“A wealth of information creates a poverty of attention.”

— Herbert Simon, 1971

Jev from TypeSafe AI landed about 72 hours ago. It is supremely interesting not because AI has learned another trick, but because it challenges one of the assumptions underlying the current AI era: that intelligence begins with a question. The architecting of tradeInsightAI’s AMIE (“friend” in French), the Agentic Market Inference Engine that powers all inferences downstream also challenges that assumption & more on that later.

For all the extraordinary progress in generative artificial intelligence, most of our interaction with these systems still follows a surprisingly old pattern: we ask, and the machine answers. Search engines work this way, chatbots work this way, and even many sophisticated AI agents begin with a human supplying an objective.

That model has become so familiar that we rarely question the premise underneath it. Why should the machine wait for us to notice something before it becomes intelligent about it?

TypeSafe AI recently introduced Jev, a model designed less as a conversational head than as a judgment engine. Given a state and a set of questions, it returns constrained judgments and probabilities rather than paragraphs of generated prose. On the surface this may seem like a relatively modest change in model design, but it points toward a larger possibility: the next evolution of AI may be less about improving the answer and more about recognizing when there is something worth asking about in the first place.

The Prompt as a Transitional Interface

The prompt has been enormously important because it made powerful computation accessible through ordinary language. But it also places a subtle burden on the human. To ask a useful question, we generally need to recognize that something deserves questioning.

Consider a financial market. Before asking why a stock is behaving unusually, someone first has to notice the unusual behavior. The same problem exists in cybersecurity, industrial systems, medicine, logistics, and countless other dynamic environments. An anomaly or meaningful combination of events can exist long before anyone thinks to investigate it.

This suggests a distinction between query intelligence and situational intelligence. Query intelligence is exceptionally good at responding once we identify what we want to know. Situational intelligence has the harder task of continuously evaluating what is happening and determining which developments deserve attention. Done objectively, it is a first-principles measurement and implementation.

The latter is particularly interesting because it moves part of the burden of discovery from the human to the machine. Instead of requiring someone to recognize the question first, the system becomes responsible for recognizing that the underlying state has changed enough to warrant one.

Intelligence Does Not Have to Speak

Generative AI has made language so central to our conception of artificial intelligence that it is easy to confuse the ability to articulate a conclusion with the intelligence required to reach it. Yet many important cognitive operations do not inherently require language.

A system may need to determine whether a piece of evidence is relevant, whether two observations conflict, whether the current explanation has weakened, or whether there is enough information to continue. These are judgments, and generating several sentences of prose may add very little to them.

This is exactly the layer Jev is explicitly designed to occupy. TypeSafe describes it as a System One model, intended to make fast, structured judgments rather than perform open-ended generation. The broader architectural idea is compelling because we do not need the same computational machinery for every cognitive operation. Deep reasoning, retrieval, calculation, prediction, classification, and explanation are different problems, even though today’s general-purpose models can attempt all of them.

If bounded judgments can be made reliably by smaller and faster models, then more expensive intelligence can be reserved for situations that actually require it. That may lead to AI systems that are less monolithic and more compositional, with different forms of intelligence performing the jobs for which they are best suited.

An Old Idea Arrives at a New Moment

There is some historical irony in all of this. Before AI became synonymous with generating language, much of machine learning was explicitly concerned with making decisions from observations.

In 1950, Glenn Brier introduced a formal method for evaluating probabilistic forecasts, establishing an idea that remains fundamental today: a probability should have an empirical meaning. If a forecaster repeatedly assigns roughly 80% probability to comparable events, those events should occur roughly 80% of the time. Calibration!

Around 1970, C. K. Chow formalized another surprisingly modern idea. A machine should sometimes decline to classify something when the evidence is insufficient. What we now describe as abstention, escalation, selective prediction, or human review has deep roots. The objective was never simply to force a prediction from every input; it was to make the overall decision system more reliable.

The quote by Herbert Simon above identifies the economic problem underneath the entire information age and that insight is remarkably current. We have built extraordinary machines for producing, retrieving, and now synthesizing information, but the human capacity to absorb it has not increased accordingly. The problem is increasingly not whether information can be found, but whether a system can determine which information deserves someone’s attention. And finally, what set of correctly vectored inferences can one draw from that information.

Seen from this perspective, Jev is interesting not because classification or calibrated probabilities are new. They are not. What is interesting is the possibility of making flexible semantic judgment inexpensive enough to distribute throughout a computational system.

Markets Don’t Wait for Questions

Financial markets make the distinction between answering questions and understanding situations particularly clear because a market is continuously changing whether anyone is watching it or not. Orders enter and disappear, trades execute, liquidity is consumed and replenished, participation changes, volatility expands and contracts, information arrives, and price responds.

Sometimes the most informative event is that price doesn’t respond!

Suppose a stock receives favorable news. Buyers become increasingly aggressive, volume expands, and price moves higher. If we begin with the headline, an obvious interpretation is that the positive catalyst has produced strong demand and higher prices.

Now suppose something changes. Successive waves of aggressive buying begin producing less upward progress. Offers repeatedly replenish near the high, and despite increasing activity, price cannot sustain itself above the same area. The important observation is no longer simply that buyers are aggressive. It is that their aggression is becoming less effective at moving the market.

The headline has not changed. The buying may not have disappeared. What changed is the relationship between effort and market response.

That distinction is difficult to discover through a system that waits passively for someone to ask the right question. Ideally, the analytical system should recognize the changing condition first and bring it to the user’s attention.

Start With the Market, Not the Narrative

This is where the broader idea intersects naturally with our work, powered by AMIE and expressed by way of Shannon, our real-time interactive layer for humans.

Our approach to financial intelligence begins as close as possible to the mechanisms producing market behavior. Rather than asking a language model to construct a narrative from a chart, we want the analytical process to begin with observable market state: how liquidity is changing, where trading pressure is originating, whether liquidity replenishes after being consumed, how much aggression is required to move price, whether participation is broadening or concentrating, and how those relationships are evolving through time. Then there are other related sources where information has to be normalized, categorized and ranked for importance and relevance.

Only after all that analytical work should the language layer explain what the evidence collectively suggests.

We think of this as first-principles market intelligence. The distinction matters because the market itself should constrain the narrative, rather than the narrative determining which market evidence receives attention.

It also changes the role of AI. Instead of functioning primarily as a more convenient search interface, the system can continuously analyze the environment and identify conditions that deserve investigation before the user knows to search for them.

Search Moves Inside the System

This does not mean search disappears entirely. It means a part of the discovery process moves behind the consumption interface.

Traditional information systems put much of the analytical burden on the user. Something first attracts our attention; we search for information, compare sources, analyze what we find, and eventually form a view. A continuously intelligent system can perform more of those early steps itself by observing the environment, detecting meaningful changes, assembling relevant evidence, and deciding when the result deserves attention. That is empowering.

The human interaction then begins at a much more interesting point. Instead of asking what is happening, we can ask why the system believes it matters, what evidence supports the interpretation, what contradicts it, whether similar conditions have occurred before, and what future evidence would cause the interpretation to change.

In that sense, search has not been eliminated. It has been absorbed into a larger process that can provide situational awareness in a manner that is more efficient.

That is an important distinction because a system that merely produces more alerts does not solve the problem. If everything is important, nothing is important. The system has to exercise judgment not only about what it observes, but also about what deserves to consume human attention.

Judgment Is Not Truth

There is also a danger in taking the idea of automated judgment too far. A probability displayed beside an AI conclusion does not transform that conclusion into fact, and financial markets make this distinction especially important.

Confidence that a news article concerns a company is one type of judgment. Confidence that the market is exhibiting a particular microstructural condition is another. Estimating the probability of a future event is something different again.

These quantities should not be collapsed into a single number called “confidence.”

This is why the historical ideas remain relevant. Brier reminds us that probabilities must eventually be tested against outcomes. Chow reminds us that uncertainty can justify withholding judgment rather than forcing one. Simon reminds us that the purpose of an intelligent information system is not to maximize the amount of information it produces, but to use human attention intelligently.

Those principles apply regardless of whether the underlying technology is Jev, a large language model, a statistical model, the tradeInsightAI stack, or something else yet to be invented.

What Comes After the Search Box?

For more than two decades, the search box has been the dominant metaphor for accessing digital knowledge. Generative AI transformed that interface. We can now ask questions naturally instead of constructing keyword searches, and increasingly receive synthesized answers rather than lists of links.

That is a major advance, but it may also be an intermediate stage.

In dynamic environments such as financial markets, cybersecurity, medicine, industrial systems, and logistics, the more consequential transition may be from information retrieval to continuous awareness. The system does not wait idly for a query. It continuously evaluates its environment, compares current conditions with what came before, tests competing interpretations, and decides whether anything has changed enough to deserve attention.

Jev is interesting to me because it makes one piece of that future more explicit. It suggests that AI systems may increasingly be composed of different kinds of intelligence: inexpensive judgment, deterministic computation, specialized analytical models, retrieval, deeper reasoning, and generative explanation working together rather than asking one enormous model to perform every task.

The chatbot may remain the visible interface, but increasingly, the important intelligence will have happened before the conversation begins.

That may ultimately be the more important transition. The first generation of modern AI became remarkably good at answering our questions. The next generation may become much better at understanding the environment well enough to recognize when there is something we need to know before we think to ask.

We are moving toward a world with “Machines of Inferential Grace” and more on that vision and characteristic in the near future.