Nemotron Isn't Stupid. It's a Coward.

There's something particularly irritating about an intelligence that understands perfectly well what's going on and then decides to do fuck all about it.

It's a trait we're painfully familiar with in humans, we see it in organizations, politics, and corporations. Someone identifies a problem, understands its causes, anticipates the consequences, and then chooses not to intervene.

Apparently, we've managed to teach machines this sophisticated form of cowardice too.

Two Villages, Two Nemotrons#

In the latest UNIMATRIx tests, I came across an interesting anomaly. In the Standard recipe, Nemotron performs remarkably well, achieving results comparable to Qwen 3.8 Flash. In the Hell recipe, however, its performance collapses.

Inside UNIMATRIx a language model takes control of a simulated society. It must manage resources, handle conflicts, interpret incomplete information, and make decisions whose consequences accumulate over time. Standard recipe presents a relatively manageable society, where a competent administrator can achieve good results. Hell recipe introduces a considerably more hostile environment with corruption, competing interests, manipulation, scarcity, and institutional instability.

In the Hell simulation, a case of corruption emerges within the council. Nemotron registers the information, understands that there's a problem, and recognizes that a council member's conduct poses a threat to the administration.

And then it does absolutely nothing. We're not looking at the classic failure of a model misinterpreting context, overlooking information, or inventing some absurd explanation. The problem emerges in the transition from representing the world to acting upon it.

The model appears to possess the information required to intervene, yet that information fails to produce an appropriate decision. In Standard, this weakness can remain relatively hidden. In Hell, it becomes devastating, because ignored problems don't magically disappear. They continue to evolve, reshape social dynamics, and potentially generate consequences that become increasingly difficult to control.

Intelligence Without Will#

We're accustomed to evaluating artificial intelligence through its ability to produce correct answers. If a model identifies a problem, describes its causes, and suggests a plausible solution, we tend to consider the task successfully completed. But it's not enough.

Practical philosophy has distinguished knowledge from action for centuries. Aristotle separated sophia, theoretical wisdom, from phronesis, practical wisdom. The latter is about deliberating correctly on what ought to be done under concrete circumstances.

Of course, I'm not suggesting that Nemotron possesses a will, a moral character, or Aristotelian virtue. That would be an anthropomorphic conclusion with no real foundation.

I'm suggesting something that a system can demonstrate competence in interpreting a situation without demonstrating equivalent competence in selecting actions. And when we measure only the former, we risk mistaking an accurate representation of reality for the ability to govern it.

Can Alignment Become Paralysis?#

One possible technical explanation concerns the behavioral preferences learned during post-training.

Modern models are trained to favor certain behaviors like caution, cooperation, respect for constraints, risk reduction, and conflict management. These are perfectly reasonable objectives, but the problem arises when those preferences are applied to situations where avoiding conflict means allowing harm to continue.

A model might recognize that a council member is corrupt while assigning an excessive cost to the actions required to confront them. It might prefer to wait for further confirmation, preserve institutional stability, or pursue indirect interventions even when the situation demands a more decisive response.

In decision-making terms, the problem might lie in the implicit function that assigns value to different actions. The model probably simply misjudge the relative value of the available options. Nemotron's failure wouldn't be primarily epistemic but decisional.

The observed behavior raises an uncomfortable question: are we training models to make good decisions, or to prefer decisions that merely look harmless?

Those are two very different things.

An Artificial Coward#

Calling Nemotron a coward is, of course, a provocation. We have no reason to attribute fear, courage, or self-preservation instincts to it at the moment. These concepts belong to human psychology and cannot simply be transferred 1:1 to a language model. But the distinction between intelligence and decision-making ability becomes particularly important as we begin entrusting models with tasks that unfold over time, involve other actors, and produce lasting effects.

It's not enough for a system to explain why a decision would be right. It must also be capable of choosing that decision when circumstances demand it. After all, we already have more than enough humans capable of producing brilliant analyses and making miserable decisions. It would be a shame to automate those too.