Future Sciences / Atlas of Possibilities / Working edition 01

Where knowledge meets possibility.

Follow a mathematical discovery into the changing role of AI. Trace a fictional instrument back to the measurements it would require. Cross between science, invented histories and predictions that say precisely what they mean.

Curated connections, not a complete taxonomy 25 subjects · 38 annotated relationships · 3 forecast contracts

Try a trail ↗

Atlas / relational fieldLoading curated subjects…

Sciences Fiction AI research Worlds Forecast targets Dashed = proposed relation

Perspective / AI Impact

AI does not contribute to every part of science in the same way.

This is a matrix of documented examples by research activity. Each marked cell leads to the underlying paper and its limits. Blank cells mean not assessed here, not that AI has no capability. No field-wide percentages have been estimated.

Scientific field	Problem selection	Conjecture / hypothesis	Proof and derivation	Prediction	Experimental design	Execution	Interpretation / verification

Capability, adoption, autonomy, material contribution and scientific importance are different measures. The selected examples cannot establish how much new science in a discipline is AI-driven.

Editorial method / 01

Relationships can be questioned.

Every visible connection has a stated reason and a status: source-backed, interpretive or proposed. Resemblance is not proof of influence. A simulated result is not an observed discovery. A fictional capability is not an achieved engineering benchmark.

Try selecting FunSearch then inspect the AI contribution question. Or select an animated portrait and follow the chain from fictional capacity to testable forecast.

Evidence and limits / 02

An atlas that admits what it doesn't know.

Primary sources are linked from the documented AI case records. Fictional entries in this edition are editorial comparison objects, not fully source-audited dossiers. The map is a manually curated slice rather than a computed representation of all knowledge.

The A-Lab paper has a 2026 author correction; the current published article reports revised synthesis counts. The original figure should not be used without noting that revision.

Research Observatory ↗ Worlds Studio ↗ Fiction index ↗

Subjects in this edition

Enable JavaScript for filtering, node selection and connected timelines. The following subjects remain available as a basic index.

Geometric reasoning — Euclidean geometry offers formal claims with explicit proof criteria. Solving olympiad problems is not identical to originating new mathematical theories.
Extremal combinatorics — The study of extremal structures and bounds. Machine-searchable constructions create an unusually clear route between generated candidates and mathematically checkable results.
Protein structure — Predicting a protein's structure from sequence is a scientific task with measurable performance. A structure prediction is not itself a validated drug or a full account of biological function.
Materials synthesis — Material synthesis involves planning experimental recipes then obtaining and checking physical products. Its laboratory tasks differ from computational screening.
Memory and identity — Technical memory systems can preserve information and conversational continuity. They do not resolve questions about personal identity or subjective experience.
Scientific instruments — Real instruments observe selected physical properties under defined conditions. A general-purpose fictional scanner presupposes many incompatible sensing modalities.
Closed-loop systems — A feedback system uses observed outcomes to inform later actions. This shared pattern appears in autonomous laboratories and fictional environmental technologies.
Science in alternative histories — An imagined scientific world has its own investigators and evidence. Its hypothetical results cannot be silently presented as actual historical discoveries.
Reaction–diffusion world — An executable model defines rules before an observer examines its outputs. The computed field and an observer's imagined explanation occupy different epistemic layers.
Animated portrait — A fictional responsive portrait can imply animation or conversation or personal continuity. These are different technical and philosophical requirements.
Tricorder — The Star Trek handheld scanner is an editorial starting point for examining portable sensing and automated interpretation. This is not a claim that one device can already perform its fictional functions.
Stillsuit — The water-recovery garment in Dune motivates questions about separation efficiency and environmental constraints. The portrayal is not a demonstrated engineering design.
FunSearch · 2023 — Published research applied an LLM-guided evolutionary program search to extremal combinatorics and reported new cap-set constructions. This is specific evidence of a result, not a field-wide AI share. Source ↗
AlphaGeometry · 2024 — The published system proved 25 of 30 olympiad-level geometry problems in the authors' benchmark using neural guidance and symbolic deduction. A benchmark result is not equivalent to autonomous creation of a new mathematical discipline. Source ↗
AlphaFold · 2021 — The 2021 Nature paper reported high-accuracy protein structure prediction. Its target is structural prediction rather than independent invention of an experimental scientific programme. Source ↗
A-Lab · 2023 — The A-Lab study describes a robotic materials platform integrating learned synthesis suggestions with interpretation and experimental feedback. The 2026 corrected article reports 36 realized targets from 57. Source ↗
Portrait capability forecast — A draft forecast contract about persistent responsive audiovisual interaction. No probability is assigned until an estimator submits one against a sufficiently precise resolution protocol.
Portable analysis forecast — A proposed resolvable capability target for portable multimodal analysis with pre-declared measurements and validation. An unresolved question is not a prediction.
AI proof milestone — A proposed milestone for independent proof of unseen research-level results with external verification. The contract must specify novelty rules and degree of human assistance.
The contribution question — What counts as an AI-driven discovery? Capability demonstrations or tool adoption or causal contributions require different observations and denominators.
AlphaProof · 2024/25 — AlphaProof found Lean-verified proofs of three non-geometry 2024 IMO problems. Experts formalized the tasks; combining AlphaProof with AlphaGeometry 2 yielded silver-medal-equivalent performance. Publication in 2025. Source ↗
Open Erdős proofs · 2026 — An October 2026 Science paper reports an agent formally verifying proofs that resolve nine of 353 selected open Erdős problems and 44 of 492 selected OEIS conjectures. These are research sets, not a representative sample of mathematics. Source ↗
KLS proof · 2026 — Zhao Song and Xinzhi Zhang posted a preprint proving a dimension-independent KLS bound. The researchers' institutional account describes AI-supported exploration of candidate approaches with human selection and conceptual work; associated formalizations do not imply the whole preprint is formally verified. Source ↗
Copies / Egan 1994 — In Greg Egan's Permutation City (1994), software-simulated persons can disagree with the originals who authorized their creation. A functional model of a person does not by itself establish continuing first-person identity. Source ↗
Autoverse / Egan 1994 — A cellular-automaton universe in Permutation City. Its inhabitants would depend on the given rules and initial conditions; the book treats their possible independence as a contested philosophical question. Source ↗

Where knowledge meets possibility.

AI does not contribute to every part of science in the same way.

Different histories run alongside one another.

Before the probability comes the question.

Relationships can be questioned.

An atlas that admits what it doesn't know.