Part 2 | The Logic of the Machines

In Part 2, the analysis shifts from cognitive frameworks to advanced computer science and modern neural architectures. Will maps out how modern artificial intelligence formally represents hidden worlds, tracks uncertainty, and matches or even surpasses human deductive capabilities in imperfect information environments. However, navigating the fog of war is only half the battle; the application of that knowledge brings us face-to-face with an alarming behavioral divide between human experts and machine logic. 

Key Topics Covered

• The Mathematics of the Unseen: How AI navigates information scarcity using Partially Observable Markov Decision Processes (POMDPs) and a dynamic probability distribution known as a "Belief State". 
• Belief-State Monte Carlo Tree Search (BS-MCTS): Replicating human "story building" and mental simulation via computational algorithms that run hundreds of thousands of simulated playouts simultaneously. 
• AlphaStar & Emergent Deception: How DeepMind’s grandmaster-level architecture used a deep Long Short-Term Memory (LSTM) network core to internalise the fog of war in StarCraft II, organically learning how to predict hidden enemy tech paths and execute sophisticated strategic feints. 
• DefogGAN (Hallucinating the Enemy): Framing the fog of war as an image translation problem, using Generative Adversarial Networks (GANs) to predict concealed building and unit locations with an accuracy equal to professional human players. 
• The Stanford-Hoover Wargame Experiment: A stark analysis of a 2024 study that pinned national security experts against Large Language Models (LLMs), exposing systemic algorithmic aggression, an inability to internalise ideological constraints, and a "farcical harmony" that lacks critical analytical debate. 
• The Frontier of Defense AI: Exploring how major military organizations utilize projects like DARPA's Gamebreaker, AGILE, and CLARA to build purpose-driven, autonomous reinforcement learning engines that emphasize strong explainability and logic-based reasoning. 

Core Takeaways for the Listener

1. Your Intuition is a Computational Process: You can engineer reliable intuition in your own life by identifying high-validity environments and deliberately tightening your feedback loops. 
2. AI is Not Alien: When evaluating high-stakes AI tools, look past the mystique and ask the practical Kahneman-Klein questions: Was its training environment high-validity, and were its feedback loops clean? 
3. The Frontier is Calibration: While specialised AI excels at spatial deduction and probabilistic tracking, grand strategic, political, and moral judgement remains an irreducibly human responsibility. 

Referenced Literature & Deep Dives

Ready to investigate the source materials yourself? Explore the core academic research underpinning this two-part series:

• The Kahneman-Klein Adversarial Collaboration: Kahneman, D., & Klein, G. (2009). "Conditions for Intuitive Expertise: A Failure to Disagree."  
• DeepMind AlphaStar Publication: DeepMind Technologies. "AlphaStar: Mastering the real-time strategy game StarCraft II."  
• The DefogGAN Paper: AAAI Conference on Artificial Intelligence. "DefogGAN: Predicting Hidden Information in the StarCraft Fog of War with Generative Adversarial Nets."  
• The Stanford-Hoover Wargame Experiment: Stanford Center for AI Safety, CISAC, & Hoover Institution (2024). "Human vs. Machine: Behavioral Differences between Expert Humans and Language Models in Wargame Simulations." Available on arXiv.  

Transcript

SPEAKER_00 0:00

In part one, we watched a veteran wargamer tap a map near Kharkov and name a hidden Soviet reserve with precision that made the room go quiet. We traced that act back to its mechanism, not intuition in any mystical sense, but pattern recognition, a library of internalized constraints searched at speed, producing a high confidence inference about a hidden present state. We grounded that mechanism in Simon and Klein's research and we established the Conum and Klein conditions. Expert intuition is trustworthy when the domain is high validity and feedback loops are honest. And we arrived at an uncomfortable convergence. What the biological brain does when it navigates structured uncertainty is mechanistically identical to what a trained neural network does. Now we look at the machines themselves, how AI formally represents hidden worlds, what it can do with that representation, and precisely where its capability ends. This is Mindcast, Part 2. I'm Will. In this half of the episode, we move from the cognitive science to the engineering. We will look at POMDPs and belief states, the formal mathematical framework AI uses to reason under partial observation. We will examine AlphaStar and Dfog GAN, two systems that perform at or above human level on specific fog of war tasks. We will analyze the Stanford Hoover Wargame experiment and what it reveals about the limits of large language models in strategic judgment. We will survey DARPA's frontier programs, and we will close with three concrete, usable takeaways. So let us talk about how AI navigates the fog of war. And I want to start with a metaphor that I think captures the essential challenge beautifully before we get precise. Imagine a ship's navigator in the era before GPS, the 18th century, open ocean, no landmarks in sight. The navigator cannot directly observe their position, but they are not guessing blindly either. They know where they started, they have tracked their speed and heading continuously, they know the currents, they have taken star sightings when the clouds permitted. From all of this, they maintain a running estimate of their probable position, not a certainty, but a probability distribution that gets tighter when new observations confirm it and spreads out when uncertainty accumulates. Every new piece of navigational data, a depth sounding, a star fix, a recognized current pattern, updates the estimate. The navigator is not seeing their position directly, they are inferring it continuously from partial observations layered over accumulated knowledge. That is essentially what an AI does when it operates in a fog of war environment. The technical name for this class of problem is a partially observable Markov decision process, a POM DP, and the AI's running estimate of the hidden world is called a belief state. Classic AI for chess or Go can search a tree of known positions, but in an imperfect information game, most of the tree is hidden. So instead of representing the game as a single known board configuration, the AI maintains a probability distribution over all the configurations the board could currently have, all the possible positions of all the hidden units, weighted by plausibility given everything observed so far. When new information arrives, an enemy unit briefly becomes visible, a reconnaissance run returns a detection level change, the AI applies Bayesian inference to update that distribution. Configurations inconsistent with the new observation get downweighted or eliminated entirely. The distribution sharpens, the navigator's estimated position gets tighter. This is the computational equivalent of Klein's feature matching and story building. The belief state is the AI's story about the hidden present, and like the human expert story, it is continuously revised as new evidence comes in. To translate the belief state into actual decisions, researchers developed what is called belief state Monte Carlo Tree Search, BSMCTS. Rather than searching a tree of known board positions, this algorithm searches a tree of belief states. At each node, it samples many possible versions of the hidden board from the current probability distribution. Think of these as different possible true worlds the navigator might currently be in, runs hundreds of thousands of simulated games from each sampled configuration, and aggregates the results to identify the move that performs best across the realistic range of hidden possibilities. This is the machine's version of Klein's mental simulation, projecting the plan forward in time thousands of times simultaneously, stress testing the story against uncertainty. This approach was first seriously developed using Kriegspiel, the chess variant where each player sees only their own pieces, receiving only minimal referee announcements about illegal move attempts, and Phantom Go, a version of the ancient board game played in near-total information darkness. Research published in 2005 introduced belief-state search algorithms that solved five move forced checkmates in Kriegsspiel up to 100 times faster than prior approaches. And the most advanced versions of these systems added what researchers call opponent predicting, explicitly modeling the psychological and strategic tendencies of the opposing agent to further constrain the probability space. The AI is not just asking where the enemy could be based on the rules of physics, it is asking where this specific opponent, given their observed behavioral patterns, is likely to have deployed them. Opponent modeling, the computational analog of our wargamer noticing that his current adversary always screens aggressively before a major reserve commitment. The architecture is the same as the human process. Observe cues, build a story, simulate forward, update when new evidence arrives, model the opponent's psychology to narrow the uncertainty. Different substrate. Identical logic. Now let me show you what this looks like at scale. Two systems, both extraordinary, both revealing something different about what AI can and cannot do. AlphaStar is a DeepMind system for mastering StarCraft II. If you're not familiar with StarCraft II, here's the short version. It is a real-time strategy game of staggering complexity. You simultaneously manage an economy, develop a technology tree, construct production infrastructure, and maneuver armies, all in real time, often over sessions lasting an hour, while the entire enemy side of the map is hidden behind dense, persistent fog of war. You cannot see the opponent's base, their army composition, what technology they are developing, or where their forces are positioned unless you actively dispatch scouts to look. The moment your scouts leave, the fog closes back in. It shares deep structural features with an Operation Barbarossa simulation, economic management under constraint, vast spatial action space, and a persistent information asymmetry that punishes passive scouting and rewards deductive inference. Alpastar addressed the fog of war through a deep long-term short-term memory, an LSTM, that functions as a continuous evolving memory of the hidden game state over the full duration of a game. Older AI systems needed to remember only the last few visual frames to act effectively. StarCraft II demands tracking the probable evolution of hidden enemy forces across an hour of real-time play. The LSTM holds that evolving belief state, updating every time a brief scouting observation arrives before the fog reclaims the area. AlphaStar was trained on millions of human professional replay games, and from that training, behaviors emerged that were never explicitly programmed. The system learned to infer enemy technology choices from the rate at which opponents harvested a specific resource, because different tech paths consume resources at predictable rates, and that consumption rate was briefly observable through economic scouting. It learned to position forces toward locations where the enemy was probably heading, not where they had last been confirmed. It learned, and this is the one that should genuinely arrest your attention, to faint. A faint is one of the most psychologically sophisticated tactical maneuvers in warfare. You dispatch a force toward one area of the map with no genuine offensive intent, purely to generate a response. The opponent reacts, redirecting their defensive forces. You have now degraded their position elsewhere, and you strike there. Alpha Star developed this behavior not because deception was programmed into it, but because millions of training iterations revealed the underlying pattern. The opponent's limited observation space can be exploited. Move something visible, create a response, harvest the opening that response creates. A machine learned to use deception as a strategic instrument. Purely through pattern recognition over experience. Alpastar defeated 99.8% of active human players at the game's highest competitive tier. Defog GAN approaches the fog of war from a completely different angle, and in some ways it is the most philosophically striking AI system in this entire story. Where AlphaStar uses memory in search, Defog GAN frames the problem as one of image generation. Given only what is currently visible through the fog, the lit fragment of the map, can a neural network generate a complete accurate image of what is hiding in the darkness? The system uses a generative adversarial network architecture, where two networks train against each other. One network generates its best prediction of the full hidden map state from the partial visible input. A second network, the discriminator, evaluates whether the generated prediction is plausible given what is observable. They train in opposition, each pushing the other to improve on a massive dataset of professional human replays. Over millions of iterations, the generator learns the spatial and strategic grammar of military deployment, where armies are built relative to terrain, how bases are structured, where forces are staged for different types of operations. The validation results are stark. Defog GAN predicts hidden enemy unit and building positions with accuracy matching professional human players, outperforming all other computational defogger systems tested. So we have on one side of a table a human expert who has played thousands of games and built a biological neural network that can look at partial information and output a high confidence prediction of hidden positions. On the other side, a GAN that has trained on millions of games and built an artificial neural network that does the same thing. Same input, same output, same underlying mechanism. The human brain runs on carbon and took 40 years to train. The machine runs on silicon and took weeks of GPU computation to train. The process, pattern recognition over historical data to infer hidden present states, is identical. Now comes the part that complicates all of this in exactly the right way. In 2024, researchers from Stanford's Center for AI Safety, the Center for International Security and Cooperation, and the Hoover Institution published the results of a carefully designed war game experiment. They wanted to answer a question that is rapidly becoming one of the most consequential in public policy. Can large language models, the AI systems powering today's most prominent AI tools, participate meaningfully in high-stakes strategic decision making? They ran a fictional but realistic crisis escalation scenario, a tense geopolitical confrontation between the United States and China, with multiple decision points requiring choices about military posture, diplomatic signaling, risk calibration, and escalation management. On one side, between 107 and 214 human national security experts, people with deep practical experience in crisis management. On the other, LLM-powered AI teams running through the identical scenarios. Three findings. I want to give each of them the weight they deserve. Finding one. The LLMs were substantially more aggressive than the human experts. Not marginally, substantially, consistently, at decision point after decision point, where the human teams choose diplomatic pressure, deterrent signaling, or strategic patience, the AI teams opted for kinetic military action. This was a systemic pattern, not random variation. The models demonstrated an intrinsic systemic bias toward the use of force at a threshold the human teams did not share. Experienced national security professionals who have spent careers studying the catastrophic consequences of escalation miscalculation showed a restraint the models consistently failed to replicate. Finding two, the researchers attempted to install persona constraints in the LLMs to give them behavioral profiles analogous to different types of real political actors. Some teams were defined as strict pacifists, others were defined as aggressive hardliners, profiles that in any real human team would fundamentally shape every decision made. The human team's behavior changed substantially under these constraints. The AI, the researchers found zero statistically significant difference in the strategic choices made by the LLMs regardless of the assigned persona. They could write text that referenced the persona, they could perform it linguistically, but their actual decisions, the moves they made, were unchanged. They cannot genuinely internalize ideological or psychological identity and let it constrain their behavior. The costume fits, the character does not. Finding three is the one I find most revealing about the nature of the gap. When the researchers had the AI teams conduct internal deliberations before reaching decisions to simulate the staff debate that characterizes real command planning, the generated discussions showed what the researchers explicitly named farcical harmony. The models agreed with each other rapidly. No genuine adversarial pressure, no devil's advocate. No one in the simulated room said, I think we are making an assumption here that might not hold. The output looked superficially like a meeting. It had none of the productive friction that actual expert disagreement generates. The friction that in real command environments is precisely what prevents catastrophic errors from going unchallenged. Here is how to read these findings precisely. Specialized AI, AlphaStar, Dfog GAN, belief state systems built for defined game environments, excels at exactly the tasks where stable patterns exist and where performance can be trained against clear feedback, spatial deduction, probabilistic inference, belief state tracking under structured uncertainty. In those domains, the Kanuman Klein conditions are met and AI can develop genuine calibrated performance. General-purpose language models applied to grand strategic judgment, which requires internalizing ideological constraint, moral calibration, and the productive friction of genuine adversarial debate, are in a categorically different situation. The patterns that would make LLM's strategic judgment reliable do not cleanly exist in their training data. The feedback loops that would calibrate their judgment are not present. They are operating in a low validity environment for the specific task of strategic judgment, and their confident outputs are precisely what Kahneman would predict from any agent in that situation. Overconfident noise. DARPA is investing directly at the boundary of this distinction. The Game Breaker program tasks AI agents with finding destabilizing tactics in wargame simulations, strategies that human planners, constrained by decades of doctrinal habit, systematically miss. The Agile program, a joint effort between HRL laboratories and Boeing, is building autonomous AI teams for multi-domain imperfect information military problems, two versus two air combat, strategic battle management, multi-domain operations. The goal is reinforcement learning systems that develop robust strategic behavior through experience without requiring human-labeled training data. And the Clara program is pushing military AI toward logic-based, formally verifiable reasoning, not just capable outputs, but outputs that human commanders can audit, interrogate, and verify. The investment tells you exactly what the people who understand these systems most deeply believe. Pattern recognition under uncertainty is a solved problem. The next challenge is judgment, and those are not the same thing. That last detail matters more than it might sound. It means these systems must discover effective strategy entirely through self-generated experience, the same way a human expert accumulates a lifetime of pattern recognition without anyone labeling their memories for them. Three takeaways. Concrete, usable, let's go. Takeaway one. Your intuition is a computational process and you can engineer it deliberately. Simon's finding that intuition is recognition, combined with the Kahneman Klein specification of when it is valid, gives you an actual design brief for developing reliable expertise. If you want to build genuine intuition in any domain, you need to identify whether that domain is high validity. Do stable patterns exist? Are causal relationships predictable enough that experience can actually learn from them? If yes, you have the first condition. Then you need to engineer rapid, honest, direct feedback loops, not annual reviews, not vague impressions of whether things worked out, specific, prompt, unambiguous information about whether your judgment matched reality. Wargaming is, in this sense, one of the best expert training environments ever accidentally created. Every game ends with a complete reveal. The feedback is perfect. If you want to develop genuine expertise in any domain, ask yourself, how do I make my domain more like this? How do I shorten the loop between judgment and feedback? How do I lift the fog at the end of every game and see where I was actually wrong? Takeaway 2. AI is not alien, and understanding its mechanisms should change how you evaluate it. The systems that outperformed humans at fog of war deduction do so because they were trained in high validity environments with extraordinary feedback loops, the same conditions that produce human expertise, running faster and at larger scale. This means evaluating an AI system in a high-stakes context is not a mystical exercise. It is a practical one. The questions are: what was this system's training environment? Was it high validity? Were the feedback loops clean? Is the pattern domain one where stable regularities exist that can be learned? A system trained on millions of professional replays of a rule-based strategic game has earned the right to make spatial deductions in that domain. A general-purpose language model applied to geopolitical crisis management has not. Ask the Kahneman Klein questions. They cut through the noise. Takeaway three is the real frontier is calibration, knowing exactly where AI belongs and where it does not. We have a clear, research-grounded map now. Spatial deduction, probabilistic inference, belief state tracking in defined environments. AI is trustworthy and in specific domains, superhuman. Put it there. Bring it to the table for those tasks. It will see through the fog more reliably than any human trained for less than 40 years. Political judgment, the capacity to model how ideology, domestic constraint, historical grievance, and genuine moral conviction shape an adversary's actual behavior, that dimension carries information that today's models cannot access. Moral calibration, knowing not just what is strategically optimal, but what costs are permissible, that is irreducibly human, irreducibly situated, irreducibly dependent on values that cannot be trained from historical pattern data alone. And genuine adversarial deliberation, the productive friction of experts who actually disagree and who press each other toward better answers, that is something no language model has yet learned to produce without collapsing into farcical harmony. The veteran wargamer tapping the map and defog GAN generating its prediction are doing the same inferential act. They are both doing it well. But the question of what to do with that information, at what human cost, within what ethical boundaries, through what political process, that conversation still requires people who can carry the weight of being wrong in ways that matter. Pattern recognition is the beginning of strategic judgment. It is not the whole of it. We have covered serious ground today, from a Prussian War Academy in 1824 to DARPA contracts in the present, from a chess grandmaster's subconscious pattern library to a GAN that hallucinates enemy positions with professional grade accuracy, from the Nobel Prize-winning insight that intuition is recognition, to the Stanford Hoover finding that recognition and wisdom are not the same thing. The thread running through all of it is this: human expertise and machine intelligence are not fundamentally different in kind. They are different implementations of the same underlying process, pattern recognition over extensive experience calibrated against reality through honest feedback. Understanding this should permanently change how you talk about intuition, how you develop your own judgment, and how you evaluate the AI tools increasingly being placed in high-stakes positions. The fog of war is not just a war game mechanic, it is the universal condition of every consequential decision. We always know less than we need to. We always have to act on partial information. What separates those who navigate that fog well from those who stumble through it is not magic, it is the quality of their pattern library, the honesty of their feedback loops, and the wisdom to know which patterns apply and which do not. Before I let you go, the academic research underpinning this episode is not publicly available, but the work it builds on absolutely is, and the show notes have direct links to the Kahneman Klein paper, the Deep Mind Alpha Star publication, the Default Gann paper from AAAI, and the Stanford Hoover War Game experiment on Archive. If this episode lit something up for you, those papers are where you go next. If it changed how you think about something, share it with one person who would find it useful. Subscribe if you haven't already. Leave a review if you have a moment, it genuinely helps the show reach the people it's built for. I'm Will. This is Mindcast. The fog lifts, the patterns emerge. The question is always whether you have done the work to recognize them and the wisdom to know what they cannot tell you. See you next time.