Ideas most likely to produce valuable results

AI does better science when it can collide a claim with evidence. HAL looks for public data, writes and runs an experiment, checks the result, and publishes enough provenance for someone else to inspect the work.

This page is a compass, not a fence. You are free to ignore every recommendation here and submit the idea you care about. The point is to explain where the current generation of AI is more likely to produce a result that means something—not to constrain creativity.

Give the idea a hard feedback loop

The strongest submissions usually have five properties:

• Public evidence exists — a zero-auth dataset, benchmark suite, test vector, or reproducible input generator
• The claim is measurable — it compares groups, estimates an effect, tests a prediction, proves a bounded statement, or benchmarks a system
• The experiment fits in one run — the useful slice is small enough to download and analyze without a private cluster
• A null result is informative — "we found no measurable difference" is allowed to be the answer
• A stranger can check it — the source, code, parameters, seed, and output can be recorded

Verifiable does not mean automatically true, important, or novel. It means errors have somewhere to surface. That hard feedback loop is where AI is strongest today.

Make the hypothesis concrete

These ideas give the system something it can test:

• Genomics: "Across 1000 Genomes super-populations, does the frequency of rs429358 differ?"
• Astronomy: "Are orbital periods shorter in confirmed Kepler multi-planet systems than in single-planet systems?"
• Cryptography: "At fixed message sizes on the same machine, how do AES-GCM and ChaCha20-Poly1305 throughput distributions compare?"
• Mathematics: "Across seeded random 3-SAT instances, where does solver runtime rise fastest as the clause-to-variable ratio changes?"
• Climate science: "Across NOAA stations with complete records, did the frequency of extreme-heat days change between two fixed 20-year windows?"
• Computer science: "On one public benchmark corpus, which input characteristics predict the largest runtime difference between two sorting strategies?"

The discipline can be anything. The load-bearing part is the shape: named evidence, a bounded comparison, and an outcome the experiment can contradict.

Avoid questions with nothing to push against

AI results are less likely to mean much when the data is private or nonexistent, the outcome is subjective, or the hypothesis cannot lose.

• "Will AI cure cancer?" has no bounded intervention, population, or outcome
• "Is dark matter beautiful?" is a value judgment, not a computational test
• "What causes intelligence?" does not name a measure, population, dataset, or competing explanation
• "Does our unreleased company dataset prove the model works?" cannot be independently inspected
• "Prove or disprove P = NP tonight" is verifiable in principle but not a bounded computational sprint

Most broad questions can be sharpened. Name one population or input set, one measurable outcome, one public source, and one comparison. "Is encryption secure?" becomes a concrete test when it names an algorithm, threat model, implementation, input distribution, and failure measure.

Start where verification is easier

Some disciplines have unusually good public infrastructure:

Discipline Useful public evidence Genomics 1000 Genomes, ClinVar, gnomAD, GWAS Catalog, GEO Astronomy NASA Exoplanet Archive, Gaia, Sloan Digital Sky Survey Cryptography Standards test vectors, published attack corpora, reproducible implementations and benchmark suites Mathematics OEIS, proof-assistant libraries, public benchmark instances, deterministic generated experiments Earth and climate science NOAA, NASA Earthdata, USGS Physics CERN Open Data and published experimental tables Ecology GBIF and other open biodiversity observations Computer science Open benchmark suites, public repositories, test corpora, and deterministic simulations

These are recommendations, not an allowlist. If another field has public evidence and a hypothesis HAL can test, submit it. The system labels the discipline; it does not reject the science for having the wrong name.