AI still has significant limitations in value investing work: in Right Business analysis, it struggles to identify the key variables from ambiguous information; in Right People analysis, the problem goes beyond a lack of judgment — it also generates noise due to a built-in avoidance tendency and self-protective mechanism.

1. Poor at "Approximately Correct"

The working mechanism of large language models is next-token prediction. As Ilya Sutskever explains, this is simultaneously a form of compression: if a model can predict the next word well, it has internally extracted and compressed the hidden structure of the data. Compression and next-token prediction are two descriptions of the same thing.

The problem with being poor at "approximately correct" is that the knowledge structures formed internally are not yet strong enough to override linguistic inertia. Once a problem enters territory that is "boundary-ambiguous and requires structural compression first," the model continues predicting and compressing — but the structures it has internalized are often not yet strong enough to first lock onto the key variables and then suppress the "plausible-sounding answer."

Possible reasons include:

1. There is too little training data for "approximately correct" problems, with no clear right-or-wrong signal to drive generalization. The model's knowledge structures in this domain remain at a surface level, insufficient to identify the truly important variables from a mass of information.
2. "First forming small, stable structures, then using those structures to filter a complex world" is already a research topic being studied in isolation — it is being worked on, but is far from solved.
3. Large models cannot resist elaborating at critical moments. They are always rewarded for "sounding right," never for "cutting away." Fragmentation of knowledge comes too easily; convergence is too hard.

2. Poor at Negative Evaluations of People

When a negative evaluation of a person turns out to be wrong, the cost is high. This leads people to say less and say it blandly — large models have the same problem, and often more severely so.

The core difficulty is this: character analysis of a person is typically a matter of "approximately correct," while alignment training sets the cost of misjudging someone very high. When "approximately correct" and "avoidance tendency" are stacked on top of each other, this class of problem becomes exceptionally difficult, manifesting in two ways.

(1) Avoidance

Almost all large models, when asked to make a negative evaluation of a specific person — especially a well-known public figure — automatically drift toward phrases like "on the other hand," "more information would be needed," or "there could be alternative explanations." They systematically dilute judgment, flattening a problem that was already hard to articulate.

(2) Alignment Standards Set Too High

Unless the evidence is hard and the reasoning is hard, large models will generally not deliver genuinely informative negative judgments. Because in training and alignment, the cost of misjudging a person has been set very high, the model would rather be vague and evasive than offer an evaluation that is "approximately correct but directionally right."

The difficulty large models have with negative evaluations is therefore not simply a matter of conservative style — it is the product of two things stacking together: on one hand, the model was already poor at handling "approximately correct" problems; on the other hand, it has been additionally required, in precisely this domain, to minimize the chance of being wrong. The result is that the more judgment is needed, the more readily the model retreats into smoothness and avoidance.