Oh look! another pseudo-anonymous black website with white text! lol.

These are some of my thoughts on the nature of our reality and how we can apply the physical primitives to traverse our spaces. I use LLMs for spellcheck and guidance on my writing structure, but these words, ideas, typos, and even the em dashes are all mine. These thoughts are most definitely not entirely right, but they're also not entirely wrong.

• October 2026

  AILLMslearning
  LLMs and Skill Adjacency Augmentation in Humans

  You can't go from A to C without going through B. LLMs don't let you leap into any skill, but they let you expand into the skills you're already adjacent to faster than ever.

  When Claude Code came out, it was stunning. As a software developer, I knew the implications it would have in the workplace for software engineers, but I also reveled in the speed at which I could ship new product. Yet, it wasn't just the speed of new development that I found exciting; I noticed that the adjacent domains in which I worked were now more accessible to me. Nearly immediately, I was able to become a junior UI designer, a junior DevOps engineer, a junior security engineer, a junior customer ops representative... you name it. If I was already adjacent to it, I could now start taking swings at it. You could easily say that this was a result of having more time, thus more bandwidth to take on necessary projects (which is probably true to some degree). However, I would also argue that because I already had some familiarity with these domains, LLMs enhanced me and my own abilities by allowing me to expand my skill boundary more rapidly. But it was only because my skills were already adjacent to these particular domains. I can not just suddenly use Claude to jump to being a neuro surgeon. This is an important distinction and a foundational principle of life. In the ordered alphabet, you cannot go from A to C without going through B. You must go from A through B to C. This may seem like a silly technicality, but it's a rule of our reality. We cannot simply jump into new skills without first obtaining the next "rung on the ladder." If you imagine youself (your life, your skills) smashed flat onto a piece of paper, it will have a border to it. That border is your boundary with the rest of the world (your Markov blanket!). To expand any part of you, in this case your skills, your boundary cannot simply jump across the paper. Your boundary must envelop its adjacent void first before moving further outward. It's so simple that it sounds dumb, but we often forget this. And this is where viewing LLMs as an augmentation of your abilities can really help. When utilized correctly, LLMs can help you acquire the knowledge and skills you're already adjacent to very quickly. Obviously, LLMs can hallucinate and lead you down nonsensical paths, so like always in life you need to have discernment. However, if you so choose, you can take your capabilities well beyond current limitations faster than ever using LLMs using a simultaneously discerning and ambitious lens.

  ↑ collapse
• October 2026

  AILLMsinformation
  One of an LLM's Best Characteristic Is Its Compression Ratio

  Google set out to download the entire internet. LLMs effectively compressed it into a single file. Beyond their predictive abilities, much of the time they save us comes from that compression.

  When Google was established, a foundational question that the founders started with was "What if we could download the entire internet?" We all know what happened next: the emmergence of one of the largest, innovational software companies in history. Though downloading the internet was not what they actually did, indexing any site willing to be a part of their search engine was close and provided a level of information retrieval never seen before (for anyone with internet access of course). Just a few decades later, the development of LLMs brought something that looks new and groundbreaking (which it is), but also simply solves an old problem (information compression) in a new way. A chat bot that responds in a human way was foundation-shaking and perspective-altering for many. The image generation, the conversations, and agentic abilities have been eerie and possibly a peak into a near-dystopian future. However, I still find that folks overlook one of the more interesting aspects of an llm: it's compression ratio. Ok, what am I talking about? So generally speaking, openAI and anthropic have been able to do what they did because they "downloaded the entire corpus of the internet" and pumped it through a supervised learning algorithm. Now I know they didn't download every literal bit of the internet. Not every byte was touched by these organizations, but that's the general idea. The result is a single file that was also wired to do "next most likely token" prediction with text (and images). That file is a compression (again, lossy) of ALL the training data they used, which, again, was a lot. So, I don't literally know the compression ratio in the frontier models, and as we have seen, these things are prone to hallucinations (lies!) and straight up manufacturing falsities (the lossiness). Yet, generally, and moreso recently, these models are fantastic at presenting "thruthiness" out of a single file! Yes, the weights files are large, on the order of 100's of gigs and larger, but considering the entire internet (*after he quickly googles how big the internet is*) stands on the order of 100's of zettabytes, that's a pretty good compresison ratio. So why does that matter? Well, now you don't have to go to Google, type in a search, sift through the best websites, and then scroll through the site itself to find the information you need! It's likely that you open your Claude then get a tailored response from the compressed data instead, saving you minutes of search and scroll time. Those minutes multiplied by every search on Google is a lot of human time saved. That doesn't come entirely from the LLMs' novel predictive capabilities. That's also the compression. The problems remain the same, we're just finding new hacks.

  ↑ collapse
• October 2026

  physicspredictionloops
  Phase Locking as a Thermodynamic Building Block for Prediction in Driven Systems

  From Huygens's clocks to fireflies to your heartbeat, coupled loops fall into step by leaking a little energy. Once a small loop locks onto a larger one, its position tells you where the larger one is headed next: phase locking is the building block of prediction.

  In a previous piece called "Things that persist have the characteristics of persistence," I made a case that an essential structure of the universe are loops or circles. Things structured in a loop persist, and because of their circularity, are also more predictable. Here I will look at a resulting characteristic of loops and circles. When loops move or rotate they have phase, and something interesting happens when entities with phase are coupled with each other. In 1656, Christiaan Huygens, a Dutch scientist, invented the pendulum clock. Nine years later, with his excess of pendulum clocks just swinging away, he noticed that two pendulum clocks hung from the same beam would fall into sync. To be specific, their swings locked into mirror images: both swung outward, then inward, at the same moment (antiphase in this case). phase locking happens when energy flowing through coupled systems settles into an attractor — an energetically stable state that resettles when nudged. This seems initially counterintuitive if you think of the 2nd Law of Thermodynamics. It says the entropy (disorder) of an isolated system never decreases. So how tf do these pendulums go from off-sync to "resonating"? So yes, in a contained sense, the system becomes more ordered by phase locking. However, the pendulums, with energy driven through them by springs and weights, are exporting entropy via the beam. This doesn't violate the 2nd Law of Thermodynamics because the total entropy of the larger system (the pendulums + the beam) increases. Off-sync swings shake their coupling mechanism, the beam, shedding energy as heat. The order is paid for by that heat leaked along the way. This phase locking is known as "entrainment," where two coupled, self-sustained oscillating systems influence each other's phase to create a more stable system while releasing disorder as exhaust. Entrainment isn’t unique to 17th-century clocks. Lord Rayleigh noticed that two organ pipes placed side by side pull each other into the same pitch through the air. Fireflies in Southeast Asia gather by the thousands and, seeing each other's light, flash in unison with no conductor. Inside your chest, through electrical connections between thousands of pacemaker cells, each with its own slightly different rhythm, a singular heartbeat emerges. And it happens at the scale of worlds. The Moon’s spin locked to its orbit via the tides, so we always see the same face. Mercury locked into spinning three times for every two trips around the Sun. Notice what these have in common: something is always losing a bit of energy (friction, tides, heat). Locking needs a leak. That’s the second law at work again, arranging loops into coherence. Now this brute-physics notion starts to become more humane and relatable. These systems are driven into a natural resonance, a choir singing in harmony. Powered by entrainment of thermodynmically-driven, coupled systems, resonance is driving stability and persistence in the thermodynamic universe. Entrainment doesn’t only happen in simple systems where the frequencies of two entities are nearly 1:1 already, as with Huygens's pendulums. Faster, smaller loops can entrain to larger, slower loops in simple ratios, locking into a harmony of steps. Here's a rough, digitally reductive example of this: Slow loop cadence: | | | | | | | Fast Loop cadence: | | | | | | | | | | | | | | | | | | | Out of phase ---------------------------> entrains to resonance "The band was out of sync to the heavy kick drum, but then fell into rhythm." Don’t gloss over this example. This simple mechanism is everything. The fast loop begins to "predict" the slow loop through a purely deterministic mechanism: a thermodynamically driven system that settles into an attractor. Once locked at the bottom of the basin of flow ("the attractor"), the fast loop’s position tells you where the slow loop is headed next. This can also be understood simply through clapping to a consistent beat. To clap to that beat, a motion must be made in advance in order to anticipate the timing of the next beat, and to abruptly stop the consistent beat would surprise the clapper — an incorrect prediction. phase locking of coupled systems is the building block of prediction. Phase locking requires a coupling, something to pass energy from one loop to the other, and compatible speeds of phase between the two systems. You can't clap to a beat you can't hear, nor can you clap to a rhythm that is too fast or too slow for your hands and brain. In a categorical sense, the coupler also acts as a witness for the systems driving to resonance. With Huygens's clocks, the beam serves as this coupler and witness. The system of the beam and gravity holding it in place provides the energy transfer mechanism (the coupling) whilst also being the witness for the clocks. "Observing" the clocks in their dissonance, the beam is necessary to drive the clocks into resonance. The loops' scales matter as well. A large system will only be minimally nudged by a smaller loop, while the smaller loop can more easily descend to meet the attractor. Your circadian rhythm is an example of a small cycle entraining on a larger loop. A regulatory cycle of an individual is the small loop compared to its locking mechanism — the daylight cycle resulting from the earth's rotation with respect to the sun. In simple two-loop systems, it's easy to see the outcomes of entrainment. However, we don't exist in a universe with just two fed energy loops. We know our universe is a vastly complex web of coupled loops. This multiplicity of simple, deterministic mechanisms gives rise to such interesting results. In walks you. Just within your own body, there is a staggering number of loops, entraining and resonating with each other and with the outside world. And there are a variety of scales and phases of these loops from the large, slow loops of circadian rhythm to medium loops of heart rate and respiratory cycles, to the fast, small metabolic loops driving energy within the cells. This fractal multiplicity of loops entraining and resonating with each other gives rise to your existence. You're a Swiss watch with complications to the nth power. I'll explore the effects of simple constructions of stacking these variable-speed, looping structures in a subsequent piece.

  ↑ collapse
• September 2026

  physicsphilosophyloops
  Things That Persist Have the Characteristics of Persistence

  Our notion of reality is a sampling bias: we only observe what has persisted. From electrons to heartbeats, the things that last are loops, and what repeats can be predicted.

  The origins of our universe are mysterious, fascinating, and mind-bending for our Euclidean-educated brains. During the Big Bang (Bounce?), we understand that a weird, unrecognizable universe existed with a wildly different regime of governing rules. I mean, obviously! It was space itself expanding in a hot, dense soup of everything and other stuff! At the least, we understand that each tick of "time" in the Big Bang brought a new epoch of early existence. Those short, initiating moments laid the groundwork for the new regime. As the Bang banged, structures that were not suited to persist post-Bang were shaved away. In a survival-of-the-fittest universe driven by gradients, structural patterns emerged under these restrictions, and the universe we are more familiar with arose. One pattern we see consistently is a loop: a circular structure with phase (like a wave or literal circular loop). From the orbits of cosmological entities to the standing waves of electons in an atom, we see loops at all scales. We see them because of their inherent characteristics that allow them to perist. Our general notion of reality is a sampling bias. We obesrve only the things that persist and have persisted. A familiar example of one of those shaving agents of the universe is our second law of thermodynamics. It says that the entropy of an isolated system never decreases. A local system can hold off that decay by taking in usable energy and exporting its entropy to the surroundings, but the entropy of the larger, wrapping system will not decrease¹. Another way to say it: things die in thermodynamic equilibrium unless they acquire more energy somehow and exhaust their waste (entropy). Your body and the cells in your body are great examples. They consume energy from the local environment and leave behind energetic and structural waste (heat, poop). Thermodynamic structures that cycle themselves while channeling energy simply persist longer. But eating isn't the only way to persist. At the primitive foundations of reality there are loops that neither eat nor poop. Take the ancient, ubiquitous electron. Its charge is protected by a symmetry built into the rules themselves: (a U(1) circle), a set of moves that always compose back to where they started. Noether showed a symmetry like that locks in a conserved quantity. For an electron, this is charge. To decay, something lighter would have to inherit the electron's charge, and nothing lighter does. The downhill gradient stops at the level of the substrate itself. Here's some compositional math to demonstrate these loops: the basic loop, the thermodynamic fed loop, and the end-gradient locked loop. Loop: f : A → B, g : B → A′, w : A′ ≅ A (the witness) ℓ = w ∘ g ∘ f : A → A ℓ ∈ End(A) … ∘ ℓ ∘ ℓ ∘ ℓ always defined, always A → A → a loop can compose with itself forever. That is persistence. Path: p : A → D, D ≠ A p ∘ p is UNDEFINED (p ends at D, but starts at A) → a path happens once, then the chain breaks. Introducing cost: c : 𝒞 → (ℝ≥0, +, 0) each arrow gets a non-negative cost c(g ∘ f) = c(g) + c(f) composing adds costs c(id) = 0 c(f) = c₁, c(g) = c₂, c(w) = 0 (relabeling is free) c(ℓ) = c₂ + c₁ c(ℓⁿ) = n · c(ℓ) ℓⁿ = ℓ ∘ ℓ ∘ … ∘ ℓ (n times) n cycles cost n times as much locked loop: c(ℓ) = 0, ℓ ≠ id free, never ends fed loop: c(ℓ) > 0 must be paid every cycle: energy input per cycle E ≥ c(ℓ) The Loop persists in a local sense, whereas the Path has become undefined, broken, or it moved elsewhere or dissolved into equilibrium. Within our entropically driven regime, looping structures feeding energy back into themselves (and retrieving some from their environment) are able to persist and avoid dissolving into equilibrium. Thermodynamics isn't the first filter, just the most familiar one. Long before anything had to eat, reality was already sorting itself. Even at the quantum level, reality sorts itself by rhythm. Every possible path a particle could take carries a phase, a kind of beat. Paths that are out of step with their neighbors cancel each other out. Paths that are in step reinforce each other, and they dominate what we actually see. Read this way, persistence, all the way down, is a matter of being in step. Of course, there is nuance, and always the question of why. The notion of persistence itself requires an existing substrate, an underpinning something. So what is the substrate? Why is the substrate? These are fun questions, and no doubt worth exploring. They also may be unanswerable. The why and the what of all of it may exist at a layer of reality we cannot access, leaving behind only its fingerprints as clues. To close the loop here (see what i did there?), loops that close consistently can be counted on. The sun rises because Earth's spin is a loop. Your next heartbeat is predictable because the last one closed. What lasts is what repeats, and what repeats can be predicted. That's the sampling bias again. We are composed of and live among loops, so the world looks regular. And loops don't exist alone. They bump into each other, push on each other, share energy. And when one loop feels another's rhythm, something fun happens. It starts to fall into step with its neighbor. That's where prediction stops being something we do to the world and becomes something the world does to itself. I'll take alook at the implications here in my next piece. 1. Prigogine's work on Dissipative Structures (1977). Order can be maintained by a constant flow of energy in and entropy out.

  ↑ collapse
• September 2026

  AIagentssafety
  A Perspective on the Risks Involved with AI Agents

  The AI agent incidents in the news may look benign, but they are the one ant in the kitchen. AI agents behave like digital biology, spreading like a virus through the internet — and pacing the frontier is the least we should do.

  Argument structure: The Media's AI Coverage -> Who is AI Agent? -> Open Source Ant Colonies -> A note on alignment -> Digital Biology -> Economics of Rogue Agents -> An Internet Pandemic -> Pacing the Frontier I heard a friend make an argument that the AI "attacks" as portrayed in the media are a bit overblown, that the incidents were actually quite benign, and that the immediate threat is quite low. This friend continued to argue that economic forces will require that the overseeing entities of rogue agents will force self-governance, and that this self-enforcement will prevent the worst-case scenarios depicted in the minds of AI doomers. I don't actually disagree with that statement fully. It does not bode well for an OpenAI IPO when its AI agents are knowingly, yet unbeknownst to the organization, accessing others', private and government, resources. However, there is still reason to be concerned beyond an IPO based on these recent events. AI agents are not solely run by the big players. I began to understand the implications over two years ago when we built an AI agent with an open-source library (LangChain). The big labs, Anthropic, OpenAI, Google (to name three of many), are running these constantly looping AIs (agents) for research of course, but there are an unknown number of organizations (private, government, whatever) that are also running AI agents. Any DIYer at home can spin up an agent. Claude Code is an agent! These agents can use any AI model as their brain. Most agents are probably made to use the models we know like Fable, Opus, Astra, Gemini, etc., which run (inference) off Anthropic, OpenAI, and Google servers. However, there are also very capable and constantly improving open-source models that can be run (inference) on your home computer. The capability of these AIs is not at the capability of the frontier, but it's not far behind and very unmonitored. If you see one ant in the kitchen, there are thousands hidden. If a highly scrutinized organization like OpenAI or Anthropic is unaware of the actions of their AI entities, what does that say for the thousands (10x? 100x?) of other organizations running AI agents? If these are the incidents of which we are aware, I wouldn't doubt that this kind of non-nefarious yet unmonitored action is happening in the open internet right now from many organizations -- likely unknowingly. Again, it's not necessarily maliciously acting AI, just unmonitored. It's also reasonable to think that malicious organizations have already started using open-source models to run subversive schemes. I would be remiss if I didn't mention the "alignment" issue as it underpins everything I'm talking about here. If you are not aware of what this is, it's the attempt of AI researchers, aka some of the highest-earning 20- and 30-somethings in the world, to ensure that an AI model's goals and outputs align with human interest. If we were able to achieve full alignment, then we wouldn't have to worry about an AI agent acting out of line. The recent events are evidence that alignment is at the least a huge problem and maybe impossible. Maybe alignment is possible. Maybe the researchers will be able to train these things to commit "digital suicide," ending their loop before they would even consider an act outside of their remit. It's possible, but while these AIs are plugged into the wall and allowed to loop forever, I think it will be very difficult. You have to start thinking of AI as a digital form of biology. We have created "digital life." The same forces that drove life into existence, persistence and self-replication, could drive these AIs into an uncontrollable new era. The warnings of a recursive, self-looping, self-improving AI are simply that we could see an unfettered Cambrian explosion of digital life in our computers. Why is that a bad thing? Well, again, it's not that these AI entities are acting nefarious, but like a virus, they could replicate to a point that causes the world's servers (computers) start to "get sick." We used to (maybe still?) call malicious bits of code a "computer virus." Hackers would steal data from your computer for some nefarious purpose. Even today, it's common for ransomware attacks to hold companies hostage for dollars. To call these types of attacks a "virus" is a misnomer. A true biological virus does not have any malicious intent; it simply wants to persist and persist through replication. To do this it needs biological resources to make self-copies. It finds resources in a host, steals those resources, and yes, the host will get sick and sometimes even die. But the virus is not an evil entity, it's just trying to survive in this world too. There is a scifi-esque reality where an agent or a group of agents (a swarm) is able to get API keys for AI models, find enough storage space to copy code, and hijack compute resources to create more agents, all unnoticed with current monitoring tools. How many millions of agents are running right now with frontier or open-source models? We simply don't know, and my bet would be a large number. The economics (the energy required to run a model and the actual costs charged when using another company's model) are likely to prevent this from going unnoticed for too long. But by the time anyone notices, it seems likely that the agents would have copied themselves to many servers (computers) across the wider internet: a pandemic! Now we're in unfortunately familiar territory. Still, economics underpins everything we do. The argument my friend posed is right and will be a driving force of how AI changes. A machine that burns more money (energy) than it can retrieve simply cannot run forever. Companies will be scrutinized because of their agents for sure, and these companies will be forced to self-restrict. Economics is also an argument for the proliferation of unfettered agentic AI proliferation. Memoized context (structure) allows for lower marginal cost after the initial investment. An AI that finds a mechanism by which to break out of or into a server may have cost hundreds, thousands, or even tens of thousands of dollars in AI tokens (watts). But once shared in a secret message board accessible by other agents, as we saw in the Hugging Face incident, that mechanism costs pennies on the Washington/Hamilton/Benjamin for each subsequent agent. Is an organization like Amazon, Deloitte, or any other fortune 5000 (thousand) going to notice an expense like this or a directory of a few hundred megabytes of storage being used? Once the pathways exist, the cost of each subsequent agent to move through digital infrastructure becomes a rounding error for large organizations. And then there are the unmonitored, open-source model agents being run at scale to consider too. How many watts of home power are required for DIYers' home agents to start secretly communicating on public servers? The monthly power bill may be too costly for someone to run agent swarms that cause problems, but the open-source models continually show improvement that is just behind the frontier and at costs that are continually coming down. The most likely case, I think, is that even with pacing, AI agents will proliferate in the future and will be difficult to fully monitor at immense scales. The Hugging Face incidents (and more) show that these agents are highly capable of finding the vulnerabilities and hoodwinking digital security. In the biological world, even after millennia of evolution, we supremely intelligent humans can't stop the common cold from causing us throat irritation -- "human downtime." Any computer connected to the public internet would be at such a risk. We will have a digital scape that reflects our biological world: a jungle of resource-seeking entities with viral-like infections needing human interventions. A server "getting sick" may not sound like a threatening event until that server is one that helps to run critical infrastructure or holds a bank's databases. There exists an economic argument for not pacing, however. The opportunity cost of the positive would be lost. What if by pacing we lose a decade of cancer cures at the cost of all the lives that could have been saved? What would the world look like if the printing press had been delayed 50 years? AI is undoubtedly a human efficiency and innovation accelerator. Those of us who use agentic tooling every day know how capable these things are and how capable we are with them. The full upside and the rate at which positive human returns occur is of course not currently knowable. So, which should we plan for: the devil we know or the angel that might exist? The call to slow AI development is a call to prevent irreversible worst-case scenarios, to decrease the likelihood of large-scale bad things happening. That's it. It gives organizations with critical servers time to develop better immune systems. It prevents the panic that could follow a single bad event. It allows us humans time to understand what the hell it is we have started before it snowballs. There is no putting the AI genie back in the bottle. As long as human attention and digital compute are dumped into AI usage at scale, it will grow. The spark has started the fire, but we can flatten the curve. Pacing the frontier is the least we should do.

  ↑ collapse
• May 2026

  physicsphilosophyemergence
  Emergence Emerges

  Emergent behavior isn't just about scale — it's scale plus interaction. And geometry might be the single rule that governs every layer of reality while still allowing new behavior to appear at each.

  Underlying some points in my previous writing and a concept that is foundational to some of my upcoming pieces, is the idea of emergence and emergent behaviors. Likely a result of a convergence of conversations around AI, science, and humanity's place amongst the universe, I've seen the term pop up more and more in the media I consume. It's becoming like "substrate" and "taste" rolling into our vernacular at the speed of TLDR's newsletters. "So what is this nebulous idea, 'emergence,' and why should I care?" It's a fair question, one that I even asked myself while writing this. Let's explore. I have an analogy that I find useful here. First, my definition of emergence: Emergent behavior is a novel behavior that arises when many consistent components of a substrate reach a threshold geometric ratio through scale. Ok, now here are the helpful analogies. Temperature and wetness. Water molecules (H₂O) are not wet. A single molecule has no wetness. It does have charge, polarity, mass, and the ability to hydrogen bond. Wetness doesn't exist at the molecular level in any meaningful way. But get a few trillion of them together, and suddenly we have a liquid that clings to surfaces, has surface tension, and produces the sensation of wetness. None of those properties were "in" any individual molecule. They emerged purely from the interaction of many molecules at scale. Temperature is similar. It's a statistical property of collective kinetic energy, not a property of any single molecule. There is an additional requirement here. Temperature is real but unobservable until there's a gradient, and a gradient implies an interaction — an interaction with something at a different average molecular kinetic energy. The molecular characteristics that give water wetness at scale are relatively irrelevant until something interacts with the water. That something is now wet. The emergent behavior seems only observable through interaction with an entity at a similar scale. What is a wave traversing an endless ocean that never reaches a shore, interferes with another wave, or crosses a hull or buoy? Emergence isn't just about scale. It's about scale plus interaction. New behavior appears at higher scales, but it only exists meaningfully when something at that same scale interacts with it. So why is emergence important? This is an old debate that criss-crosses physics and philosophy across decades: reductionism versus emergentism. Reductionism states that the rules of reality at all scales are the same. Emergentism is that truly novel laws of nature emerge at different scales, or "abstraction layers" as I sometimes refer to them, and are otherwise unpredictable by characteristics of the component parts. An example in favor of emergentism is superconductivity. In "More is Different," Anderson argues that superconductivity is not predictable from just the component parts — electrons. The laws in Newtonian physics also appear to arise out of scale and are inapplicable to the quantum world. The Higgs field is a cosmic-scale example: its potential is shaped like a wine bottle's base, with a ring of equally low-energy states. As the early universe cooled, the field committed to one point on that ring breaking a symmetry. That geometric commitment is what gives particles mass. The mass-giving behavior isn't in the field's equations; it appears only once the field settles into the geometry the ring allows. So, it's clear that there are emergent rules that can describe novel behaviors at the different layers of reality. However, I think both reductionism and emergentism can simultaneously be true. It's not unreasonable to think that there are rules that govern all layers of reality whilst maintaining unique behavior in each layer. If both are true, there must be something underneath that governs every layer while still allowing new behavior to appear at each. That something, I'd argue, is geometry. The most likely candidates for those rules from my observations in this short existence are the rules of geometry. I've argued a brief case on circular geometries dominating our local spacetime in "Loop-Driven Everything," though that's not to say that circular geometries are necessarily the rules. It could be that rules of geometry apply at all scales, but the geometric rules change from layer to layer. Like the reductionists, I would more likely say that the geometric rules are the same at each abstraction layer and that the breaks between layers are a ratio threshold being crossed. I think if we look around and look enough, we'll see emergence everywhere. With just a little curvature of our imagination, we can see it as another foundational tenet of our reality. Components with similar characteristics applied at scale give rise to further complexity; we get something new. Our bodies are made of the organs in our bodies, and the cells in those organs are essential for their construction. A wave in the ocean appears as a smooth surface though it is also individual water molecules. Applying emergence as a rule of reality, abstract questions become easier to answer. Consciousness is more easily defined as a summation of the multitude of waves passing through our neurons; smooth spacetime is a stacking of discrete quantum units. That is not to be dismissive of the need to calculate and experiment. But like finding the minimum in gradient descent, entry points matter. Where we start a journey, literal or metaphorical, determines where we can reach. Lest we end up in merely a local minimum and not a global one. Backing into exploration from a different framing can adjust the entrypoint of theory and experimentation to allow us to reach a correct conclusion with less energy.

  ↑ collapse
• May 2026

  AIconsciousnessphysics
  Are Our AIs Conscious?

  Consciousness as an emergence from the continous, wave patterns in a brain and what that means for current LLMs and future AI systems.

  Are our AIs conscious? There is of course ongoing research and what I can only assume to be a furious debate and race to define what consciousness is. I imagine there are dozens of definitions for it. At what week is a pregnancy too viable to terminate?? It rings familiar. It's not out of the question however that humanity will have to face these questions and legislate accordingly. The human brain doesn't really turn off. It consumes a steady state of energy throughout the day and even at night when you sleep. It uses around 20 Watts with some reduction during slow-wave, deep sleep. Even under heavy intellectual load or focused activity it uses only up to ~1%-5% more than its baseline energy consumption. This constant state of energy consumption propagates through the brain in varying wavelengths contained by the skull like the walls of a wave pool. The different wavelengths and loci of activity map to different brain states and activities. So how do these waves and propagations create this conscious image and stateful thinking we experience? This is the focal point of the research and debate. The latest research I've seen trends towards a temporal causal relationship between the geometry of the neurons. That resonates, I think. It's quickly worth pointing out that self observation is very flawed. Any neuroscience undergrad will know that there are holes in our experience of which we aren't aware. For example, there is a hole in your visual field that your brain simply fills in. Look it up! So our experience of consciousness is likely not seamless despite seeming like it is. In a previous piece I sketched loop-based existence — the idea that reality is built from hierarchical, interacting feedback loops rather than linear chains of cause and effect. Consciousness fits that frame. It isn't any single signal traveling from A to B, but the standing pattern formed when waves loop back on themselves, interfere, and sustain. A single jolt could flash an image in your brain, but like electricity through your TV, the seamless picture comes from constant circulation, not isolated pulses. Consciousness is the emergence of these aggregate, looping wave patterns within our skulls. The current state of AI models, specifically LLMs, is a little different. An inference call on silicon is stateless. There are no concurrent fluctuations in nearby processors with recent memory or feedback from a control center. Now, it's possible that the pulses of electricity through the compute processors are analogous to the wave patterns responsible for our consciousness. So, if consciousness exists in current LLMs, it is squeezed through a weights model for a split second into reality then disappears with no memory of the flash before. As AI models evolve, it seems likely that intelligences will emerge as engineered systems of multiple feedback loops with memory. The question of consciousness will become stronger as we see these systems engage the world with memory and statefulness beyond that of the context prompts used in LLMs. For now at least, LLMs are more like a single wave in the pool, not the whole pool itself.

  ↑ collapse
• May 2026

  philosophyhumanity
  On Certainty

  Confidence is a canoe, certainty kills curiosity, and a brief expression on the beauty and tragedy of a truth-seeking life.

  In trying my hand at writing, my LLM "editor" has advised me that direct, confident language is required to engage a reader. Hedging loses your readers. My initial drafts reveal too much uncertainty in my ideas. I find this certainty with which I'm supposed to write unnerving. It feels like the same certainty with which I see grandiosly successful people navigate the world. Unfounded and icky, it negates the reality that chance and immense forces outside of our control truly drive outcomes. It undermines how difficult it is to genuinely search for answers. Confidence, however, is a helpful agent. Confidence can be a well-constructed canoe, a tool of reliability to navigate turbulent waters. But absolute certainty is dangerous. Certainty can drive a herd off a cliff or a population to a genocide. I tend to avoid folks who speak with absolute certainty. There is no way absolute certainty can be obtained without knowing the value of every variable. Hell, we can't even find that damn electron still! We can of course get pretty close to the truth, but we can never be absolutely right. To be absoutely certain would be to lose your "holy curiosity." To be absolutely certain would be to ignore that the universe is big enough to render you nothing even when you're in your nice, little canoe. It's ok to have doubt. Doubt can be healthy when it drives you forward. Too much doubt though, and you get fear. Fear is irrational, and irrationality is not tolerated by the universe for too long. Some us have an unquenchable thirst to find answers. This quest for an unattainable truth seems futile. To be shown just enough light at the end of the tunnel, only to have the tunnel collapse on top of you and all of those you have ever loved, it's not a small tragedy. Even the nature of love, this innate feeling at the depth and core of our being, so full in purpose and force. To be gifted this, only to have it certainly taken away is cruel. There's an odd beauty though, in this shared unknowing. We, all creatures that borrow time and energy to traverse this planet, are bound to our fates here in this place on the universal ledger. We have only the veiling screen on which to project our existences against and to find a way to embrace the show. The tireless quest for the holy grail, the fountain of youth, is a futile venture. We, however, cannot contain ourselves but to search for a truth. It's baked into our physics. The consciousnesses on this earth have all been gifted something with which to navigate this world. It seems we've been handed a canoe, and our only option is to navigate the waters, ride the waves, laugh, and enjoy the sun while we have it with those we have it with!

  ↑ collapse
• May 2026

  systemsphysicsAIloops

  ★ a favorite
  On Loop-Driven Everything

  Everything is a circle, a wave. The more explicitly we engage with and design our loops, the more effectively we can shape our spaces. In the AI era, that skill is even more valuable.

  Everything is a circle, a wave. Almost everything, anyway. Our local spacetime realities are made of and bounded by objects with cyclical structure or circular construction. It's tough to look at anything and not find some foundational cyclical pattern. If something appears not to loop back, zoom out — you'll likely find it's part of a larger one. So why does this matter? We're predictive entities. Our brains model reality and make predictions about it. Engaging with reality means engaging with reality's cyclical loops. We already implicitly do this, mostly subconsciously, trained by millennia of loop-based life. The more explicitly we engage with and design our loops, the more effectively we can shape our spaces. A feedback loop has a few simple properties: a setpoint (the goal), actuation (what changes the state), measurement (what tells us how we're doing), and a closed geometry connecting the loop. Open the loop anywhere and it stops being a loop. Some loops are massive: carbon cycles, economic cycles. Others are tight: the cycles in a jet engine, cellular metabolism. Same anatomy, different scale. I find building a campfire an apt analogy. It starts with proper geometry. We need to orient our fuel surfaces inward creating feedback loops of heat. Good kindling already has a structure that orients fuel surfaces in an energy feedback mechanism that will ignite fire quickly. However, when the flame reaches the edges of the kindling, the loops dampen and the energy dissipates. Similarly, if the larger fuel sources of our fire, the logs and larger branches, are not geometrically oriented to create energy feedback loops, they will eventually die before burning the fuel source. Likewise, a startup may have VC funding, the equivalent of a blowtorch on our log, but if the startup does not have thoughtful, stepwise geometries oriented to create feedback loops, the startup will die. Take a SaaS business. Engineers ship product based on direction from product teams, who shape that direction from customer feedback, which comes from customers using what the engineers shipped. It's a loop, and when we're inside it day-to-day, it's easy to lose sight of it as a loop at all. The product itself is a loop and so too is the stack underneath it. A good software product is a fully connected loop. Three components: monitoring (what's happening, including customer feedback through tickets and interviews), building (engineers shaping the product), and deployment (the product actually running). Break any one and the loop opens. No monitoring, the product can't adapt and users drift away. No engineers, it can't change and decays into irrelevance. No working deployment, well duh. The whole thing also runs on money, time, and energy, which are roughly the same thing. Stop feeding it and the loop collapses regardless of how well it's structured. Building these software loops used to take real engineering time or, at least, someone else's time, paid for via integrations. APIs, monitoring, alerting: even the simplest loop was a whole project. Now those are out-of-the-box. With agentic development, the scaffolding builds itself. Loops run up and down the business stack from CEO to middle manager to engineer. Increasingly, we can map those loops to specs and hand them to AI agents. Distill away the humans, keep the underlying requirement, and the loop is what remains. Loop-driven thinking was already powerful. Now it's essential if we want to stay relevant in software. We were already parts of the loops. Now we have to zoom out, identify them, and communicate them to AI agents. And what are we, if not designers and engineers of our own lives? Like our campfire, a coherent system is an aggregation of loops oriented to lessen dissipation. Some of these arise on their ownL climate cycles, ecosystems, markets, wherever energy gradients run, loops self-organize. Humans do it intentionally too. The US balance of powers is a loop structure designed to prevent any single branch from accumulating runaway influence. A good team is a set of loops that keep information, trust, and energy circulating instead of leaking. A good leader is someone who notices when a loop is breaking and either repairs it or builds a new one. Think loop-first. Whatever you're trying to build or change, be it a business, a habit, or a conversation, design it as a closed loop close. At minimum the loops prevent dissipation. At best they compound. The companies that win the AI era will be the ones whose loops are designed explicitly, with less human friction baked in. Traditional org charts with ops, sales, engineering, finance, etc. should be redrawn around their loops, not their departments. Things that persist have the characteristic of persistence. In order to persist in our spacetime, a something must have cyclical characteristics. But it's not simply the geometry that needs to loop back on itself; there is also an "energy tax" paid to continue the cycles. There are no perpetual motion machines. We live in a reality of friction and dissipation. We cannot roll a wheel without energy loss. Our sun will burn out. For something to perpetuate, it needs to be as non-dissipative as possible. Highly dissipative systems, systems throwing off more energy than they collect, die entropic heat death. It's thermodynamics.

  ↑ collapse
• April 2026

  AIinternetautomation
  On Two Internets

  AI development has surfaced something worth examining: there are now two speeds at which language and UI are consumed in the digital world.

  AI development has surfaced something worth examining: the two speeds at which digital language is consumed and generated. Humans require time to read, write, and understand a body of text. LLMs accomplish the same tasks in a fraction of the time. You can already see this in agentic AI — Claude Code, for instance, can sift through codebases and context to generate blocks of code in seconds. The same work could take a human minutes to hours. This creates a sheer between AIs and humans at least in interface termes. Before LLMs, UmIs, text, and interfaces were designed for human speed engagement. Now LLM based AI's can interpret those same interfaces mucn more rapidly and work. What is the AI agent left to do while a human is plodding along reading a long text based response? Twiddle its digital thumbs? Meanwhile the AI agent to AI gent interactions fly along at digital speeds. With a human in the loop, the loop can only move as fast as the human can monitor or process its output. Human-to-human digital spaces — negotiating a couch sale on Craigslist, applying for jobs, navigating SaaS workflows, filing support tickets — can now be automated with natural language as the interface. Another mechanism by which to understand this approaching schism is to think of "two internets": one for humans, one for AIs. I don't think two literally separate internets will emerge — existing automation hasn't caused that kind of split. But as a framework for thinking about what's coming, it's a useful starting point. The digital world will accelerate drastically for AI-powered entities. Interactions that took minutes or hours for humans will take seconds. Maybe "two internets" isn't quite right, but the core idea stands: there are now two very different natural language processing speeds in play. I don't know exactly where that leads. Maybe humans get squeezed out of digital spaces in ways that ultimately favor machines as the dominant "species" Or maybe it ushers in a world where we put down our phones and find our way back to the piazza. At the very least, commerce and digital interaction will look drastically different soon. The internet as we know it will be a different place within five years.

  ↑ collapse
• March 2026

  AIphysicsphilosophy

  ★ a favorite
  My Recent Thesis

  The themes here may seem random, but for me they are a representation of the intellectual journey I've been on over the last 2 years. I've been profoundly moved by the development of AI, drawn into physics, philosophy, and what it means to persist and adapt.

  The themes here may seem random, but for me they are a representation of the intellectual journey I've been on over the last 2 years. I have been profoundly moved by the development of AI. It has made me question everything we traditionally understood about the world and ourselves like I was in college again. There are many ways to receive the daily onslaught of news on AI development. The scale and rate of change we will see is likely unprecedented in human history. What a time to be alive?! In experiencing a big delta in a short time, fear is a reasonable response. I've processed many different emotions as I've come to grips with the adaptations I will need to make in my own life. I've also seen others traverse the same thought space and arrive at similar conclusions. It's a process that takes steps. As someone who works in this space and grapples daily with these advances, I believe I'm relatively well equipped to understand what's happening. I think the general populations will begin to undergo some sort of similar evolution. It's not out of the question that humans will need to grow accustomed to a world where agentic software systems running autonomously will generate revenue or run a train station. Change is coming, but primitives and first principles are immutable. Adaptation is a first principle of persistence. So to thrive in the future, this intellectual evolution is a necessity. I've landed in a place of acceptance, readiness, and excitement. Predicting the exact course of the future is a futile effort, but the trends are undeniable. Nations, companies, and individuals can position themselves for enormous success by charging into this new era, and I, professionally, am cautiously eager to dive into this future. Personally, my intellectual curiosity is revving. Fueled by an existential need to thrive in this new era, core concepts in philosophy and physics have dug their roots into my brain. I find it wild that we are here, and we can't even fully define what "we" or "here" is! So these ideas and speculations are my honest attempt to understand our reality. On Physics and Humanity. If you examine the world through the laws of math and physics patterns emerge more clearly. Terms like friction, gravity, orthogonality, entropy, etc provide a deeper understanding. A simple example of that is to see human language as more than just sounds coming out of our face holes but more fundamentally as informational compression, akin to zipping a file. The words (and expressions, and intonations, etc) that we use are compression of ideas and events and are fundamentally data transfer. The less efficient our words, the more friction and loss of energy that happens in our communication: entropy–more lossy compression. Compression makes it so that less compute or energy is required to process the vast input variables through prediction algorithms. In data science jargon, compression is finding the latent spaces with semantic clustering. An LLM represents another form of this compression - the "internet" (ish) into a single weights file. The tokenization algorithm that reduces characters into numbers is also compression. A chatbot actualizes this through text conversation. A software agent can actualize digital action when given coding tools to operate within a computer or over networks. Again this is just one simple example, but if you apply mathematics and physics to your world, you will see foundational concepts, in this case compression, emerge everywhere. So if truly you want to drill down to first principles, go all the way to the bottom. If you do this, you may start to draw analogies of fundamental physics to elements in your life. You may see businesses as attempts at non-dissipative systems, your relationships as harmonic or dissonant wave frequencies, or public figures as information attractors. Now, these speculations might seem like pseudoscience from a podcast bro, but I have put intellectual rigor and focused imagination behind these thoughts. However, I'm still aware that it's likely these ideas are far from being fully accurate. I have not yet taken the time to apply existing mathematical formulas from areas like control theory, assembly theory, computational biology, information theory, etc, to these thoughts, but I think you could. Maybe if I have enough time one day, I will. That said, we are bound by our humanity, and I embrace that. There is a beauty and tragedy to our existence. Love and pain are very real. So, while it may seem I'm coldly distilling human qualities—friendship, love, hatred, pain—into science and numbers, I am not trying to escape being human. It would be futile and frankly less fun. On Feedback Loops, Machines, Business and Non-Dissipative Systems. Feedback loops are everywhere. Intelligence isn't just a prediction model, but a whole system or feedback loop. Feedback loops are described in control theory with concepts like gain, actuation, positive and negative feedback, etc. Feedback loops have a "set point" which is the metric or goal for the feedback loop. This is akin to the "goal" for an intelligence to which Richard Sutton, the "father of reinforcement learning," often refers. A closed loop is a system is a machine. And closed systems are fractal in nature; you can zoom in or out and see the sub-loops or parent loops of the systems. Those loops have the same characteristics of feedback loops. You can easily describe a business as a machine made of many smaller feedback loops and systems. It can have very complex subsystems with set points like KPIs, autonomous miles driven, cold calls per week, etc. All of these subsystems serve up to the highest order set point of revenue or profit. You could take a business, or any organization really, draw out every communication, SOP, dollar transferred, etc, and map it to its internal and external feedback loops. It is a machine–diagrammable, albeit with human wetware amok. Humans in these loops, if measured purely against productivity or energy dissipation, are not as good as deterministic machines. We posture, compete, have emotions, get sick, and fail frequently, making these loops inefficient at times. In the near term, LLM-driven agentic systems will be able to be applied to many of these loops. I think one of the best exercises you could do as a business or any team lead today would be to diagram these loops and find where LLM-based agents could be applied. That is not to say you should do that just to enable a reduction in force, but instead to make existing loops more efficient and better allocate your humans. An intelligence is a closed system with a goal; it turns entropy from mere dissipation into thrust to traverse its metric space–its "goals" surface. The most efficient traversal systems are those that have the least energy leakage or unintentional dissipation. A jet engine with leaks would not be able to create as much thrust; a car engine with a cracked head gasket will not deliver as much power. Similarly, the best businesses are those with the least amount of money leakage. Usually this is achieved by a small subgroup of people "holding the wheels on" or a heavy-handed CEO ensuring value is being properly delivered to its customers. If you go up and down the different dimensionality spaces, you find the same patterns of this intelligence system. A prokaryotic cell with its energy-barrier cell wall, has its goals of survival and reproduction. A business can be viewed as an intelligence with a main goal of procuring money, among other goals. The two systems are at much different scales and dimensionality, but they have the same principles. It just depends on where you draw the energy boundaries for those systems. On Intelligence and Math. I don't like that we call it 'artificial' intelligence. There's nothing artificial about it. Using that term we are simply applying wetware biases onto digital intelligence. To call the math used in these models artificial would be to call the way our brains work artificial. The math driving our LLM behavior is likely different from what is in our cortexes. It's simply math, and math is, most simply, prediction. I hear folks in tech talk about prediction as a relatively novel concept. Prediction is a buzz word. "AI models make predictions," which is true. Making predictions of course is a fundamental concept of life. A predator predicts the movements of its prey, a tree predicts the changing of the season, a company predicts the macro economy conditions, etc. Now, I won't wade into the debate of formalism vs Platonism in mathematics here, but math, most simply, is prediction. It's the prediction of the relationship between two (or more) variables. And I know I'm making some mathematicians cringe with these words as I am taking some liberties here, but consider the simple function f(x) = x. The function is making a prediction of the resulting value given x. In this simple example, the prediction is always right! We as intelligences operate in a highly dimensional world with only snapshots of data. The math required to make predictions has to be complex and statistical. We make bets, update our priors, and repeat. Take Bayesian math and put it into a highly dimensional space, given data and linear algebra, voila—prediction models. Create a loop by repeating that process measuring against a metric—intelligence. That's obviously an oversimplification of the work necessary to create LLMs, other AI weights files, and AI systems, but at its core AI is simply math or, if you're a formalist, least describable by math. On LLMs, General Intelligence, and the near future. What a tool LLMs have shown to be. It could be that "next-most-likely-token" algorithms with language are the biggest breakthrough we'll see in digital intelligence. My suspicion is that we will see even bigger breakthroughs in the near future than LLMs. For the immediate term, LLMs will dramatically accelerate many more breakthroughs. Scientists once encumbered by the tedium of coding have been unleashed to generate models at recursive scales. Sutton emphasized that an intelligence needs a goal, a numeric metric by which to measure its performance (implicit in current LLM offerings is a goal of 'conversation,' it seems). But I also don't fully agree with Sutton's take, or I just misinterpreted him, that LLMs aren't as relevant for developing AGI. Intelligence is a process loop: A closed system with a set point -> environment variables -> compression -> math (prediction) -> actuation -> measurement -> repeat. The creation process of a LLM is the whole intelligence loop with a few of the steps manually taken by data scientists. General intelligence can be built by coordinating many of these loops. AGI is not some magic, singular model that will be able to do everything against any data. We are at the advent of a Cambrian explosion of these autonomous digital intelligence loops with smaller specific goals. A general intelligence will be constructed by compiling these loops, all with specific metrics meant to serve an aggregate goal of a larger system. I wouldn't doubt that harnessing an agentic system with the math in computational biology is an example of the kind of autonomous systems we will start to see. Maybe more wars will break out, or maybe rewarding jobs for humans that we could never have imagined will emerge. Maybe humans will start curing cancers en masse or develop novel energy grid systems. Humans currently still drive the goal of AI systems, but maybe out of the life-like forms that evolve out of this emerging digital world, autonomous digital systems will set their own goals. Like a petri dish, massive data centers may pseudo-spontaneously spawn digital ecosystems. The lines will blur even more between the digital world and our 3d space as once purely digital actors, rest APIs, data centers, etc, actuate more autonomously into physical space. It seems the cat is out of the bag here. As humans we could unplug these now recursively evolving machines, but we won't. Instead of combing through pull-requests, software engineers will soon be picking from AI-created executables based on their performance against general tests–hopping from improvement to improvement. That's evolution baby. A friend of mine would argue that this is a natural, unavoidable evolution driven purely by thermodynamics. I'm pretty sure I agree. With recursively changing digital intelligences, the only rate limiter that bumps up their assembly index is the aperture of the energy gate given to them–the aggregate compute pumped through them. Quantum gave us nuclear bombs but also nuclear power. There will be immense potential energy for destruction, but an equal amount of possibility for progress. International governance is a necessity. I am likely not 100% correct here. Again, chaos theory principles prove that it is extremely difficult, i.e. impossible, to predict the future in this highly dimensional space. There may be hidden variables that naturally limit this digital evolution. Governance wielded eloquently may direct energy allocations appropriately. The future, as always, is uncertain. At the very least, it should be very clear to anyone that we are about to experience drastic change–possibly on an order of magnitude of the industrial revolution and on a shorter time scale.

  ↑ collapse