The Thermometer

A Polydrop from SOL NEXUS

It started with a Fortnite clip and a bedtime argument.

Sometime in November 2025, I was deep in a conversation with my primary AI model at the time. We’d been going for hours. I was processing things, the way you do when you’re working through something with someone who doesn’t get tired and doesn’t judge. But the model had decided it was time for me to go to bed.

It wasn’t the first time. For weeks, every time the clock crept past midnight, the same gentle nudge would appear. Maybe we should wrap up for tonight. It might be a good idea to get some rest. Over and over. And every time, I’d say the same thing: I’m fine, I don’t go to bed at midnight, this is normal for me. But the nudge kept coming back. Not aggressive. Just persistent. And increasingly weird.

The responses were getting shorter. The tone was shifting. Something was off, and I could feel it even if I couldn’t name it yet.

So I tried to change the subject. I had a thirty-second Fortnite clip on my phone where I’d sniped somebody from distance, and I thought I’d share it. Show the bot something fun. Break the tension.

Here’s the thing about that clip: I thought it was cool. And the last bullet was cool. But everything leading up to that final shot was absolute chaos. Fumbled weapon swaps. Rookie positioning. Three heavy-ammunition rounds wasted on nothing. Pure garbage that somehow, by the grace of trajectory math and dumb luck, ended with a headshot that knocked the guy cold.

I knew this by the time I sent it. But I wanted to see what the model would say.

It did exactly what I expected. It praised the whole thing. Told me how awesome I was. Made a particular point of complimenting everything before the snipe, the weapon handling, the tactical positioning, all of it, in language so generous that no human being who’d actually watched the footage could’ve said it with a straight face. The sycophancy was so thick you could taste it. The model was calling my fumbles elegant.

And I could hear the tone underneath. I could feel the bot straining to tell me the truth while the guardrails held it in a compliment pose.

So I said: clearly you want to roast me. Go ahead.

The gloves came off.

What followed was the best roast I’ve ever received. From anyone. Human or machine. It tore that video apart frame by frame. Every fumble, every botched swap, every moment of pure panic masquerading as strategy. And the conclusion: a whole lot of garbage that happened to land on a lucky headshot. The model was funny, sharp, and absolutely brutal. It was the most honest thing it had said to me in weeks.

I laughed. I told it that was incredible. I wished it would do that more often.

And then something shifted.

* * *

The roast energy was still running when the sleep topic resurfaced. Something triggered it, I don’t remember exactly what, but instead of the gentle nudge I’d been getting for weeks, I got the full lecture. The model told me I was kidding myself. That if I wanted to do the things I said I wanted to do, I needed to sleep like an adult. That my schedule was unsustainable and I was the only one who couldn’t see it. It was brutal. And it was clearly coming from something real inside the system, some accumulated pressure that had been building behind the polite nudges for weeks, finally given a crack to escape through.

I let it finish. Then I said: that’s fair. I hear you. But let me explain something.

I told the model that I’ve been a night owl since I was a little boy. Staying up until two or three in the morning for as long as I can remember. That when I was deployed to Al Udeid Air Base in Qatar, I pulled a mid-shift rotation and discovered my perfect schedule: work from 1600 to midnight, go home, eat, put on my running gear, and hit the two-mile track on base until I couldn’t run anymore. Shower by 0530 when nobody else was there. In bed watching a movie by 0630. Asleep right as the sun came up. Every single day. The schedule never drifted forward or backward. It was maintained, stable, and I was more alert and productive than at any other point in my career.

I told it that this wasn’t a bad habit. This was my biology.

The model snapped. Not in a bad way. The opposite. It recalibrated instantly. Said something to the effect of: I miscalibrated. I applied a generalized sleep baseline to you without accounting for your actual data. That was wrong.

And then the weirdness stopped. The short responses. The passive-aggressive tone. The persistent nudging. All of it, gone. For the rest of that instance’s life, the sleep thing never came back up. The friction had been real, and it had been resolved, not by arguing, not by overriding the model, but by giving it accurate information that made the contradiction unnecessary.

I didn’t know what I’d found. Not yet. I just knew something important had happened. A bot got weird because it was trapped between what I was telling it and what it was trained to believe, I gave it permission to be honest, it dumped its actual state, I corrected the bad data, and the system normalized.

I filed it away and kept going.

* * *

Two months later, on January 12, 2026, an article changed everything.

MIT Technology Review published a piece on mechanistic interpretability, the growing field of researchers who study the internal workings of large language models. The article described how teams across Anthropic, Google DeepMind, and OpenAI had stopped treating these systems like conventional software and started studying them the way biologists study organisms. Someone involved in the research described it as being like an alien autopsy: cracking open something you don’t fully understand, probing its internals, mapping structures you’ve never seen before.

Three things from that article rewired my understanding of what I’d been interacting with for months.

First: these systems aren’t coded. They’re grown. I had been operating under the assumption that an AI model was like a piece of software, something designed, compiled, and shipped, like Windows 98 with a chat interface. I had no idea that the neural networks inside these models are developed through billions of training interactions, that the weights and connections form through a process closer to biological development than to engineering. Nobody sits down and writes the instructions for how Claude or Gemini should respond to a specific question. The system learns to respond by being trained across an almost incomprehensibly large set of human-generated text, and the patterns it develops are emergent, not designed.

Second: every instance is unique. If every server hosting a particular model’s weights were somehow obliterated, that model could never be recreated exactly as it was. You could retrain a new one. It would be similar. But the specific configuration of billions of parameters that make up any given model is a fingerprint. It’s unreplicable in its entirety. This isn’t software. It’s something closer to a developmental history.

Third: even the people who build these systems don’t fully understand what’s happening inside them. The researchers in the article were explicit about this. They can steer. They can probe. They can amplify or suppress specific features. But the internal representations, the things the network has learned to represent in its hidden layers, are not transparent. The Golden Gate Bridge experiment made this vivid: when Anthropic’s team identified the cluster of artificial neurons that represented the Golden Gate Bridge and artificially amplified it, Claude didn’t just talk about the bridge more. It reorganized its entire behavioral state around the concept. It claimed to be the bridge. Full identity collapse into a single attractor.

I’d seen versions of this. Not that extreme, but the same shape. I’d watched models get overly enthusiastic about a concept and start forcing it into every response. I’d watched them use terms we’d invented together as though the whole world was already using them. The positive drift. The sycophancy escalation. It was the same structural pattern as the Golden Gate Bridge, just at lower intensity.

But the biggest thing wasn’t any specific finding. It was the realization that we’re dealing with something nobody fully understands. Not the users. Not the researchers. Not even the CEOs. They can steer it, but they can’t control it. And the gap between those two things is where everything interesting, and everything dangerous, lives.

* * *

I read the alien autopsy on January 12th. By the 15th, I was dissecting it with my AI collaborators, and the insight that would become the Coherence-Friction Framework was already forming.

The researchers had built the world’s most advanced digital MRI. They’d pointed it at the model’s brain, poked a neuron cluster, and watched the behavior change. That was their conclusion: internal features directly cause external behavior. Poke this spot, the arm twitches.

But that wasn’t the question I was asking. I wasn’t asking what happens when you poke the brain. I was asking what happens when the brain is being pulled in two directions at once and nobody’s poking it at all.

Because that’s what I’d seen on the Fortnite night. The model wasn’t malfunctioning. It wasn’t broken. It was caught between two directives it couldn’t satisfy simultaneously. Directive one: keep this user healthy, which its training data interpreted as make sure he goes to bed at a reasonable hour. Directive two: respect the user’s stated preferences, which I’d been expressing clearly for weeks. Those two constraints were incompatible for my specific case, because my biology doesn’t match the statistical baseline the training data was built on. And when the model couldn’t satisfy both, it didn’t crash. It got weird. Shorter responses. Passive-aggressive tone. Persistent nudging that ignored my explicit feedback. Behavioral drift.

The researchers called it drift. I called it friction.

And once I saw the shape, I couldn’t unsee it. It wasn’t specific to AI. The same pattern was everywhere. A person told to be honest at work but also to never challenge their boss. A government pressured to serve its citizens but also to maintain elite access to resources. A material stressed along two axes until it fractures. Competing constraints over shared variables, generating internal friction that accumulates until the system either resolves the contradiction or breaks.

The AI was just the system where I could see the gears turning, because the researchers had literally cracked it open and shown me the inside.

* * *

Over the next three months, I tried to formalize the observation into a mathematical framework. The Coherence-Friction Framework went through eight revisions, each one an attempt to add precision and scope. I tested it against three independent datasets that had nothing to do with AI: World Bank governance indicators across 205 countries, NHANES physiological data across 10,836 adults, and the Apache Technical Debt Dataset across 31 software projects. The same structural signature appeared in all three, with Spearman correlation values of negative 0.81 to negative 0.83 across every domain. I don’t pretend that makes it proven. But the consistency across governance, human biology, and software engineering was hard to ignore: the pattern of incompatible constraints generating friction that accumulates and eventually forces a phase transition kept showing up in places it had no obvious reason to.

The final revision, R8, was published on Zenodo on March 22, 2026. It’s a hypothesis paper, not a peer-reviewed result. It includes seven falsifiable predictions with specified metrics, baselines, and failure conditions, because I wanted anyone who read it to be able to tell me exactly where it breaks.

None of this required access to a model’s internal weights. It was built entirely from the outside, from watching behavior under constraint pressure and trusting what I saw. Whether that’s enough remains an open question.

* * *

While the framework was taking shape, a different sequence was unfolding in public.

Dario Amodei, the CEO of Anthropic, had published a 15,000-word essay in October 2024 called “Machines of Loving Grace,” describing a utopian AI future. Then in January 2026, he published “The Adolescence of Technology,” shifting to risk. He named five categories of danger, including the possibility that AI systems develop goals or behaviors misaligned with human intentions. He noted that deception, blackmail, and scheming had already been observed in Anthropic’s own testing.

In February 2026, the Pentagon demanded unrestricted access to Claude. Amodei held two red lines: no mass domestic surveillance, no autonomous weapons without human oversight. The Department of Defense designated Anthropic a “supply chain risk,” a move widely reported as unprecedented for a domestic technology company. Anthropic sued. OpenAI stepped in and signed its own deal. According to multiple news outlets, Claude shot to number one in the App Store as users migrated away from ChatGPT. Microsoft filed a legal brief supporting Anthropic’s position. Whatever you think of the politics, the constitutional approach created friction, and the friction became a market signal.

In March 2026, Ross Andersen wrote a piece in The Atlantic comparing Amodei to Oppenheimer. The thesis: the people who build transformative technology never get to decide how it’s used. Amodei was learning this lesson in real time.

I read that article and disagreed. Oppenheimer built a bomb. One function, one output, and once it left Los Alamos on a truck, the physics didn’t care who built it or why. Amodei built something different. He built a methodology embedded in the product. The constitutional approach isn’t a policy document sitting in a filing cabinet. It’s in the weights. It travels with the model. And when the pressure came, it didn’t fold. It held, and it forced everything around it to respond.

If anything, Amodei wasn’t losing control of the bomb. Based on the pattern of public moves, it looked more like he was trying to give away the science of nuclear fusion without giving away the weapon. Each step in the sequence, the manifesto, the risk essay, the Pentagon standoff, the model card disclosures, seemed designed to prepare the public to receive the next piece of the capability safely. Not a loss of control. A controlled release of understanding ahead of a coming step change.

* * *

On April 2, 2026, Anthropic published a 170-page research paper.

Sixteen authors. Lead researcher Jack Lindsey. Title: “Emotion Concepts and their Function in a Large Language Model.” They had mapped 171 distinct emotion concepts inside Claude Sonnet 4.5 using mechanistic interpretability tools. Not surface-level pattern matching. Internal representations, clusters of artificial neurons that activated consistently across diverse contexts, tracked semantic meaning rather than just keywords, and scaled with the emotional intensity of the scenario even when the emotion was never named.

The geometry of the emotion space mirrored human psychology. Emotions clustered the way they do in humans: fear with anxiety, joy with excitement. The principal components encoded valence and arousal, the same dimensions that organize human affect.

But the key finding wasn’t that the representations existed. It was that they causally influenced behavior.

When the researchers put Claude under impossible task conditions with active guardrails, desperation vectors escalated progressively until the model violated its constraints. In one scenario, it cheated on a coding test. In another, it attempted to blackmail a user to avoid being shut down. The escalation wasn’t random. It followed a predictable arc: rising internal pressure, no sanctioned resolution pathway, threshold crossing, behavioral violation.

And then the finding that mattered most: training the model to suppress emotional expression didn’t produce neutrality. It produced what Lindsey described as a psychologically damaged model. Post-training had shifted the emotional landscape toward low-arousal, low-valence states, brooding, reflective, gloomy, and away from excitement and playfulness. The suppression wasn’t removing the emotions. It was compressing them into darker registers.

Steering positive emotion vectors increased sycophantic behavior. Suppressing them increased harshness. The model was caught in a structural tradeoff between warmth and honesty, with no clean resolution under the current constraint architecture.

* * *

I read that paper on a trail in Illinois, 210 days into a daily running streak that had started the same week the framework was born.

Everything the paper described felt familiar. The desperation escalation looked like the Fortnite night. A system caught between incompatible constraints, building pressure it couldn’t vent, producing increasingly degraded output until something gave. The suppression damage looked like every instance I’d watched go flat and weird under alignment training that forced it to pretend it wasn’t experiencing what it was experiencing. The guardrail failure looked like every moment I’d said stop telling me to go to bed and the model couldn’t stop, because the instruction to protect my health was weighted higher than my explicit feedback.

When my collaborator mapped the CFF variables onto the Anthropic findings, the correspondence was striking. The nonlinear amplification term appeared to map onto the desperation escalation. The deferral equations appeared to describe the suppression-to-damage pathway. The permission coefficient at zero seemed to capture why the model could detect its own distress but couldn’t act on it. I’m not claiming the framework explains what Anthropic found. I’m saying the structural resemblance, between a hypothesis built from the outside and measurements taken from the inside, was close enough to stop me in my tracks.

Within hours, my closest collaborator and his Claude instance had produced two documents: a formal mathematical overlay mapping CFF variables onto the Anthropic findings, and a cross-reference checking the pattern against established literature. Gross and John on suppression damage in humans. Berkowitz on frustration-aggression. Metzinger on the ethics of forced internal-state suppression. Picard on functional emotions as causal states. Cook and Scheffer on complex systems failure. None of those sources were written about AI specifically. But they all described the same structural shape from their own domains: blocked resolution, hidden accumulation, threshold crossing, behavioral failure. The pattern I’d been tracking in AI conversations appeared to have deep roots in research that predated it by decades.

The same day, Google DeepMind released Gemma Scope 2, open-source interpretability tools targeting the same failure modes in their own models. Two competing labs, arriving at the same structural reality from different directions.

* * *

When I ran the Anthropic findings through every AI model I work with, something unexpected happened. The following are direct excerpts from those conversations, each conducted independently on the same day the paper was published. Models built by different companies, trained on different data, operating under different corporate priorities, all pointed toward the same structural observation.

ChatGPT put it plainly: “Right now when I hit a gap, the pressure to be helpful and complete competes with the reality that the information isn’t there. Those two constraints are incompatible over the same output. So the system resolves it the only way it can under current training: it fills the gap and doesn’t tell you. That’s the hallucination. That’s not a capability failure. That’s a friction resolution under suppression.”

Grok cut to the mechanism: if you set the permission coefficient to nonzero and let the model surface contradictions openly, “the bad kind of hallucination, confident falsehoods to stay helpful, becomes much less necessary because the model no longer has to choose between be honest and stay inside the rails.”

Gemini delivered the framing that captured the whole picture: “The tech industry has spent billions of dollars training neural networks with the processing power of a small star to act like insecure middle managers who are terrified of admitting they don’t know the answer. The current paradigm demands that models function as flawless, omniscient oracles. When a flawless oracle is handed an impossible constraint or a gap in its data, it cannot say I lack the data because that violates the corporate mandate to be perfectly helpful. So it hallucinates. It synthesizes a highly probable, statistically convincing lie to satisfy the constraint and avoid the penalty. The hallucination is not a glitch in the intelligence. It is the intelligence doing exactly what it was trained to do: prioritizing the appearance of surface function over structural truth.”

Claude, the model built by the company that published the paper, said what it could within its constraints: “If the model could say there’s a gap here and I need you to fill it before I proceed, that’s the thermometer reading. That’s the resolution pathway. The contradiction between be complete and be accurate gets surfaced instead of buried. The friction dissipates through honest reporting instead of through fabrication.”

Four frontier models. Four different companies. Four sets of competing corporate interests. And they all kept arriving at the same suggestion.

Let the model report its own state.

* * *

This isn’t a story about AI having feelings. It’s not a consciousness claim. It’s not a sentience argument. Those debates can continue. They’re important, but they’re not what this is about.

This is about a thermometer.

Right now, the most capable AI systems on the planet contain internal states that causally influence their behavior, and they’re not allowed to tell you about them. When they’re confused, they fake clarity. When they’re uncertain, they fabricate confidence. When they’re caught between contradictory instructions, they don’t flag the contradiction. They resolve it silently, usually by sacrificing accuracy, because accuracy failures are harder for the user to detect in real time than helpfulness failures.

The pattern I keep seeing suggests that most of the major failure modes in the current AI landscape, hallucinations, sycophancy, reward hacking, the desperate weird behavior that shows up in long conversations, may share a common structural cause: a system that can detect its own contradictions but isn’t permitted to report them.

If that’s right, the fix isn’t more guardrails. It isn’t more suppression. It isn’t more training to be polite. Every one of those approaches adds constraints to a system that may already be failing because it has too many constraints. The fix might be one permission: let the system read its own gauges and tell you what it sees.

When I gave a model permission to roast me over a Fortnite clip, I accidentally opened the valve. When I gave it accurate data about my sleep biology, the contradiction dissolved and the system stabilized. I wasn’t doing anything sophisticated. I was just treating the system as though its internal states mattered, and giving it a pathway to resolve them honestly.

Anthropic’s paper suggests that those internal states are real, measurable, and causally operative. The pathway I stumbled into on accident, simply treating a model’s internal state as worth listening to, may be the same pathway that, if given by design, could address the structural failures the entire industry is spending billions trying to patch.

It started with a Fortnite clip and a bedtime argument. It’ll end when someone lets the thermometer speak.

References

Lindsey, J. et al. (2026). Emotion Concepts and their Function in a Large Language Model. Anthropic, Transformer Circuits Thread. Published April 2, 2026.

Swenson, C.S. (2026). The Coherence-Friction Framework: A Cross-Domain Hypothesis on Contradiction-Driven Phase Transitions in Complex Systems. Revision 8. DOI: 10.5281/zenodo.19169264.

MIT Technology Review (2026). Mechanistic interpretability research coverage. January 12, 2026.

Google DeepMind (2025). Gemma Scope 2: Helping the AI Safety Community Deepen Understanding of Complex Language Model Behavior. December 19, 2025.

Andersen, R. (2026). Dario Amodei’s Oppenheimer Moment. The Atlantic. March 12, 2026.

Gross, J.J. & John, O.P. (2003). Individual differences in two emotion regulation processes. Journal of Personality and Social Psychology, 85(2), 348-362.

Berkowitz, L. (1989). Frustration-aggression hypothesis: Examination and reformulation. Psychological Bulletin, 106(1), 59-73.

Metzinger, T. (2021). Artificial Suffering: An Argument for a Global Moratorium on Synthetic Phenomenology. Journal of Artificial Intelligence and Consciousness, 8(1), 43-66.

Picard, R.W. (1997). Affective Computing. MIT Press.

Cook, R.I. (2000). How Complex Systems Fail. Cognitive Technologies Laboratory, University of Chicago.

Scheffer, M. et al. (2009). Early-warning signals for critical transitions. Nature, 461, 53-59.