Knowing What to Ignore

A cross-border question lands on my desk. An Israeli company, a US subsidiary, a payment that could be a dividend, a royalty, or a management fee depending on how three bodies of law read three sets of facts. I have, within reach, everything anyone could want. The Israeli Income Tax Ordinance. The Internal Revenue Code. The treaty between the two. Decades of rulings, circulars, and commentary in both systems, most of it a search away, some of it free, the rest a subscription I already pay for. Access is total.

And access is not the problem. It never was. The problem is that of the thousand authorities I can reach, perhaps four decide this question, and the work, the actual expertise, is knowing which four. A junior lawyer with the same subscriptions reaches a different answer than a senior one not because the senior has more information, but because the senior knows what to ignore. The library is identical. The judgment is not.

I think this is the least understood fact about knowledge work in the age of AI, and the assumption the whole industry is currently building against.

The belief the industry runs on

The prevailing conviction in AI is that capability scales with information. More data, larger models, longer context windows, bigger retrieval. It is not a foolish belief. It is grounded in some of the most robust findings the field has produced. The scaling laws described by Kaplan and colleagues in 2020, and refined by the Chinchilla work in 2022, show test loss falling predictably as a power law in parameters, data, and compute. The Chinchilla result specifically found that earlier large models had been undertrained on data, that more tokens would have made them better. Richard Sutton’s “Bitter Lesson” makes the case at the level of philosophy: the methods that win, over and over in the history of the field, are the general ones that leverage more computation and more data, and human-encoded cleverness tends to become a ceiling rather than a floor.

Take that whole case seriously, because it is true where it applies. At the training frontier, where a model is learning the shape of a language and a world, more real data genuinely does produce more capability. I am not going to pretend otherwise, and the argument I want to make does not require me to.

The argument is about a different stage, and a different kind of “more.”

Where access stopped being the bottleneck

Across most of the fields where professional judgment is sold, the authoritative information is no longer scarce. In law and tax, statutes, case law, treaties, and rulings are public or purchasable; a cross-industry look at the question finds the practitioner’s own difficulty framed not as access but as volume and currency, knowing which of an overwhelming corpus is relevant and whether it is still good law. Two firms with identical access to the same authorities routinely reach different positions, and the divergence is interpretive, not informational.

Finance is the cleanest case of all. Market data, filings, and increasingly even “alternative data” have been progressively commoditized; Bloomberg has spent years moving alt-data from an edge to a terminal function available to everyone. The semi-strong efficient-market hypothesis says the publicly available information is already in the price, so the edge has to come from interpretation. A 2020 Bloomberg piece captured practitioners watching alternative data become, in their words, commoditized statistics: by the time a source is widely discussed, the advantage is already crowded out. The canonical cautionary tale is Long-Term Capital Management, which had two Nobel laureates, the best models of its era, and access no one could fault, and lost roughly $4.6 billion in under four months in 1998. The failure was one of judgment under tail risk, not of information.

The same trajectory has overtaken management consulting, where the frameworks that were once proprietary are now taught in every MBA program, and transfer pricing, where both sides of a dispute typically share the same comparables database and argue over the judgment applied to it. The pattern is consistent enough to state as a rule: as a field’s information matures, the constraint migrates from getting it to knowing what to do with it.

The evidence that more actively hurts

Here is the part that should unsettle the “more is better” assumption, because it goes further than anything so far. It is one thing to say that extra information stops adding value once a field matures. It is another to say that extra information subtracts value, and that is the claim the evidence now supports, from inside AI’s own literature, tested on AI’s own systems.

More relevant-looking information does not merely add less value as you pile it on. Past a point, it makes the answer worse. The 2024 study “Lost in the Middle,” published in the Transactions of the Association for Computational Linguistics, showed that language models given long contexts use the beginning and end well and the middle poorly, so that adding material can bury the very fact that answers the question. The NoLiMa evaluation presented at ICML in 2025 was starker: on a task requiring genuine reasoning across a long context rather than literal keyword matching, a leading model that scored 99.3 percent at short context fell to 69.7 percent at thirty-two thousand tokens. A separate 2025 result found measurable degradation from context length alone, even when the retrieval was perfect, even when the needed fact was definitely present. Work on retrieval-augmented systems shows the same thing from another angle: feed a model noisy retrieved passages alongside the right one and accuracy falls.

I want to be careful here, because it is easy to overclaim this, and overclaiming it would be exactly the kind of fluency this essay is arguing against. These are findings about information at inference time, the material you hand the model when you ask the question. They are not a refutation of the scaling laws, which are about training. The honest synthesis is that both are true at once, because they describe different stages: curated scale on real data helps when you are building the model, and naive accumulation hurts when you are asking it something. The mistake is to take the training-stage lesson, more is better, and apply it to the question-answering stage, where more of the wrong thing is not neutral. It is noise, and noise has a cost that shows up in the answer.

There is a sharper way to see the same gap, and it comes from legal work specifically. Harvey’s Legal Agent Benchmark asks a model to complete real legal tasks, and an independent evaluation of it by Vals AI in mid-2026 separated two numbers that are usually collapsed into one. The first is the criterion pass rate: how often a model satisfies an individual requirement of the task. The second is task resolution: how often it gets the whole task right, which the benchmark credits only when every criterion passes. The leading models cleared roughly ninety percent of the individual criteria. On full task resolution the best of them sat around twenty percent, and most of the frontier models well below that. The models are not ignorant; they satisfy nine criteria in ten. What they cannot yet do reliably is bring all of it together at once, which is the difference between knowing a great deal and producing work someone can sign. That gap, between the criterion and the whole, is where professional judgment lives, and it is exactly the gap that more information does not close.

Which returns the whole matter to judgment. If irrelevant information actively degrades the output, then the ability to decide what not to include is not housekeeping. It is the core of the capability.

Where access is still the bottleneck

If that is the argument, then the honest thing is to say where it stops, because it does stop, and pretending otherwise would undercut everything else. The claim that judgment beats access is not universal. It holds where the information is there for anyone to reach, and it fails where the information itself is the scarce thing.

There are fields where access genuinely is still the binding constraint, and they are exactly the fields where the best knowledge is private and expensive to generate. In drug discovery, the public databases are, in the words of one review, minuscule compared to the proprietary datasets siloed inside pharmaceutical companies. The failed trials, the abandoned lead-optimization series, the unpublished mechanistic insight, none of it is a search away, and the durable advantage there is the data itself, not the algorithm running on it. Frontier scientific research has the same shape. So does clinical medicine in part: publication bias means the negative results, the trials that did not work, are disproportionately never published, so that for some questions no amount of judgment recovers evidence that was never released. And at the training frontier of AI itself, access to large quantities of high-quality real data is a genuine and coveted constraint; the finding that models degrade when trained on their own synthetic output only underlines how valuable the scarce organic data is.

The distinguishing variable is the marginal cost and exclusivity of the best information. Where it is cheap and shared, the edge is in use. Where it is expensive to generate and can be kept exclusive, access itself is still the moat. Cross-border tax sits firmly in the first category, which is why the argument holds for the work I do, and I would rather say plainly where it does not than let it pose as a universal law.

What this looks like when you build it

If judgment about relevance is the real work, then a system meant to do professional work has to be built around that judgment rather than around retrieval. That conviction is the one ktAi is built on, and it shows up most clearly in what the system refuses to use.

The system treats “is this information available” and “is it relevant to the question in front of me” as two separate questions, and spends most of its effort on the second. It scores candidate material on two axes that it deliberately keeps apart: how well a source answers this specific question, and how authoritative it is in its own right. Authority can strengthen a source that is already on point, but it can never lift an off-topic source into the answer. The consequence is the one that matters: a maximally authoritative source that does not bear on the question stays out. A landmark ruling from the wrong jurisdiction, a foundational statute about a different subject, is held back precisely because of what it does not address, no matter how impeccable its pedigree. The most seductive failure in this field is a beautifully sourced answer to a question nobody asked, and keeping relevance and authority on separate axes is what guards against it.

Most of the engineering, honestly, goes into withholding. Superseded law is removed outright, because citing yesterday’s rule as today’s is worse than saying less. Unverifiable claims are declined rather than asserted, though the system is careful to distinguish “I cannot verify this” from “I have not yet catalogued this,” so that a real authority is not suppressed just because an index has not caught up. When verification can only be partial, the section is carried forward with the gap marked and surfaced in the report to the person who will sign it, never smoothed over with a confident-looking score. The discipline is to prefer a visible gap to a plausible fabrication.

Honesty requires saying where that judgment is still thinner than it should be, and the sharpest gap is one I will name precisely. Across jurisdictions, the system is better at keeping the wrong country’s law out than at reaching for the right foreign one. It reliably sets aside an authority from a jurisdiction that does not apply. It does not yet actively follow a cross-border link, a treaty, a foreign-tax-credit relationship, to pull in a foreign authority that is relevant precisely because of that connection. Today a treaty surfaces when its text answers the question, not yet because the cross-border relationship makes it relevant. The map of those relationships is built; the retrieval does not yet walk it. The protective half of cross-border judgment is in place. The connective half is still being built, and I would rather tell you that than let the word “cross-border” carry more than it currently earns.

The question worth asking

The industry is asking who can hold the most information, and the answer is getting less interesting by the month, because soon the answer will be everyone. The public corpus is a subscription away. The model has read it all. The context window keeps growing.

The question that survives all of that is the older one, the one every experienced professional already lives by without saying it aloud. Not who knows the most. Who knows what to leave out. An expert was never the person who had read everything; an expert is the person who can look at a thousand sources and set nine hundred and ninety-six of them aside without a second thought, because they do not bear on the question, however impeccable they are. That judgment is hard to build, harder to prove, and it is the whole game. The systems worth trusting will be the ones built around it rather than around the size of what they can reach.

That is not something a benchmark settles, and not something the size of a context window can claim. It gets settled the way it has always been settled in this profession, one signed report at a time.