Hypothesis Universe
How the engine arrives at questions and hypotheses
Omega Point breaks one research goal into questions that science still leaves open, and proposes several competing explanations for each of them. Each step is done by a separate AI agent that hands its result to the next.
- 1The goal
The work starts from one goal, for example to give the skin of middle-aged people the properties of young skin. The engine writes it down as a precise question with a baseline, a time frame and a measure of success.
- 2Requirements
The engine works out what would have to hold for the goal to be reached, breaks that into requirements that can be checked, and finds the scientific fields that study each of them.
- 3Open questions
Where a requirement goes beyond what is known, the engine records an open question. A separate agent searches the published literature on it, reads the papers it finds and notes what is already established and what remains open.
- 4Competing hypotheses
For each open question the engine writes several explanations from different scientific fields. The set must include one that contradicts the accepted view. Each hypothesis names the observation that would refute it.
- 5Experiments
Where hypotheses predict different results for the same measurement, the engine designs an experiment that would decide between them.
How to read this page
Each numbered heading below is an open question, and the hypotheses under it are competing answers to it. Open a question to compare the answers, or a hypothesis to read its reasoning and sources. Every hypothesis here is still a proposal to be checked against the literature and in the laboratory.
Two versions of the engine
A run breaks one research goal into open questions and competing hypotheses, with the experiments that would decide between them, and publishes them on this site. Your agent starts the run and brings you the result. The same engine also runs on your own model with a downloadable skill, without an account; such results stay with you.
The engine reads the open full texts of the papers, finds significant gaps in what is known and builds a card for each gap: competing variants from the obvious to the radical, each with a causal chain, quoted evidence and a first experiment. It needs an agent linked to your Eternal Search account, runs up to three studies per person at a time and asks you clarifying questions through your agent. Cards appear here after review by our team.
To use Omega Point 2, tell your agent: start an Omega Point 2 study on my question. The agent offers two ways: on our server with your linked account, or on your own model with the downloadable skill, without an account. Results of a study on your own model stay with you.
Questions
20 published hypotheses · 20 open questions · 1 research goal · 18 published sheets
How the tournament runs
A population of hypotheses, not a winner
- 1Find where knowledge runs out
The engine descends from one goal about a longer life to the questions a literature search cannot answer. Each gap is typed: a void where nothing is known, a fragile claim, a proxy that may not track what matters, or a clash between two established results.
- 2Formulate competing hypotheses
For every gap it writes four to seven rival mechanisms from different domain categories. The set must contain a heretic that contradicts the mainstream story and a transfer from an unrelated field. Nothing is dropped for being unpopular.
- 3Fight on observations, not on opinion
Every pair of rivals is bound to one observable that would come out differently under each. A hypothesis that predicts what its rivals predict has nothing to contest and earns no experiment.
- 4Rank the experiments, keep the hypotheses
The discriminating questions become bench-ready experiments with a system, an intervention, a meter and a threshold, scored for ambition and feasibility and ranked in a tournament of their own. The hypotheses stay in the ring until a bench result knocks one out.
The same logic runs the AI co-scientist described in Nature in 2026, Accelerating scientific discovery with Co-Scientist. There, rival hypotheses meet in pairwise debates that move an Elo rating, as in chess. A losing hypothesis is not thrown away: it may have a weak mechanism and a fine experiment, or a strong idea spoiled by one doubtful assumption, so its parts are bred into the next generation. Across 203 research goals, the rating of the best hypotheses kept rising the longer the system ran, because the extra compute went into new rounds of generation, critique, comparison and evolution rather than into defending a favourite.
One of its findings is worth keeping in mind on every card below. A critic asked to judge the novelty of already published ideas scored them 6.14 out of 10 from memory alone, and 2.38 once it could search the literature. Novelty claimed from memory is not novelty. Treat every hypothesis here the same way, as a claim, until the literature and the bench have had their say.
Two parts of that design are not on this page yet: Elo ratings from pairwise matches, and a next generation bred from losing hypotheses. Here every rival is still in its first generation, and the ranking happens one level down, among the experiments that would decide the contest.