A fascinating piece of research from MIT CSAIL recently caught my attention: Sparks of Weight-Space Programming: LLMs Can Natively Read & Write Neural Network Weights, published November 2, 2026, by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab.
It’s a fascinating read, and very much worth the time. This blog isn’t a review of that paper per se. For a “dedicated “pure play” review of that paper, see Asian Dad Energy’s nice video: The Black Box of AI May Have Just Been Cracked?
The researchers explored whether an advanced LLM, GPT-6 Astra, could examine the weights of a neural network, infer something about what the network does, and even modify those weights to change its behavior.
This reminded me of an idea I developed nearly twenty years ago using SQL Server Analysis Services (SSAS), which eventually evolved into my Game of Dōh.
The connection isn’t in the underlying algorithms. It’s in how we might look inside a complex system, understand how its parts influence its behavior, and deliberately change those parts to achieve a desired outcome.
Looking Inside a Neural Network
A neural network is, in a mathematical sense, a function. Inputs are transformed into outputs through layers of calculations governed by weights. Training determines those weights, but the resulting function can be extraordinarily difficult for humans to interpret.
The MIT researchers explored whether an LLM could reason about those weights directly. In effect, a broadly capable neural network was being asked to examine, understand, and potentially repair another neural network.
In one set of experiments, Astra was given the weights of neural networks trained to play four different games: Breakout, Pong, Pac-Man, and Freeway. Despite the networks having identical architectures, Astra correctly identified which game each had learned to play by examining their weights. In other experiments, it constructed small game-playing neural networks from scratch, including one for Go. These experiments are particularly interesting because they demonstrate both interpreting a learned function and constructing one to perform a specified task.
In one experiment, a neural network controlling a car could not get over a ramp. It needed to back up first to gain enough momentum. Astra identified the problem and modified approximately 3.5% of the network’s weights. The revised network succeeded in all 1,024 evaluation episodes.
In another experiment, Astra examined the open-weight model, Qwen 3.5-0.8B, which had a tendency to become trapped in repetitive reasoning loops. By modifying only 0.016% of its weights, Astra dramatically reduced the looping behavior. The improvement extended beyond mathematical reasoning to instruction following.
These are fascinating results because the changes were not produced through conventional gradient-based retraining. Astra reasoned about the model and constructed targeted modifications. The researchers call this emerging capability Weight-Space Programming: treating neural-network weights as something a general-purpose intelligence can inspect, reason about, and deliberately modify.
Naturally, I have questions about scale. How effectively can such reasoning work with much larger neural networks? How thoroughly must a modified model be tested to ensure that improving one behavior hasn’t inadvertently damaged another? Can an LLM identify and navigate interactions among many competing behaviors rather than addressing a relatively isolated problem?
Interestingly, the researchers also encountered cases where targeted weight modifications did not produce meaningful improvements, suggesting that correcting a recognizable failure and creating a capability that doesn’t already exist may be considerably different challenges. These aren’t critiques of the research. They’re questions inspired by the possibilities it opens.
I also see this research in the historical context of AI’s progression from impressive accomplishments in narrowly defined problems toward increasingly general capabilities. In 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov. It was a remarkable accomplishment, but one confined to a game with well-defined rules and objectives. It would be another 25 years before ChatGPT arrived in 2022, demonstrating the ability to converse and reason across an enormous range of subjects. The progression wasn’t a direct one, of course, but it illustrates the considerable distance between mastering specific problems and developing broadly applicable intelligence.
This general-purpose capability is especially interesting to me because of something I’ve explored in my two most recent articles, Abductive Calculation and Eliciting Creative Questions from AI. One of the greatest strengths of LLMs is their extraordinary breadth of knowledge across domains. Unlike a chess-playing system designed for a particular problem, an LLM can draw upon concepts from physics, biology, economics, engineering, and countless other fields to propose relationships that weren’t explicitly anticipated when the problem was formulated.
In Abductive Calculation, I explored how AI might construct plausible calculational pathways when the relationship between available information and an unknown quantity isn’t already established. In Eliciting Creative Questions, I took this further, exploring how an LLM’s breadth of knowledge could help discover unexpected relationships and entirely new lines of investigation.
This brings us back to the MIT research. What intrigues me isn’t merely that Astra can modify neural-network weights. It’s that a broadly capable intelligence is being used to examine another function, hypothesize how it behaves, and devise a targeted intervention. I see these experiments as potentially an early demonstration of a much broader capability. What might happen when such reasoning is applied to increasingly complex systems, involving many interacting behaviors and objectives across multiple domains—as an enterprise does?
Looking Inside an Enterprise
During the long weekend of Thanksgiving 2008, while working in Microsoft Consulting Services’ Global BI Practice, I further developed a concept I called “Just Right Targets”.
At the time, I was deeply involved with SSAS and its data mining capabilities. Data mining was still not the mainstream, at least as it would be when “Big Data” and “Data Science” became household terms just a few years later. I wanted to explore applications beyond using multidimensional cubes as fast sources for analytical and reporting tools.
SSAS offered an interesting combination of capabilities: KPI calculations, what-if analysis through writeback, and analytical processing. Why not combine them to search for better configurations of an enterprise?
My approach was based on what I later described in Levels of Pain. An enterprise has many goals, often competing. Improving one KPI may worsen several others. But a little pain might be tolerable if it relieves something much worse. That’s the nature of investment and sacrifice.
The idea was to modify values through SSAS writeback, recalculate KPI statuses, classify their levels of pain, and use a genetic algorithm to search for combinations that improved the overall situation without exceeding acceptable pain thresholds.
For example, we might tolerate additional labor costs or longer working hours to meet an important deadline, provided we don’t cross thresholds that cause unacceptable financial consequences or employee burnout.
In December 2025, I expanded this work into The Complex Game of Planning, introducing Dōh. The basic idea is to treat enterprise planning as a move-based game of changing conditions, consequences, and tradeoffs.
Dōh treats enterprise planning as a game, somewhat analogous to chess or Go, except that its board is an enhanced strategy map representing goals, KPIs, configurations, and the relationships among them. Each configuration represents something we can change, such as marketing expenditures, staffing levels, pricing, or infrastructure capacity. KPIs measure the consequences, while their levels of pain indicate how tolerable those consequences are. A move consists of changing one or more configurations, recalculating the affected KPIs, and evaluating the resulting state of the board. In this sense, Dōh is less about isolated optimization than about seeing how one move reshapes the larger situation.
For example, suppose an e-commerce business is suffering from poor sales conversion. Increasing marketing expenditures might improve traffic and sales, but it could also overwhelm infrastructure, increase operating costs, or damage customer experience. Dōh evaluates the resulting configuration against the pain thresholds of all relevant KPIs. If improving conversion creates unacceptable pain elsewhere, that isn’t really a solution. The game must search for another configuration, perhaps one that first increases infrastructure capacity before pursuing additional traffic, or that approaches the problem from a different angle altogether.
The larger challenge is that a move can change the conditions governing subsequent moves. There may be no acceptable solution within the current configurations, requiring changes to the relationships, constraints, or even the rules themselves. The search, then, is not simply for a better number or even a better immediate move, but for a viable path through an enormous space of interacting possibilities. That is why I think of Dōh as a generalized planning problem rather than merely KPI optimization.
In a sense, an enterprise is also a function. It takes inputs—resources, customer demand, information, employee effort, and countless other variables—and transforms them into outputs. Its behavior depends on configurations, processes, constraints, and interactions among its components.
Unlike a neural network, the enterprise isn’t one completely specified mathematical function. Much of its behavior is only partially understood, its relationships change over time, and human decisions introduce additional uncertainty.
Nevertheless, we can construct a functional representation of significant portions of it through a strategy map, KPIs, relationships, and configurations.
The correspondence becomes clearer when we consider the components of a neural network alongside those of an enterprise performance management system.
| Neural network | Enterprise performance management |
|---|---|
| Inputs — Values supplied to the network | KPI inputs — Enterprise measures, operational data, and configuration values |
| Weights and computational structure — Parameters and connections that determine the function’s behavior | KPI calculations and dependencies — Value, target, and status formulas, their parameters, and the KPI Status Relationship Graph |
| Interpreting learned relationships — Investigating which components influence behavior | Effect Correlation Score (ECS) — Measuring whether intended relationships between KPI statuses hold in observed data |
| Objective or reward — Evaluating whether the network produces desirable behavior | Targets and Levels of Pain — Determining desirable outcomes, acceptable sacrifices, and unacceptable consequences |
| Weight-space programming — Deliberately modifying weights to change network behavior | Just Right Targets / Dōh — Modifying enterprise configurations, recalculating KPIs, and searching for better outcomes |
The correspondence isn’t exact. Neural-network weights are learned numerical parameters, while enterprise KPI formulas and relationships are generally explicitly defined or empirically discovered. Nor does changing a KPI calculation necessarily change the enterprise itself. The important similarity is that both approaches seek to understand a complex computational representation well enough to identify promising interventions and evaluate their consequences.
And then we can ask a question analogous to the one that fascinated me about the MIT research:
Can we look inside that function, identify which parts to change, and predict how those changes will affect the behavior of the larger system?
Computers surpassed the best human chess players decades ago. IBM’s Deep Blue defeated Garry Kasparov in 1997, and DeepMind’s AlphaGo defeated Lee Sedol at Go in 2016. These were remarkable accomplishments, but both games have fixed rules, well-defined moves, and unambiguous conditions for winning.
Enterprise planning is a rather different game. We don’t necessarily know all the rules, the relationships among the pieces aren’t completely understood, and there isn’t one definition of winning. Improving one goal may damage another, sometimes in ways we didn’t anticipate. Worse, changing the enterprise can change the rules governing its subsequent behavior.
That’s part of what Dōh attempts to address: searching for moves while continuously reconsidering the board, its relationships, and the acceptable tradeoffs among competing objectives.
This is where the general-purpose nature of LLMs becomes particularly intriguing. In my original Just Right Targets experiment, I supplied the KPI calculations and rules governing acceptable levels of pain. A genetic algorithm then searched for better configurations. But understanding the relationships within an enterprise, especially across domains, is a much larger challenge.
That’s already a central purpose of the Tuple Correlation Web (TCW) and Time Molecules. The TCW investigates statistical relationships among metrics across enterprise domains, guided in part by observations retained in the Insight Space Graph (ISG). Time Molecules approaches connections through sequences of events, recurring patterns, and transitions across processes. Neither establishes causation merely by discovering a relationship. But what if a general-purpose LLM could reason across these different forms of evidence, propose additional connections through abductive reasoning, and help identify interventions that hadn’t been explicitly anticipated?
Dōh could then provide a framework for exploring the consequences of those interventions, including their effects on other goals and subsequent moves. The enterprise itself would ultimately provide feedback about which hypothesized relationships and interventions actually hold.
The Bigger Question
The underlying mechanisms are quite different. MIT’s experiments involve directly interpreting and modifying neural-network weights. Just Right Targets and its later evolution into Dōh involve evaluating enterprise configurations through relationships, KPI calculations, and search algorithms.
But both invite us to go beyond treating a complex function as a black box. We don’t merely want to know what the system will produce under its current configuration. We want to understand how it works well enough to deliberately change what it produces. And in both cases, a successful intervention must ultimately be judged against more than the immediate problem it was intended to solve.
That’s what makes the MIT research so interesting to me. It suggests the possibility of moving from observing a complex function’s behavior to reasoning about how to modify it.
I’ve been exploring a related ambition from the enterprise perspective for almost two decades. I don’t know that there is a really good answer somewhere out there for complex problems, but I keep on the lookout for it.
Different functions. Different mechanisms. But a remarkably similar question: What should we change, and what happens to everything else when we do?
