history stays on your device and can be cleared; local data can be exported from My library.
Sources and independence.
Readable Intelligence is an independent prototype. It is not an OpenAI product or an official review of the collection. Paper content comes from the OpenAI mathematics repository, with its Apache-2.0 license and provenance preserved.
Find a paper to read