Methodology

This map reflects my personal interests: it covers the areas of AI I follow and work in, not the whole field. It is an independently curated synthesis, not an official scientific consensus. Assessments are scoped to available evidence, may be incomplete, and are revised as new results emerge.

Current status of this map

The map is at an early stage. It holds 50 entries backed by 8 sources that the curator opened and read: one survey was read in full text, and the others at abstract level (each source record says which). The initial content was drafted with AI assistance from those sources. 0 of52 assessments have been human-reviewed. 22 assessments are explicit placeholders for areas the map has not yet examined. Nothing here is labelled “established”: that state requires review and at least two independent sources.

What the map represents

The AI Knowledge Frontier tries to show what AI systems can currently do, under what conditions, what limitations remain, and how the boundary of that knowledge changes. It is a map of knowledge about AI, not of Adam Slucki’s personal opinions, although which areas it covers follows his interests. He is its independent author and curator; the scientific claims come from the cited sources, and the interpretation of how they fit together is the curator’s and may be wrong.

Why it is not a literature map

Papers are not nodes. The units of the map are capabilities, claims (specific, scoped statements), limitations, open problems (phrased as questions) andhypotheses. Publications, benchmarks and replications are evidence attached to assessments of those units. The number of papers on a topic is never used as a signal of progress.

Assessments depend on conditions

A capability is rarely “solved” or “unsolved” in general. Whether adaptation works can depend on how large the shift is, whether labels or feedback are available, the batch size, the architecture, and how long the system runs. The map therefore stores each assessment together with its scope (the conditions under which it applies), and a single capability can have several assessments that differ by condition. Conditions are shown as qualitative tables; the map does not invent numeric boundaries.

Four separate dimensions

State of knowledge: established, demonstrated (restricted), partially solved, open, hypothesis, or insufficiently documented.
Evidence strength: strong, moderate, preliminary, insufficient, or contested.
Map confidence: how confident the curator is in the map’s own interpretation of the evidence.
Practical maturity: research, prototype, deployed, or unknown.

These are never combined into one score.

What counts as evidence

Peer-reviewed studies, preprints with reproducible experiments, independent replications, benchmark evaluations, formal proofs, public reproducible experiments, author-reported experiments, and well-documented negative results. Each record notes its source type, date, what was actually read (full text, abstract or neither), the exact finding used, the experimental setting, caveats, and whether independent validation exists. Author-reported results are not treated as independently established. Records the curator has not read cannot support an assessment.

Safeguards against self-reinforcement

The map’s own entries are never evidence; evidence and map nodes live in separate records and the validator rejects citing one as the other.
AI-generated interpretation is not verification. An AI draft cannot mark itself reviewed or establish a finding.
Secondary reports of one result count as one source, not several replications.
Contradictory findings are represented, not removed; contested evidence is labelled as such.
A claim does not become more certain because it was restated; status changes need a recorded reason.

Relationships

Links are typed (for example requires, addresses, limited by, hypothesized to enable) and each carries a rationale. Links backed by evidence are drawn solid; conceptual or speculative links are marked hypothesized and drawn dashed. “Addresses” never means “solves”. Structural similarity between two problems is not evidence that they share a solution.

Absence of evidence

“Insufficiently documented” means the map has not collected enough evidence to assess something responsibly. It does not mean the capability is impossible, nor that nobody has studied it. The map covers a small set of areas and omits most of AI.

Revisions

When an assessment changes, the old one is kept and a new one supersedes it, together with a revision entry recording what changed, the previous and new assessment, the reason, and the supporting evidence. Each entry’s detail panel lists its revisions. Dates distinguish when evidence was published from when the map examined it: a new paper does not by itself change the map.

Update process

Candidate results are turned into machine-readable proposals (extracted claims and conditions, how they relate to existing entries, and the resulting assessment changes). A proposal is validated and then accepted or rejected by a named human reviewer before publication. AI tools may draft proposals but cannot declare a problem solved. Currently updates are made by hand via data files and scripts; there is no automated scraping.

Limitations and corrections

The map is incomplete, simplifies a large and fast-moving field, and involves editorial judgement. Errors in transcription, scope or interpretation are possible, especially in entries read only at abstract level. Corrections are welcome via the contact details on the Profile page.