Imagine a company preparing to enter a new market. Months before an announcement, it recruits a regulatory specialist, changes the language in a patent filing, and signs an agreement with a distributor. Each development has several plausible explanations. Together, they may suggest a strategy that nobody has yet described in public.
A competitor deciding where to invest cannot necessarily wait for the announcement. By then, the best locations may be taken, the most useful partners committed, and the opportunity more expensive. The decision has to be made while the evidence is incomplete.
This is an ordinary condition of consequential decisions. Businesses allocate capital before demand is established. Governments prepare for threats whose likelihood is disputed. Researchers pursue explanations they cannot yet prove. Much of human progress depends on the ability to form useful judgments before certainty is available.
AI needs to become much better at this kind of reasoning. Its development should be judged partly by whether it helps us understand what is emerging, while there is still time to respond. This becomes more important as AI moves from answering questions to acting on our behalf. The more responsibility we delegate, the more important it becomes that these systems can represent uncertainty, anticipate change, and revise their beliefs when reality disagrees.
Large language models give us an extraordinary starting point. They make recorded knowledge accessible, connect ideas across disciplines, and contribute to systems that can investigate questions and produce useful forecasts. The fact that a model learned from historical data does not prevent it from drawing novel inferences. Every human expert also learned from a past that differs from the future.
The harder question is what makes an inference worth acting on. A compelling explanation can accommodate the evidence we already have while making the wrong prediction about what comes next. A system can produce a plausible account of a company’s strategy without having identified the causes that will determine its behavior.
The explanation a model supplies is not sufficient on its own. In controlled experiments, Anthropic researchers found that models could follow hints that influenced their answers without acknowledging those hints in their reasoning. The experiments do not settle whether language models reason. They establish that a readable explanation cannot automatically be treated as a faithful account of how an answer was reached.
Return to the company preparing to expand. An AI might connect its hiring, patent activity, and distribution agreement and conclude that a launch is imminent. But perhaps the hire replaced someone who left, the patent protects an existing product, and the agreement serves a market the company already occupies. A persuasive narrative is easy to construct in either direction.
Useful reasoning would ask what evidence distinguishes those explanations. Would a regulatory application be necessary for a launch? Would the distributor need capabilities it does not currently possess? Is the hiring unusual relative to the company’s normal behavior? These questions turn an interpretation into an investigation.
This is a form of abductive reasoning: developing candidate explanations for incomplete observations. Its value depends on what happens next. An explanation earns confidence when it survives attempts to distinguish it from credible alternatives. The system should be able to specify what it expects to observe, seek relevant evidence, and reduce its confidence when that evidence fails to appear.
Even missing evidence needs interpretation. A regulatory filing’s absence means little if the relevant database is updated irregularly. Ten articles repeating one anonymous source are still one underlying observation. Any system studying the world must also understand something about how the world becomes visible to it.
This creates a substantial research problem. More information can reinforce an error when the sources share the same origin or when the system keeps looking in places that confirm its first hypothesis. An autonomous investigator needs a way to choose observations for their ability to resolve uncertainty. It also needs to recognize when further investigation is unlikely to change the decision.
The investigation becomes harder when people respond to one another. A rival may abandon an expansion after an incumbent cuts prices. A government subsidy may attract enough new capacity to undermine the economics that justified it. A warning may prompt action that prevents the event being forecast.
For that reason, a useful model of the world must represent how outcomes depend on behavior. Who can act? What do they want? What constrains them? What would cause them to change course? A collection of facts about companies or governments becomes strategically useful when it supports judgments about how those actors might respond under different conditions.
Simulation can help explore those possibilities, provided we remain clear about what it establishes. A thousand simulated executives agreeing that a market will grow does not constitute a thousand independent observations. They may share the same assumptions and the same error. Simulations earn credibility through their relationship to observed behavior, and their conclusions should change when the assumptions supporting them change.
These requirements point toward a persistent model of the world rather than a sequence of isolated research tasks. A useful system would maintain an evolving representation of entities, relationships, behavior, evidence, and uncertainty, updating that understanding as new observations arrive and as earlier predictions resolve. The purpose is not to eliminate uncertainty, but to preserve it explicitly and learn from reality over time.
There is already progress to build on. Bridgewater AIA Labs researchers reported that a system combining language models, information search, forecast reconciliation, and calibration matched human superforecasters on ForecastBench. Their system underperformed market consensus on a separate benchmark drawn from liquid prediction markets. That mixed result is informative: capabilities need to be established for particular settings, with comparisons that reveal where they work.
But answering a forecasting question leaves another difficult problem unresolved: deciding which questions deserve attention.
A company might ask whether a competitor will launch a product in the next year. The more consequential development could be a supplier quietly gaining the ability to serve that entire market. A good answer to the assigned question would leave the larger change unexamined.
Most AI systems begin with a prompt, question, or task. But some of the highest-value intelligence exists before anyone has formulated the right question. A system that continuously maintains an understanding of a changing world could identify when assumptions break, unusual relationships appear, or several weak signals begin converging toward a consequential outcome.
This is why I believe autonomous intelligence must include the capacity to discover problems. It should continuously study the external world, connect changes across domains, and identify developments that could alter someone’s decisions. Doing that well requires an understanding of both the evolving situation and the interests of the person or institution affected.
Those interests will never be perfectly inferable. Public information may reveal a company’s markets, dependencies, and commitments while leaving its internal priorities unknown. A responsible system should preserve that uncertainty and seek clarification when it matters. Useful autonomy includes knowing when another observation, or a conversation with a person, is necessary.
At Eqlipse, this is the research problem we are pursuing: how to build systems that continuously accumulate an understanding of a changing world, reason about what may happen next, and autonomously identify developments that could change a decision. That requires connecting observation, persistent world modeling, forecasting, simulation, and evaluation across time, and making each component answerable to reality. We call this Autonomous Intelligence.
The ambition requires a demanding account of what would count as success.
Forecasts should be recorded before their outcomes are known, with clear deadlines and criteria for resolution. They should be compared with sensible baselines, including how often an event normally occurs. A record of results must include failures and false alarms. When a system assigns a 70 percent probability to many comparable events, roughly 70 percent should occur; selecting a few memorable successes tells us little about whether its confidence is justified.
We also need to measure whether the intelligence arrived early enough to be useful and whether it deserved the attention it consumed. An accurate warning delivered after the last practical opportunity to respond may have little value. A system that issues hundreds of plausible warnings can impose a considerable cost even when several prove correct.
The opportunity is to make informed anticipation more widely available. A smaller business could examine a dependency it lacks the staff to monitor. A city could recognize a developing infrastructure problem while it still has affordable options. A research team could investigate a neglected explanation before committing years to its current approach.
As AI systems assume greater responsibility, being able to act will not be enough. They will also need a grounded understanding of the world they are acting in, an explicit treatment of uncertainty, and a way to learn when reality proves them wrong.
Consider the company entering a new market once more. Before its intentions become obvious, there is a period when the evidence is scattered, explanations compete, and decisions can still change the outcome. That is where intelligence has some of its greatest value. We should build AI capable of working there.
— Iman