AI trained on history repeats history. To eliminate work, it first has to unlearn the assumption that the work is necessary.
A frontier model was recently handed the raw weights of a small open-weight model that kept falling into repetitive reasoning loops. No training loop. No gradients. Fifty forward passes.
It traced the habit to two layers, removed a single direction from 64 columns in each, and changed 0.016% of the weights. Math accuracy went from under 8% to over 48%.
Most of the coverage went straight to AGI. That is not what caught my attention.
What caught my attention is simpler: a habit inside a model can be located and removed. That matters to anyone building intelligence meant to do something other than repeat how work is done today. Before AI can eliminate work, it has to unlearn the assumption that the work is necessary.
AI learns yesterday’s operating model
Every model we put into enterprise operations learns from history: tickets, alerts, runbooks, resolutions. History is a record of how we worked, not how we should have worked. So the model gets very good at doing yesterday faster.
Here is a simple example. Enterprises invested heavily in control planes: virtualization, converged, hyperconverged, cloud. Every one of them ships with built-in monitoring. Yet most organizations still layer open-source or commercial monitoring tools on top, because we were taught that operations needs monitoring tools.
When we built historical intelligence on our platform, none of the dispositions it produced asked why the monitoring tool existed at all. The data never asked. No one in the data ever asked. The model learned the modality as a given.
This is not a monitoring problem. Every domain has its version: approvals that a policy engine already enforces, reconciliations a system already guarantees, handoffs that exist only because two teams were organized that way years ago. A model trained on that history will automate all of it, beautifully.
Labels teach instances, not principles
So we taught it. We labeled event-driven alerts and tickets from a monitoring tool so the model would question why that tool existed. It worked, for that tool.
The next time the pipeline ran, a different tool showed up. The model did not recognize the same pattern. We labeled again. The same happened with platforms. We kept building known machines and known behaviors, one at a time.
What we had taught was “question Tool X.” What we needed was “question any layer that duplicates what the platform underneath already does.”
We experimented with weights and temperature in the harness to get consistent dispositions. Temperature was never the real lever. It changes how randomly a model samples, not what it believes. The lever is the weights, or the activations flowing through them. That is why this research matters to me.
Question the modality, keep the gap
Unlearning does not mean rejecting. A crude version would eliminate every monitoring tool, including the ones that matter.
Application performance monitoring is the clean example. Control planes see infrastructure: host health, capacity, VM state. They do not see code-level traces, transaction flows across services, or what the end user actually experiences. APM covers a real gap. It survives the question.
So the target is not “monitoring tools are unnecessary.” The target is the shortcut: assuming the tool is needed without asking. For every layer, the model should ask three things. What does the native layer already cover? What does this add? Is the remainder worth paying for?
That gives three outcomes, not two:
- Eliminate when the overlap is complete.
- Keep when the tool covers a real gap.
- Shrink when it does both: keep the tool, drop the overlapping modules and licenses.
Put the disposition in the model, keep the facts in retrieval
The coverage line keeps moving. Cloud platforms now ship native tracing and application insights. Part of what justified APM five years ago may be redundant today. If you train “APM is required” into the weights, you have just created the next status quo to unlearn.
So the architecture splits:
- The disposition to question lives in the model. Teach it with contrasting reasoning: one trace that accepts the modality, one that questions it. On open-weight models, that difference can steer the model at inference or train a small adapter. Use synthetic scenarios across domains so it learns the pattern, not the products: monitoring over native telemetry, ticketing over self-healing platforms, approvals over policy engines.
- The facts live in retrieval. A capability map of what each platform natively does, refreshed as platforms change.
- New tools map to capabilities, not labels. A tool the model has never seen gets classified by what it does and compared against native coverage. No relabeling. Only the capability map changes, and that is data, not training.
This is where the weight-space research earns its place. If “skip the question” is a localized habit, like the looping in that study, it can be found with a small set of contrasting examples and suppressed. And you can check inside the model that it is gone, instead of hoping your tests catch every case.
What this doesn’t solve
- It removes habits, not beliefs. The looping fix worked because the behavior was concentrated in a few places. A belief spread across a model’s world knowledge is much harder to remove.
- It only unlocks what is already there. The same procedure on a smaller model produced no gain, because that model never had the capability. The model must already be able to reason “this layer is redundant.”
- It is early. The study is unreviewed, used one frontier model, and worked on small targets.
- Someone still has to define what questioning looks like. That is the leadership job. These tools make unlearning cheaper, more targeted and verifiable. They do not decide which modality deserves to be questioned.
The industry is focused on teaching AI to do today’s work faster. The harder and more valuable problem is teaching it to stop accepting today’s work as a given.
If your AI learned everything it knows from how your organization worked yesterday, who is teaching it to question yesterday?
This article is part of the Enterprise Intelligence series on building intelligence that drives peak business performance.