Coming out of stealth · October 28Out of stealth · October 28
Cognitive infrastructure
← ArticlesResearchResearch

On the Discovery of What Matters

Discovery under uncertainty: deciding which questions should be asked in the first place.


A black hole, centered, its bright ring and disk against dark streaks.

1. The Retrieval Revolution

Google became one of the most consequential companies in modern history by solving one very specific problem: retrieving the exact piece of information someone is looking for. For the first time, information became accessible at scale, not by browsing randomly, but by querying: a user could ask a question, and the system would point to the most relevant answer. That was a true game changer.

When Large Language Models were released, the world treated them like the next revolution, and it’s hard to deny that LLMs changed, once again, the way we consume information. But fundamentally, they were an extension of the same paradigm. Instead of retrieving a single document or link, they synthesize across many. The user no longer gets just one piece of information, but a structured combination of everything that appears relevant to their query. In that sense, they are definitely more powerful: they offer more context, more coverage, more fluency. But the underlying assumption remains unchanged: one still needs to know what to ask in the first place.

And that is where the system breaks down, because in most real-world situations, the hardest problems are not about retrieving known information. They are about uncovering what is missing, identifying signals that were never explicitly queried, inconsistencies that were never surfaced, questions that were never formulated. In other terms, we became extraordinarily good at answering questions, but never solved the inverse problem.

2. Search vs. Discovery

The limitation of retrieval-based systems is subtle, but fundamental. They assume that the user knows what they are looking for. And the problem, of course, is that in practice, this is rarely the case. In many situations, the problem is not to find a specific piece of information, but to determine whether something important is being overlooked, such as a risk not accounted for, a contradiction not surfaced, or a signal buried in noise. It is the classic needle-in-a-haystack problem. Except perhaps that you don’t know what the needle looks like; you don’t know how many there are; and you don’t even know if there is a needle in the first place.

Today, entire professions exist to solve this problem. Data scientists formulate hypotheses and test them against data, trying to uncover patterns that were not explicitly queried. Fraud teams investigate whether something “doesn’t check out,” often without knowing in advance what they are looking for. Risk managers attempt to identify exposures that are not yet visible. Quants search for signals in markets where most patterns are indistinguishable from noise. Paralegals sift through large volumes of documents to surface what might matter when it is not obvious.

In all of these cases, the task is obviously not retrieval: it is discovery under uncertainty. And today, this process remains fundamentally manual. We compensate for the limitations of our systems by inserting human judgment and people whose job is to continuously ask: What are we not seeing?

3. The Limits of Query-Driven Systems

Modern AI systems are built on a simple loop:

query → retrieval → response

Large Language Models generate answers to prompts. Retrieval systems fetch documents based on queries. Agentic systems orchestrate sequences of actions but still rely on an initial instruction to begin. Despite their differences, they share the same underlying assumption: the system only acts because it is asked to in the first place. This works remarkably well when the problem is well-defined or when the user knows what to look for, how to phrase it, and when to stop.

But it breaks down in the exact situations where intelligence matters most: the cases when information is incomplete, when signals are weak or conflicting or when the structure of the problem is not yet known. In these environments, the quality of the outcome is not limited by the system’s ability to generate answers: it is limited by the quality of the question. And if the right question is never asked, the system never engages with the relevant information at all. And no amount of scale, data, or model improvement changes this dynamic: a system that depends on queries cannot surface what was never queried in the first place.

4. Intelligence Is Not Just Generation

The recent wave of generative AI has led to an unexpected but undeniable shift in how we think about intelligence: we have started to equate intelligence with the ability to produce content, to generate text, code, images, or answers. And in some ways, that makes sense indeed: the ability to synthesize information and express it fluently is a powerful capability.

But that is only half of the equation. Intelligence is not just about generating possibilities, but also about selecting which ones matter. In human reasoning, these two processes are inseparable. We generate ideas, hypotheses, interpretations, and then we evaluate them, discard most of them, and focus on the few that deserve attention. This process of selection is not secondary. It is the core of intelligence. Curation is intelligence.

This becomes even more critical in environments where information is abundant and ambiguous. The challenge is no longer to produce more possibilities, but to identify which signals are truly meaningful, which inconsistencies matter, and which directions are even worth pursuing. A system that can generate endlessly but cannot discriminate between what matters and what does not will tend to amplify noise rather than reduce it. And in ambiguous settings, nothing distinguishes what matters from what doesn’t, unless there is a process that actively distinguishes between the two.

5. The Breaking Point: Ambiguity and Noise

In well-defined environments, retrieval and generation are often sufficient. If the information is structured, the objectives are clear, and the questions are known, then a query-driven system performs well. It retrieves, synthesizes, and responds with sufficient accuracy.

However, most real-world environments do not exhibit this level of structure; they are inherently ambiguous, characterized by incomplete information, weak or contradictory signals, and relevant facts dispersed across sources without a clear organizing framework, often in situations where the problem itself is not fully specified. In such conditions, the difficulty does not lie in producing answers, but in determining where to direct attention in the first place. The system is faced with a combinatorial space of possible interpretations, most of which are irrelevant. Without a mechanism to discriminate between them, it has no way to prioritize where attention needs to go.

This is where current systems reach their limit: although they can generate plausible responses to almost any prompt and retrieve and synthesize information from vast corpora, they lack a native mechanism to determine what matters in the absence of a well-formed query, leaving them unable to discriminate between what should be ignored, what should be investigated, and when something is fundamentally “off”. As a result, they tend to smooth over ambiguity rather than resolve it, producing coherence at the surface level (fluent, internally consistent outputs) while leaving deeper inconsistencies untouched, even though, in complex environments, those inconsistencies are precisely where the signal lies.

6. From Pull to Push

All of the limitations we’ve described stem from a single assumption: that intelligence is something you query. But in ambiguous environments, the central problem is not answering questions. It is deciding which questions should be asked in the first place.

This is where a different class of systems emerges. Instead of waiting for a query, these systems operate from an internal state by continuously evaluating information, identifying inconsistencies, and surfacing what requires attention. They do not rely on explicit prompts to engage: they initiate. This is the shift from pull-based systems to push-based systems: a pull-based system responds when asked while a push-based system acts when something doesn’t make sense. And that behavior simply cannot be emulated by a retrieval-first system. In a push-based system:

  • contradictions are not smoothed over but surfaced
  • gaps are not ignored but highlighted
  • questions are not supplied by a user or a prompt but triggered by the system itself

This requires a different underlying representation, not a collection of documents to retrieve from, but a structured internal view of the world that can track what is known, what remains uncertain, and where inconsistencies emerge.

7. Coherence, Dissonance and Signal Discovery

If a system is to question its own understanding, it needs a way to detect when something is wrong. This is the missing mechanism in today’s AI systems. In human reasoning, the discovery of what matters does not come from exhaustive search or explicit queries, but rather from a continuous process of evaluating coherence (the degree to which different pieces of information fit together into a consistent whole).

When something does not fit, we notice. We take note of the fact that a detail contradicts what we expected. We realize that two pieces of information cannot both be true, and can feel when an explanation feels incomplete. This experience, often described as doubt, tension, or confusion, is not a failure of reasoning. It is the signal that drives it. Dissonance is what directs attention.

That said, this process does more than signal where to look: it creates structure. When multiple interpretations coexist under uncertainty, they are not independent. They constrain each other. A piece of information that supports one interpretation may weaken another. A contradiction introduces a tension that reshapes the system: some possibilities become more prominent while others fade, and what remains is not a collection of facts, but a structured understanding shaped by the interaction of competing interpretations.

Traditional data systems do not capture this dynamic. They treat information as discrete and additive; think for example of rows in a table, documents in a corpus or tokens in a sequence. But reasoning does not operate by accumulation but by interaction. In that sense, the process is closer to interference than aggregation. Meaning emerges not from individual pieces of data, but from how they reinforce or contradict one another: uncertainty allows multiple interpretations to coexist, dissonance forces them to interact, and coherence is what remains in the end.

8. From the Prefrontal Cortex to Dyssonance

In biological systems, this process has a locus. A significant part of what we’ve described (operating under ambiguity, detecting inconsistencies, initiating investigation) is associated with the prefrontal cortex. The prefrontal cortex does not store most of our knowledge, nor does it generate raw sensory information. Its role is different: it evaluates; it maintains a working model of the situation, detects when something does not fit, and decides when to act based on whether to seek more information, revise an assumption, or redirect attention. It does not wait for a question but determines when a question needs to be asked.

This function is fundamentally different from what today’s AI systems are designed to do. It is not a matter of scaling retrieval or improving generation. It is a matter of introducing a new layer of computation that continuously evaluates coherence and initiates resolution when needed.

At Dyssonance, this is the layer we are building. We approach it as a belief-centric system that maintains an internal representation of what is known, what is uncertain, and what is inconsistent, and uses that structure to drive behavior. When coherence breaks, the system does not simply produce a plausible answer. It triggers a process. It questions, investigates, and resolves. In that sense, Dyssonance acts as a prefrontal layer for AI, not responsible for generating content, but for deciding when something requires attention, and what should be done about it.

9. Implications: A New Class of Systems

Once a system can evaluate its own state, detect dissonance, and initiate resolution, its role changes fundamentally. It is no longer limited to responding to queries but becomes capable of driving discovery independently. In practical terms, this opens up a class of use cases that are difficult (or even impossible) to address with query-driven systems. A system can continuously monitor information and surface risks or inconsistencies as they emerge, without being explicitly asked to do so. It can identify gaps in reasoning, missing evidence, or conflicting signals across sources, and bring them to attention before they lead to failure. It can guide analysis by indicating where further investigation is needed, rather than waiting for a user to formulate the right question. It can track evolving situations in documents, datasets, or environments and highlight what has changed in a way that affects understanding.

More broadly, this shifts the system from a passive tool to an active participant in reasoning, capable of continuously evaluating information, identifying inconsistencies, and directing attention toward what requires resolution. Rather than waiting for a user to formulate the right question, the system itself surfaces what does not fit, what remains unclear, and what is missing. In environments where information is abundant but structure is not, this is the difference between navigating the surface and uncovering the underlying signal, and therefore represents a fundamentally different mode of interaction with information.

Conclusion: From Answers to Questions

Over the past two decades, we have built increasingly powerful systems for retrieving and generating information, optimizing how questions are answered without fundamentally changing how they are discovered. The limits we are now encountering are therefore not a matter of scale or capability in isolation, but a consequence of a mismatch between the systems we have built and the environments in which we expect them to operate. While query-driven systems perform well in structured settings, they break down in ambiguous ones, where the central challenge is not answering questions, but determining which questions actually matter.

The next generation of AI systems will be defined by their ability to operate in that space, not by producing more content, but by identifying what deserves attention and initiating the process of understanding it. Rather than waiting to be asked, these systems will determine when a question needs to be raised, and, in many cases, will act before the question is ever explicitly formulated.

Follow the research