DEV Community: oleg kholin

The Evolution of Circuit Compression

oleg kholin — Sun, 14 Jun 2026 14:37:31 +0000

I. From Tubes to Topology
Miniaturization replaced the vacuum tube with the transistor. The currency of compromise shifted from watts of heater power to milliwatts of leakage and dissipation. The circuit remained a set of separate decisions, each of which could be touched, desoldered, rearranged. Compromises lived in four distinguishable layers, from the technical specification to the printed circuit board pattern.
Micro-miniaturization did not continue the reduction — it changed the method of packaging. Four layers were pressed into a fifth: the transfer of topology onto a substrate. Component selection, routing, and power filtering ceased to be separate actions; they became a single lithographic pattern. The price for density turned out to be not size, but loss of transparency. The overweight of the sum of compromises, previously visible on the board, went inside the crystal.
Together with compression, a bundle of feedback connections emerged. In a discrete circuit, feedbacks were explicit — they could be broken with a probe. In an integrated circuit they became distributed: thermal, substrate-coupled, parasitic capacitive. The problem ceased to localize at a point; it began to drift through the network of compromises. Drift manifests not where it is born: noise in the speaker can begin as a supply sag in another corner of the die, and bias instability as heating of a neighboring stage. The network itself becomes the channel for error transport.

II. Drift of the Problem Through the Network
Any circuit is a network of compromises linked by a bundle of feedbacks — thermal, supply, parasitic capacitive, substrate. These connections are not drawn on the schematic; they arise as a consequence of placement. They form the transport network for error.
Drift is the movement of the place where a problem manifests along this network. A problem is born at one point in the layer of compromises, and becomes visible at another.
Drift in Discrete Circuits
In miniaturization on discrete elements, drift was slow and observable. Overheating of an output transistor changed the quiescent current of the input through a common power rail, and this could be traced with a probe from point to point. The bundle of feedbacks was sparse, so the trajectory of drift was readable.
Drift in Integrated Circuits
Micro-miniaturization compressed the network. The fifth layer — the transfer to the substrate — made feedbacks dense and invisible. Heat from digital logic drifts through silicon to a low-noise input and appears as increased noise. A supply sag in one corner of the die drifts along the common ground and appears as a bias shift in another corner. Parasitic capacitance between neighboring traces carries interference from output to input. Drift ceased to be movement across a board; it became movement through a field inside the crystal.
The key property of drift: it is not eliminated by local correction. An attempt to compensate the manifestation at the observation point does not touch the birth point. Therefore in an integrated circuit, bias correction in one stage often amplifies drift in another, because the bundle of feedbacks redistributes the overweight of the sum of compromises.
With the transition to large-scale integration, the network becomes even denser. Drift accelerates, because distances are small and thermal density is high. A problem born as a short current spike can drift through the substrate and appear milliseconds later as a long-term frequency shift.

III. The Proxy Channel
The attempt to control this drift led to the idea of a proxy channel. In a superheterodyne, the proxy is the intermediate frequency — the translation of a complex task into a region where filters are stable. In software-defined radio, the proxy is digital — the translation of physics into numbers. For a circuit, a proxy means taking the sum of compromises out of the physical layer into the informational one: measuring currents and temperatures, digitizing the error, and returning correction.
While the circuit remained small-scale integrated, the proxy could live outside. The transition to a large integrated circuit hid the proxy inside. It became part of the same fifth layer, and began to pay with heat and area for the right to treat heat and area.
This does not remove layers 1 to 5, but adds a sixth above them, where the compromise is no longer in die area but in the speed and accuracy of measurement. You pay not with heat but with processor cycles and memory for the model. Partially this already exists in digitally assisted analog, when an amplifier is calibrated by digital logic every millisecond. In full form, a proxy channel would mean that the circuit ceases to be a set of fixed compromises — it becomes a system that continuously translates its own errors into a convenient intermediate form and corrects itself there.

IV. The Intermediate Form Between 2D and 3D
Between planar integration and volumetric integration, an intermediate form appears — analogous to point-to-point wiring. This is not a trace in metallization and not a through via, but a bridge over the substrate, under it, or along the edge of the die: an air bridge, a backside power delivery network, a silicon bridge between chiplets.
Such a form returns part of controllability, allows bypassing an overloaded spot in the fifth layer without a full transition to a three-dimensional stack. It pays with lower reproducibility, but gives the ability to spread compromises in space.

V. Atomization
The alternative path — conditional atomization — proposes not to compress a circuit, but to grow a material with a given function. Here the layers of component selection and routing disappear; they are replaced by a layer of crystal synthesis. Compromises move from geometry to lattice physics: to purity, to uniformity of the doping gradient, to domain stability. The path requires a different currency, which at the moment of choice did not exist in controllable form.
The overweight of the sum of compromises does not disappear — it simply moves from geometry into lattice physics. Previously, drift was visible as a quiescent current that wandered; here it would be visible as a resonant frequency that shifted because of a single dislocation. And it would be impossible to correct with a trimmer — only with new growth. Therefore a layer would appear, but controlling it would be harder than the fifth.
In atomization, drift also changes its carrier. Instead of current along a conductor, it becomes the movement of a defect in the lattice or a domain wall in the material. A problem is born as growth non-uniformity and appears as a characteristic shift after hours of operation. The bundle of feedbacks here is the internal fields of the crystal, and drift through them cannot be stopped with a trimmer.

VI. Conclusion
Thus evolution looks not like linear shrinkage, but like a sequential change of the place where the overweight of compromises is stored. Miniaturization stored it in elements, micro-miniaturization in the plane of the crystal, large integration in volume and in the built-in proxy, atomization would store it in the substance itself. Each step solved some forms of drift and created new ones; each step redistributed the bundle of feedback connections but did not eliminate it.
The choice between compression into a layer, extraction into a proxy, or growth into material remains open, because only the currency of payment changes, not the fact of payment itself.

The Intention Decompiler: Algorithm as the Stable Layer of AI Workflow

oleg kholin — Sat, 13 Jun 2026 12:14:56 +0000

The value of a professional tool is determined not by how long one works in it — but by the density of relationships that accumulate inside the container. Photoshop stores not pixels — it stores a graph of transformations applied to the original. DaVinci Resolve stores not video — it stores a node graph of relationships between clips, color decisions, and effects. AutoCAD stores not a drawing — it stores geometry plus the procedures for constructing it in AutoLISP. Exporting from any of these tools destroys not the artifact — it destroys the architecture of relationships. You get the result without the procedure that produced it.

Temporal depth — the condition for accumulating relationships — manifests in different ways. A developer sits on a single task for hours and days: every commit, every PR, every issue adds new relationships to the container. Microsoft understood this and covered the entire cycle — VS Code, GitHub, Copilot, Azure — with a single container where relationships are closed. A video editor works on post-production for weeks: Blackmagic builds DaVinci Resolve on the same logic. A musician works intensively on a single track: Ableton with Max for Live embeds a programming language inside the container.

But there is another mode of accumulation — not a long session, but inheritance between short sessions. You work intensively, but briefly. You produce an artifact. You take it as a foundation — fork it for a new task. You work intensively again. Fork again. Temporal depth is created not by the length of a single session but by the chain of inheritance between sessions. Each artifact carries within it the accumulated knowledge of previous iterations. This is the exact model of working with AI prompts.

The first approximation of the niche looks like this: a prompt is source code, the model’s response is the compiled artifact, a library of prompts with inheritance is a container with growing connectedness. Git for prompts — versioning, forking, diff, collaboration. The analogy is elegant. And it breaks at the foundation.

Source code is separated from the compiler. sort(arr) works in Python 3.8 and in 3.12 — the syntax is stable across versions. A prompt is not separated from the model. A prompt written for GPT-4 produces a different result on Claude, a different result on Llama, a different result on the next version of the same model. The prompt is the compiler call — it does not exist independently of the execution environment. Git for prompts breaks precisely here: a diff between prompt versions is meaningless if the model has been updated. The role behaves differently. The context is interpreted differently. Constraints are followed differently. The inference logic — differently. Everything is bound to the model.

But underneath the prompt lies an algorithm. And the algorithm is stable.

Write on Medium
A prompt decomposes into two layers. The first — model-dependent: formulations, style, trigger tokens, syntactic patterns specific to a given model. The second — model-independent: role, inference logic, constraint structure, reasoning chain. GPT-4 and Claude use different formulations — but the logic of decomposing a task into subtasks is the same. Python and Rust have different syntax — but the sorting algorithm is the same. The value is not in the prompt. The value is in the model-independent layer.

Vibe coding proves this thesis from the opposite direction. In vibe coding, a person does not write a prompt at all — they describe an intention, and the AI generates the prompt and the artifact on its own. If a prompt is generated automatically — it was never source code. It was compiler input. The source code was always the algorithm underneath it. The hierarchy: human intention → algorithm as the stable layer → prompt as the model-dependent wrapper → artifact as the result. The algorithm is what must be stored. The prompt is generated. The artifact is discarded.

From this follows a structural argument through symmetry with an adjacent domain. Decompilers exist: IDA Pro and Ghidra take machine code and reconstruct source code. This works. But a tool that takes source code and extracts from it a clean algorithm — in the form of a portable logic graph, a flowchart of relationships — does not exist. Code visualization exists; algorithm extraction does not. Apply the same matrix to the world of prompts: prompt → artifact exists, that is any LLM; artifact → prompt is emerging as reverse prompt engineering; prompt → algorithm exists nowhere.

The product is not Git for prompts. The product is IDA Pro for prompts: a decompiler that extracts the model-independent algorithm from a prompt, builds a graph of its logic, and makes it portable across models and tasks. Copilot in this architecture identifies the algorithm inside the prompt — the way IDA finds functions in a binary. GitHub stores the graph of algorithms and the inheritance relationships between them. Cursor transfers the algorithm into a new prompt for a different model or task — the way a patch is applied to a new binary.

Inheritance happens not between prompts — but between algorithms. A fork does not lose meaning when the model changes because what is forked is not the text of the prompt but the logic underneath it. Temporal depth is created by the accumulation of an algorithm graph: each session adds new nodes and relationships, each fork inherits the stable layer and specializes it for a new context.

The niche is defined by double absence. A tool for decomposing a prompt down to its algorithm does not exist in any current AI product. An analogous tool for extracting an algorithm from code — also does not exist. This is not coincidence. This is a structural void formed because all existing tools work with the surface: with the text of the prompt, with the syntax of the code, with the pixels of the artifact. No one works with the logic beneath the surface as the primary object of storage and inheritance.

A container that stores algorithms rather than prompts — is an uncaptured niche. And uncaptured symmetrically: in the world of code and in the world of AI simultaneously.

Apple and Keeping the Ecosystem in Premium

oleg kholin — Fri, 12 Jun 2026 06:40:00 +0000

an analytical essay based on the model from article «The Evolution of New Things: Premium, Service, Environment»
The standard framing of the question — why iPhone remains expensive — is imprecise. In the model presented in article «The Evolution of New Things: Premium, Service, Environment», premium is not defined by price. Price is a consequence. Premium is when ownership remains the sole or primary means of access to a function. The question is more precisely formulated as follows: why does Apple retain ownership as the primary mode of access — not only to the device, but to the entire ecosystem?
The unit of analysis here is not the iPhone as a device, but the container «iPhone + ecosystem». It is this container that Apple deliberately keeps in premium. The services within it — iCloud+, Apple One, Music, TV+ — do not contradict this positioning. They serve a structural function: they make the ecosystem coherent, make exit costly, and keep the top-level container resistant to sliding down.

Downward Pressure
The smartphone is a mature category. The forces pushing it toward environment are specific.
Cameras have converged — Android flagships shoot comparably, often better on specific parameters. ARM architecture has become mainstream: Apple Silicon's advantage in the mobile segment is shrinking. Most basic functions are identical — messengers, navigation, browser, streaming work the same on any platform. Hardware is commoditizing: displays, modems, and sensors are manufactured by the same factories.
In the terms of article's model, this means: the market is pushing iPhone toward environment — toward a state where the device stops being noticed and becomes infrastructure. It is against this pressure that Apple constructs its architecture of retention.

Ownership as the Mode of Access In article's model, premium is not defined by price. The defining feature of premium is ownership as the primary means of accessing a function. Apple retains precisely this structure. There is no official way to obtain iOS without purchasing a device. There is no iPhone rental like a car-share — no access to the function without owning the device. Even the iPhone Upgrade Program, which superficially resembles a subscription, leads to ownership through trade-in: after 12 payments, the user returns the device and begins a new cycle. The purchase does not disappear — it becomes a ritual with a fixed rhythm. High price is a consequence of this structure, not its defining feature. ________________________________________
Inseparability of Shell from Resource According to the principle formulated in article's article, what transitions into environment is that which can be divided into an empty shell and a heavy resource. Apple keeps them as a single whole. iOS is not licensed — the iPhone shell cannot be installed on third-party hardware. The modem, camera, neural engine, and Secure Enclave are sealed inside. Even with the transition to USB-C, Apple attempted to maintain control: for the iPhone 15, restricting accessories to MFi-certified models was discussed, as was a custom authentication chip for the port. The Hackintosh phenomenon is instructive. The community attempted to separate macOS from Apple hardware — to break the coupling in the terms of article's article. It did not succeed: Apple keeps it closed both legally and technically. Hackintosh remained a niche for enthusiasts and confirmed the rule: until the shell is officially separated, the system does not become environment. ________________________________________
Exit Friction Environment is when a thing goes unnoticed. Apple makes exit visible. iMessage, AirDrop, FaceTime, Continuity, and Handoff work only within the ecosystem. The coherence is structured so that iPhone talks to Mac, Mac to Watch, Watch unlocks Mac, AirDrop just works — as long as the user remains within the walls. Photos, passwords, health data, and smart home settings live in iCloud with seamless synchronization. Migration to Android is a weekend project. The social cost of exit is built in separately: RCS compatibility is not a priority; green bubbles in iMessage remain a marker of belonging. The user pays not for the phone — but to avoid paying the cost of leaving. ________________________________________
Managing Time Premium is maintained through rhythm. An annual release on a single day worldwide turns the purchase into an event. Pro models receive features a year ahead of standard models — a ladder is created within premium. Long software support works paradoxically: Apple confirmed at WWDC 2026 that iOS 27 will support the 2019 iPhone 11 — seven years of updates. Normally, a long support cycle pushes a product toward environment. Here it works differently: the phone does not become waste, does not depreciate sharply, remains a relevant artifact. ________________________________________
Service as a Protective Layer This is the most subtle mechanism — and it is precisely this one that makes Apple's strategy not contradictory, but coherent. Apple is actively building a service layer: iCloud+, Apple One, Music, TV+, Arcade. At first glance, this looks like a move toward service — toward selling access instead of ownership. But the opposite is happening. Apple does not forbid service — it localizes service within the ecosystem. Each service strengthens the coherence of the container: iCloud holds data that cannot be moved without loss; Apple One builds the habit of paying Apple without releasing the user; Music and TV+ add reasons to stay. Services are not a retreat from premium. They are the load-bearing structure that keeps the top-level container in premium, making exit from the ecosystem progressively more costly. In the terms of article's model: Apple permits servicification inside the container so that the container itself does not get servicified. The broad lineup — SE, standard, Plus, Pro, Pro Max — operates on the same logic. This is not democratization, but the downward extension of premium: each tier sells ownership, not access. ________________________________________ What Has Already Become Environment Inside According to the container concept from article's article, it is important to distinguish levels. Inside the iPhone, the modem, Bluetooth stack, ARM cores, and codecs have long been environment — invisible, not purchased separately. Within the ecosystem, some services are moving in the same direction. But the top-level container — «iPhone + ecosystem» — Apple keeps in premium. The container will slide when its coherence breaks: when iMessage starts working everywhere, when Continuity ceases to be exclusive, when leaving the ecosystem stops feeling like a loss. Not when iPhone becomes cheaper. ________________________________________ Conclusion According to the model from article «The Evolution of New Things: Premium, Service, Environment», what is happening with Apple is not a collection of separate decisions. It is a unified strategy: Apple allows the servicification of everything inside the ecosystem so that the top-level container remains in premium. Service here is not a threat to premium — it is premium's protective shell. As long as the architecture does not allow separating the empty shell from the heavy resource — and as long as services maintain the coherence of the ecosystem — the container will remain in premium, even if all its internals have long since become infrastructure.

AI as a Thin Client and the Crisis of Knowledge Succession: An Academic Analysis

oleg kholin — Wed, 03 Jun 2026 09:37:33 +0000

Two Hypotheses In the contemporary discussion about artificial intelligence, two distinct hypotheses intersect and are often conflated.

The first hypothesis describes AI as a thin client between intention and result. Historically, a chain of translators existed between a concept and an artifact. A person formulated a task for a programmer, the programmer wrote code, the code became a program. A screenwriter passed an idea to a studio, the studio hired a VFX team, the team produced a film. A composer worked with musicians and a studio to record a track. AI shortens this chain, allowing a result to be obtained directly from a natural language prompt.

The second hypothesis is more radical. It asserts that AI washes out not only performers but also apprentices. The main function of many professions was not the production of the current result, but the reproduction of knowledge. A junior was needed not because he is useful today, but because in five years he will become a senior. A student was needed not to create value now, but to become an engineer. A doctoral candidate was needed not for brilliant papers, but to undergo the school of scientific thinking.

The Destruction of the Apprenticeship Mechanism The classical model of competence growth was built on review. A junior wrote code, a senior dissected it, extracted the substrate of experience, and transmitted professional intuition. Each review was an act of knowledge transfer.

The new model looks different. A person formulates a prompt, AI generates the result. If code of acceptable quality appears immediately, the economic need for a junior declines. Along with it, the mechanism through which knowledge was transmitted disappears.

A structural question arises that goes beyond the labor market. Where will the next seniors come from if the intermediate link does not undergo the path of learning through mistakes and reviews. This is a problem of competence reproduction, not simply automation.

The Transformation of Education Historically, the university and school performed the function of an institution of verification. The teacher took lived experience, analyzed it, and taught how to distinguish working knowledge from noise.

Under conditions of mass AI adoption, this function shifts. Teaching increasingly concentrates not on the subject, but on the ability to work with the model: formulating queries, checking answers, assembling agent chains. Knowledge of the subject is assumed to be available on demand, therefore teaching knowledge as such recedes to the background.

Education is turning from an institution of succession into a course on interacting with a thin client.

Where the Teaching of Knowledge Goes The teaching of knowledge does not disappear completely, but is pushed to the periphery and distributed across three directions.

First direction: inside models. Knowledge is preserved in the form of statistical weights, without an author, without context, and without a witness who could explain why a solution works.

Second direction: into narrow craft communities. Small laboratories, open-source groups, workshops where the practice of personal analysis and transmission of experience is preserved.

Third direction: into nowhere. A large part of intermediate knowledge simply ceases to be reproduced because the economic incentive to transmit it disappears. There is no systemic reason to teach rotoscoping, syntax, or mixing if these operations are performed by a model.

The paradoxical effect is that access to knowledge has become instantaneous, while learning knowledge has become a luxury. An indirect indicator of this shift is the growth in requests to encyclopedic resources. It is not the number of people who learn that is increasing, but the number of agents that index.

Can AI Become a Mentor The key assumption of the second hypothesis is that AI is fundamentally incapable of performing the function of a mentor. Today this assumption has grounding. Models provide answers well, but they form professional intuition poorly. A master usually says: this solution works, but in two years the system will collapse at this point. Such knowledge is based on lived experience of consequences, not on text patterns.

Current models work with corpora, not with experience of operating solutions. This limitation is not proof of a fundamental impossibility of AI mentorship, but it records the current state of the technology.

The Problem of Selection, Not Origin The most contentious claim is that new knowledge bases will be filled with statistical noise without verification. Historically, knowledge has never undergone ideal filtration. Universities, scientific schools, and corporations also produced a significant amount of noise.

The problem therefore lies not in who generates the content, a human or a model, but in the presence of a selection mechanism. If high-quality review, testing, replication of experiments, and audit exist, knowledge can be reproduced regardless of the origin of the text.

AI accelerates the production of information faster than society creates new institutions for its verification. It is precisely this gap between the speed of generation and the speed of verification that creates the risk of accumulating unreflective content in knowledge bases.

Conclusion
The analysis shows that the real subject of the discussion is shifting. The first part of the discussion describes AI as a tool for shortening production chains. The second part points to a more fundamental process.

AI removes intermediaries between intention and result, and together with the intermediaries, the institutions through which society reproduced bearers of knowledge disappear. The issue is not so much the automation of the labor of programmers, musicians, or artists, as the possible crisis of knowledge succession.

The key question of the next decade is not whether a model can write code or generate a film, but whether the social mechanism for the emergence of the next generation of specialists capable of understanding why this code and this film work will be preserved.

The Impact of AI Agent Development on Smartphone Screen Size: An Analysis of Trends, Paradoxes, and Architectural Shifts

oleg kholin — Mon, 04 May 2026 13:44:50 +0000

Introduction
The rapid growth and development of AI agents often leads to what seems, at first glance, an obvious thought — smartphone screen size must inevitably shrink. Indeed, if an intelligent agent can perform tasks by voice, work in the background, and deliver brief summaries instead of long feeds — why do we need a six-inch display? The logic appears flawless. However, a deeper analysis reveals that behind this assumption lies a series of paradoxes, and the real trends point in an entirely different direction.
Moreover, the question of screen size turns out to be merely the tip of the iceberg. Behind it stand far larger processes: a shift in the architecture of human interaction with computing, the emergence of a new class of market players, and — perhaps most unexpectedly — a revolution that may begin not with complex tasks, but with the simplest "remind me not to miss the turn."
In this work, we will examine the arguments on both sides, conduct their critical analysis, identify fundamental trends, and show that the question of screen size, for the first time in the history of smartphones, may be decided not by the manufacturer, but by the consumer — and that the answer lies deeper than it appears.
Arguments in Favor of Screen Reduction
Proponents of screen reduction put forward a number of arguments that appear compelling at first glance. Among them:
• Voice interaction is becoming the primary channel, making tactile input on a large screen redundant.
• AI agents are capable of generating brief summaries of texts, emails, and notifications, eliminating the need for a large display area.
• Agents perform tasks autonomously in the background, reducing screen usage time.
• The development of AR glasses transfers visual information from the smartphone screen to a wearable device.
• The growing ecosystem of wearable devices (smartwatches, AI pins, AI-enabled earbuds) distributes functions across multiple devices.
• The development of neural interfaces may, in the long term, eliminate the need for a visual channel altogether.
At a surface level, these arguments form a coherent picture: AI takes over tasks — the screen becomes less essential — the device shrinks. However, critical analysis of each of these arguments reveals significant weaknesses.
Critical Analysis: Why the Arguments "For" Don't Hold Up
Voice assistants have existed since 2011 (Apple Siri), yet over thirteen years of their presence on the market, the average smartphone screen size has only grown. Voice remains a niche scenario — timers, music, simple queries — while for complex tasks such as comparing products, navigating a document, or reading, the visual interface remains indispensable.
Wearable devices do indeed take over certain functions from the smartphone: earbuds have assumed audio, watches — notifications and fitness tracking. But in practice, users are not prepared to carry five devices instead of one, and the smartphone remains a universal tool — a "Swiss army knife" of digital life.
AR/VR technologies are perhaps the most promising direction, yet mass adoption of lightweight and affordable AR glasses is a matter of at least five to ten years. Current solutions (Apple Vision Pro at 600g and $3,500) are far from the mass market. Added to this is the social stigma, well known from the Google Glass experience.
Neural interfaces, with all due respect to the Neuralink project, remain in the realm of science fiction for the mass consumer — the horizon of their practical application is measured in decades.
Finally, the trend toward UX simplification has historically not led to smaller screens. On the contrary — more whitespace in the interface, larger typography, greater visual comfort. The iPhone grew in size precisely when Apple was simplifying iOS.
Arguments Against Screen Reduction
The arguments on the opposing side rest on fundamental rather than circumstantial factors:
Growth of video consumption. TikTok, YouTube Shorts, Reels, streaming platforms — video content is growing exponentially. The average user spends more than three hours per day watching video on a smartphone. AI amplifies this trend through personalized recommendations and AI-generated video content. No one will watch video on a screen smaller than the current one.
Visual verification. The more tasks an AI agent performs autonomously, the greater the user's need to verify the result before confirmation. A booked hotel, a sent payment, a composed letter to a supervisor — all of this requires visual review. An agent's error can cost real money.
Privacy. In the office, on public transport, in a café, voice interaction is impossible or socially unacceptable. This is not a technological limitation that can be overcome by an engineering solution — it is a fundamental property of human coexistence. As long as we live among other people, the screen remains a private channel of interaction.
Visual communication. Memes, stickers, video messages, stories — modern communication is approximately 70% visual. AI amplifies this trend by generating stickers, filters, and AI avatars. People communicate through images, and for that, a screen is needed.
Identified Trends
A deep analysis of both groups of arguments allows us to identify six fundamental trends, which we propose to divide into two categories.
Trends From the Device
The arguments "for" screen reduction, upon closer examination, point not to a smaller display, but to three technological trends:

Distributed computing. The smartphone is ceasing to be the sole device — computation is being distributed among glasses, watches, earbuds, and other devices. The screen is not shrinking — the smartphone is losing its monopoly.
Multimodal interaction. The number of ways to interact with a device is increasing — voice, gestures, gaze, touch. The user chooses a channel depending on context: at home — voice, in the subway — screen, while driving — voice. The screen does not disappear; it becomes one of many channels.
Transition from executor to supervisor. The user is doing less themselves and increasingly reviewing the results of AI's work. This does not reduce the need for a screen — it changes the nature of its use. Trends From the Human The arguments "against" reveal three trends rooted not in technology, but in human nature:
Explosive growth of generated content. AI endlessly creates visual content — images, video, charts, tables. The volume grows exponentially, while consumption remains a visual process that requires a screen.
Deficit of trust in AI. The more autonomy agents receive, the greater the need for transparency and verification of their actions. Verification is a visual task, and it requires screen space.
Privacy as a permanent barrier. Social norms and the need for confidentiality limit the spread of alternative interfaces — voice-based, AR glasses with cameras, neural interfaces. This barrier is not technological and cannot be overcome through engineering. Interconnection of Trends The most significant finding of this analysis is the discovery of systemic interconnections between the two groups of trends. Trends from the device and trends from the human do not contradict each other — they complement and mutually reinforce one another. The tendency toward the user's transition into the role of supervisor is amplified by the deficit of trust in AI. The more autonomous the agent, the greater the need for a screen to visually verify its actions. Multimodality of interaction collides with the barrier of privacy: new channels — voice, AR — are constrained by social norms, and the screen remains the primary private interface. The distribution of computing across devices cannot keep pace with the explosive growth of content: the volume of AI-generated visual material grows faster than the device ecosystem can distribute it. Who Determines Screen Size: The Manufacturer or the Consumer? At this stage, however, it is necessary to ask a question that calls into doubt all the preceding logic: does consumer behavior actually determine screen size? Historical practice suggests otherwise. Before 2007, any consumer survey would have shown absolute loyalty to physical keyboards: BlackBerry was iconic precisely for its tactile feedback, the Nokia E-series sold in the millions. The consumer did not ask for a virtual keyboard — Steve Jobs imposed it, and within a few years, physical buttons on smartphones disappeared as a class. The same story played out with the headphone jack, the removable battery, the SD card slot — all things the consumer "wanted," until the manufacturer decided otherwise. TikTok did not make screens large — Apple and Samsung made screens large, and TikTok emerged as a product optimized for the already existing vertical six-inch display. First the hardware — then the content to fit it. Apple killed the mini lineup not because the consumer "didn't want a compact smartphone" — the consumer wanted it and bought it — but because the margins were lower. In this logic, screen size is determined not by user needs, but by the manufacturer's product strategy: OLED panel costs, patent wars, camera-driven chassis thickness requirements, supplier agreements. And any analysis of user patterns — how many hours they watch video, how they verify AI's work — turns out to be methodologically fragile. The Turning Point: The AI Agent as a New Player And here we arrive at what is perhaps the most important conclusion of the first part of this study. The manufacturer of an AI agent is not Samsung competing with Apple over display brightness. It is a player from a different industry altogether, one that changes the very object of consumption. OpenAI, Anthropic, Google with Gemini — they sell not a device, but the ability to perform a task. And if that ability is accessible through any carrier — a smartphone, glasses, a speaker, an AI pin — then the hardware manufacturer's monopoly on shaping demand collapses. Previously, "needs" were shaped by the smartphone itself: Apple defined vertical video, and the industry followed. Now the AI agent is an independent product that the user selects separately from the device. For the first time, a reverse movement emerges: a person chooses an agent for their task, and then selects a carrier for the chosen agent. The hardware manufacturer finds itself in the position of follower, not dictator. This explains the failures of the Humane AI Pin and Rabbit R1 not as "the consumer wasn't ready," but as "the consumer was given a real choice for the first time" — and chose. Previously, such a choice did not exist: you bought a BlackBerry — you used the keyboard; you bought an iPhone — you used the glass. When a real choice of form factor for the same AI function appeared, it turned out that AI Pin and Rabbit were not what people needed — they needed a screen. This is a market vote that did not exist for the keyboard in 2007: back then, no alternative was offered. Beyond the Screen: AI as an Operating System However, the analysis would be incomplete if we stopped at the question of screen size. Behind it, an architectural shift of far greater magnitude comes into view. From Apps to the Agent Layer The history of computing has seen several fundamental transitions: DOS gave way to Windows, the desktop web to mobile operating systems, web search to app ecosystems. Each time, what changed was not merely the technology, but the foundational model of human access to computation. The next possible transition is from an app-centric OS to an agent-centric OS. The user interacts not with a set of applications, but with a unified agent layer, where "apps" become invisible backend tools. The precursors are already visible: AI browsers are partially replacing search and navigation, intent-first UX allows the user to articulate a task instead of opening a specific application, and cross-app orchestration promises to become a superstructure over the fragmented app economy. In this scenario, the smartphone ceases to be a "container for applications" and becomes a terminal for accessing the agent. The screen, microphone, camera, sensors — all remain, but the value shifts from iOS/Android to the agent system. AI is potentially capable of not merely weakening device manufacturers, but of creating a post-OS paradigm, where the traditional mobile operating system becomes the same kind of "invisible layer" that BIOS became for most users. Control of the market in this case may pass to whoever builds the dominant agent OS — be it OpenAI, Meta, Google, or a yet-unknown player. Zuckerberg was premature with the Facebook Phone, attempting to turn a social network into a device shell. But the bet back then was on the social graph, whereas today's AI agent operates on a cognitive graph — it can simultaneously become the shell, the interface, the coordinator, the search engine, and the workflow. This is potentially far more powerful. The Main Barrier Is Not Hardware However, the primary barrier on the path to an agent OS is not the hardware implementation. Creating an "AI phone" is technically possible today. The real barrier is orchestration, trust, and ecosystem depth. The agent must reliably execute actions and have access to payments, identity, messaging, APIs, and security systems. The winner will not be whoever makes a "smartphone with AI," but whoever creates a new computational environment of trust. The Revolution Begins With "Remind Me" And here we arrive at what may be the most unexpected turn of this entire study. Virtually all futuristic models overestimate complex scenarios — "organize a vacation," "manage my finances," "replace the OS entirely" — and systematically underestimate micro-mundane attention management. Cognitive Scaffolding Instead of Superintelligence The real mass AI-native experience may begin not with the automation of complex tasks, but with the simplest requests: • "Remind me not to miss the turn." • "Remind me when my grandson gets home." • "Remind me when it's 7 o'clock." Yes, even that. Not an alarm — but "remind me." The difference is fundamental: an alarm is a tool that needs to be configured. "Remind me" is the delegation of an intention to an agent that will figure out the method of execution on its own. This is a shift from task execution to cognitive scaffolding — not "do something complex for me," but "hold my context better than I can myself." Not a command executor, but a keeper of unfinished intentions — what can most precisely be called an ambient guardian of intention. Why "Remind Me" Is More Powerful Than "Organize" Micro-mundane scenarios possess three critical advantages over complex agent tasks. First, frequency: such requests arise hundreds of times per week, not once a month. Second, a low threshold of trust: "remind me to turn" carries no financial risk, unlike "send $3,000," making delegation psychologically comfortable. Third, habit formation: if a system reliably maintains everyday context, it becomes a cognitive prosthesis that is difficult to abandon. Historically, technologies win not through maximum complexity, but through the minimization of minor frustrations: autocomplete, GPS, push notifications, autosave. AI may win the mass market through anticipatory reminders — proactive nudges that connect geolocation, time, family graph, habits, calendar, and behavior into a unified contextual memory. Agent OS as a Replacement of Forgetting From this perspective, the agent operating system begins not as a replacement of applications, but as a replacement of forgetting. The first true AI revolution may turn out to be not in the automation of labor, but in the automation of memory and attention. And if this happens, then the simplest "remind me…" scenarios may become for the agent era what the alarm clock was for the early mobile phone: not a spectacular feature, but an everyday point of dependency. What This Means for the Screen At the same time, the revolution is first logical, then form-factor-driven. An AI OS will more likely first change the structure of interaction — kill app navigation, remove part of the UI complexity — than shrink the physical display. In the "reminder" scenario, the screen is needed less as a workspace, but more as a point of confirmation and trust calibration: "You asked to be reminded before the turn — now," "Grandson is home," "7:00." Brief, contextual, minimal messages. Paradoxically, this brings us back to the original question — but on a different level. The screen does not shrink because of AI agents as such. But if the agent OS wins through micro-mundane scenarios, the nature of screen usage will change so radically that the question of its size may be reformulated anew — no longer by the manufacturer or today's consumer, but by a new model of interaction in which the screen becomes not a workspace, but a window of confirmation. Conclusion The initial assumption that the development of AI agents will lead to a reduction in smartphone screen size finds no confirmation upon deep analysis. AI creates more visual content than ever before. Humans need to verify the agent's work more and more. Alternative interfaces run up against social and cultural barriers. But the main conclusions lie deeper than display size. First. The AI agent, existing above devices and platforms, breaks for the first time in twenty years the hardware manufacturers' monopoly on shaping the user experience. The consumer receives a real choice of form factor for the first time — and the early results of this vote (the failure of screenless AI devices) speak in favor of the screen. Second. Behind the question of screen size stands an architectural shift on the scale of DOS→Windows: a transition from an app-centric to an agent-centric operating system, where AI may become not a feature within the smartphone, but a new level of operational logic, calling into question for the first time in the mobile era the centrality of iOS and Android. Third. Mass adoption of the agent paradigm will most likely begin not with complex automation scenarios, but with the simplest cognitive scaffolding — "remind me," "warn me," "don't let me forget." The first true AI revolution may turn out to be a revolution not of labor, but of memory and attention. The smartphone screen will likely not shrink. But the world in which we look at it will change beyond recognition.

Adaptive Company: A CSS-like Language for Describing Organizational Structure Dynamics in Crisis

oleg kholin — Wed, 29 Apr 2026 10:42:06 +0000

Three Phrases Every Consultant Hears
"I can't see how the crisis is affecting the company"
"I can't see how the company behaves in a crisis"
"I can't see how we can change during a crisis"
The key word here is can't see. Not "don't know," not "don't understand." Specifically — can't see. The problem isn't a lack of data. The problem is the absence of a language capable of describing and showing what happens to a company's structure under pressure.

The Problem: Company Structure Is Described as a Photograph
Classical tools — org charts, UML diagrams, process descriptions — capture a single state. A snapshot. Here are the departments, here are the connections, here are the functions.
But in a crisis, the structure moves. People become overloaded, roles blur, departments contract, business directions die off. A static diagram won't show any of this. It becomes outdated the moment it's created.
What's needed is not a snapshot, but rules for how the snapshot changes. Not a description of the structure, but a description of how the structure transforms under pressure.

Theoretical Framework: Taleb and Three Types of Systems
Nassim Taleb identifies three levels of system response to stress:
• Fragile — breaks under pressure
• Robust — withstands pressure, maintaining functionality
• Antifragile — grows stronger from pressure
What is described in this article is not antifragility. A company doesn't become stronger from a crisis on its own. This is about robustness through managed structural reorganization: the ability to maintain functionality by reorganizing from within.

An Analogy from Frontend: Adaptive Layout, Not Fluid
In web development, there are two approaches to how a website responds to screen size changes:
Fluid layout — everything changes smoothly, continuously, totally. Blocks stretch, compress, overflow. There are no fixed modes. If the rules are poorly defined — elements break out of bounds, the interface falls apart.
Adaptive layout — the system operates in clearly defined modes with specific boundaries. When a threshold is reached — the layout restructures: the composition changes, blocks appear or disappear, the placement logic becomes different.
A company that responds to crisis in a "fluid" manner — without clearly defined modes — becomes unreadable to external players. Clients, partners, suppliers, regulators don't understand: who is responsible for what right now? What commitments are in effect? Where are the boundaries?
A company built on the principle of adaptive layout presents clear states to the outside world. Flexibility lives inside. A readable operating mode is what's shown outside.

Two Levels of Change
A company's structure in crisis doesn't change in just one way. There are two fundamentally different mechanisms:
Continuous Level (Within the Structure)
Roles, functions, employee workload, task distribution — all of this changes smoothly, within ranges. As long as the structure holds — the system compensates for pressure by redistributing the load.
Example: a department of 10 people, each with one professional profile. The crisis reduces headcount — the number of profiles per person grows. People take on more.
Discrete Level (The Structure Itself)
Departments, business directions, organizational units, connections between them — these change abruptly. This is reassembly: merging departments, removing management layers, shutting down business directions.
Example: employee profiles are a reflection of business directions. When a department shrinks below a critical threshold — it's no longer about overloading people. The budget can't sustain it, adjacent departments can't cope. This means the business directions themselves must be cut.
The key point: first, the system compensates internally. When the internal resource is exhausted — structural reassembly occurs.

Crisis Is Not a Single Parameter
A single cause of crisis generates multiple parallel consequences for a company:
• Financial pressure (cash flow, budgets)
• Staffing shortages (layoffs, overload)
• Operational overload (processes can't keep up)
• External environment pressure (regulators, market, clients)
A crisis doesn't strike along a single axis. It hits the entire system simultaneously.

Cascading Threats: Not a Gradation, but a Screen Rotation
When multiple threats coincide — the company's response is not sequential, not "step by step." It's not a gradual deterioration with a transition to the next level.
It's an instant mode switch. An analogy from the same frontend world: not a window resize, but a screen orientation change — portrait → landscape. Everything restructures at once and entirely.
A company that lacks pre-defined rules for such a switch loses controllability at that very moment.

Language Architecture: Two Layers
Layer A — Structure (UML Level)
The base description: departments, people, roles, functions, connections, business directions. This is the company's "skeleton" at a given moment. Classical UML handles this well.
Layer B — Deformation Rules (CSS Level)
A description of how the structure changes as pressure shifts:
• At certain environmental parameters → headcount changes
• At certain parameters → departments are reorganized
• At certain parameters → new functions and responsibilities are introduced
• At certain parameters → business directions are cut
This is not a static description, but a set of cascading transformation rules — analogous to how CSS defines rules for changing the display when conditions change.

Why the Visual Layer Matters
An important note: the company's business model doesn't change through this process. What changes is the internal organization. But to manage these changes, they need to be visible.
When the structure and its dynamics are visualized:
• Assessment criteria can be assigned to every element's state
• UI dashboards can be built for monitoring and analysis
• It becomes possible to see in real time: where the overload is, where the gaps are, where the structure is at its limit, where reassembly is already needed
Visualization is not decoration. It is a management instrument. Without it, the executive returns to those same three phrases: "I can't see."

Conclusion
A company is a system with two layers of dynamics:
• Internal → continuous redistribution of roles, functions, workload
• Structural → discrete reassembly of departments, connections, business directions
What's needed is a language that describes not a single state of the company, but the rules of transition between modes. A language in which:
• The UML level defines the structure
• The CSS level defines the rules of its transformation
• The visual layer turns this into a manageable, observable system
Then the executive stops "not seeing" — and begins to manage not by intuition, but by architecture.

The Evolution of the 3D Printing Problem: From Technological Optimism to Structural Deadlock

oleg kholin — Sat, 11 Apr 2026 11:08:43 +0000

The development of 3D printing over the past decades has been accompanied by a persistent expectation of its inevitable mass adoption. The logic appeared straightforward: the technology matured, hardware became cheaper, materials became widely available, and software gradually improved in usability. Within this framework, it was assumed that further cost reduction and simplification would eventually make the 3D printer as common a household device as a paper printer or a microwave oven.

However, the actual trajectory has been different. Despite technological maturity and accessibility, 3D printing has not become part of everyday domestic life. This discrepancy between expectation and reality is often explained through familiar arguments: the lack of a “killer use case,” high barriers to entry, poor economic competitiveness compared to mass-produced goods, and inferior product quality. Yet these explanations remain superficial and fail to address deeper structural causes.

The Initial Misframing: False Universality

The core issue begins with how the question itself is framed. The assumption that any mature technology must become mass-market ignores a fundamental distinction between classes of tasks. Some technologies serve daily or regularly recurring needs, while others address rare, highly variable, and context-specific problems.

Low cost and accessibility are not sufficient conditions for mass adoption. There are many examples of inexpensive, highly capable devices that never become household standards because they do not correspond to everyday needs. The ability to use a tool does not imply the necessity of using it.

In this context, 3D printing was incorrectly positioned from the outset. It was treated as a potential mass household technology, whereas by its nature it belongs to the category of specialized tools—similar to equipment used in workshops or production environments.

Reframing the Context: From “Every Home” to “Every Workshop”

Correcting the framing leads to a different interpretation. A 3D printer is not a household appliance in the conventional sense. It is a tool designed for solving problems that arise irregularly but require a high degree of customization.

From this perspective, it becomes clear that the technology has already found stable domains of application. Jewelry production, custom components for technical devices, advertising and promotional items, and educational construction kits all demonstrate effective use of 3D printing. These domains share a common characteristic: small batch sizes, high variability, and the absence of economic justification for traditional industrial manufacturing.

Thus, the issue is not the absence of demand, but its nature. The demand is not mass-market—it is niche, yet stable and reproducible.

The Illusion of Technological Limitations

Many arguments against broader adoption of 3D printing rely on outdated assumptions. Claims about insufficient precision, strength, or functionality increasingly fail to reflect current reality. Modern desktop systems are capable of producing working mechanical components suitable for practical use without additional finishing.

Other limitations, such as water resistance or consistency of output, are often interpreted as inherent to the technology. In practice, however, these depend heavily on process parameters. Their resolution lies in standardization and reproducibility of settings, not in altering the underlying physics of the process.

Thus, many perceived “limitations” are not technological but infrastructural.

The Ecosystem as a Consequence of Task Structure

Another commonly cited issue is the lack of a developed ecosystem—unified model libraries, standardized print profiles, and user-friendly tools. However, a deeper analysis shows that an ecosystem cannot emerge independently of a structured understanding of tasks.

In mature engineering and software systems, the primary layer is not the toolset but the ontology of objects and operations. Users work not with abstract geometry, but with entities that have parameters and behavior. This allows systems to scale through extensions, reuse, and accumulation of knowledge.

In 3D printing, the situation is reversed: tools exist, but there is no shared understanding of what tasks are being solved or how. As a result, each user constructs an individual workflow, and accumulated experience does not scale across the system.

Under these conditions, an ecosystem cannot be built directly. It can only emerge as a byproduct of task systematization.

The Representation Problem: From Geometry to Parameters

The dominant model exchange format—static geometric files—limits reuse and adaptability. These models contain no information about purpose, constraints, or functional parameters.

A parametric approach, by contrast, defines objects through relationships and constraints. This enables adaptation to specific conditions without breaking functionality. However, adoption of this approach is constrained by the lack of accessible tools aligned with real-world workflows.

The gap between existing CAD systems and practical user behavior remains one of the central barriers.

The Role of Adjacent Technologies

The evolution of 3D printing is closely tied to the maturity of adjacent technologies. One of the most critical missing components is affordable, accurate 3D scanning. The ability to quickly capture the geometry of existing objects would significantly simplify many practical workflows, particularly those involving replication or repair.

The absence of such tools increases labor costs and reduces accessibility, further limiting adoption. In this sense, 3D printing remains partially constrained by the immaturity of its technological ecosystem.

The Limits of Generative Solutions

Attempts to compensate for the lack of models through generative approaches encounter a fundamental limitation. Generative systems are oriented toward creating new forms, while many real-world tasks require accurate reproduction of existing objects under functional constraints.

Without embedded engineering logic, generated models may appear plausible but fail in practical use. This highlights the distinction between form synthesis and engineering design. The former may assist the latter, but cannot replace it.

The Absence of a Dominant Use Scenario

Another defining feature of 3D printing is the absence of a dominant, unifying application scenario. In successful technological domains, development is typically organized around a small number of clearly defined use cases, which drive standardization and infrastructure.

In contrast, 3D printing is characterized by a wide range of fragmented applications without consolidation. This fragmentation hinders standardization and slows ecosystem development.

The Non-Obvious Cause: The Absence of a Risk-Bearing Actor

The deepest layer of the problem lies in the distribution of risk. Building a fully functional ecosystem requires long-term investment, coordination across multiple layers, and acceptance of uncertainty. Yet the benefits of such an ecosystem are distributed across many participants, while the costs are concentrated on whoever initiates it.

Hardware manufacturers are incentivized to protect proprietary advantages rather than standardize. Software companies focus on high-margin enterprise markets. Open-source communities lack the resources to deliver robust, production-grade systems. Investors are reluctant to engage with long-term, uncertain, and weakly monetizable opportunities.

As a result, no actor emerges for whom building the ecosystem is a rational decision. This creates a structural deadlock: the technology exists, demand exists in niches, partial solutions exist—but integration does not occur.

This distinguishes 3D printing from cases of successful technological scaling. In those cases, there is always an actor for whom the cost of inaction exceeds the cost of building the system. That actor may be a company, a consortium, or a public institution—but it exists.

In 3D printing, such an actor has not yet emerged. Moreover, the current distribution of incentives actively discourages their appearance. Benefits are diffuse, while risks are concentrated.

Therefore, the absence of an ecosystem is not the root cause, but a consequence. The root cause lies in the economics of risk. As long as the cost of integration exceeds its expected return for any individual participant, systemic solutions will remain unrealized.

The Resulting Picture

3D printing is not a failed mass-market technology. It is a mature tool for a specific class of problems that do not align with everyday consumer use.

Its limitations are not primarily technological, but structural:

incorrect framing of mass adoption as a goal;
absence of a formalized task space;
inadequate model representation formats;
mismatch between tools and real workflows;
immaturity of adjacent technologies;
lack of dominant application scenarios;
absence of an actor willing to bear integration risk.
Future Directions

The future of 3D printing depends less on improving hardware and more on advancing the organization of knowledge and systems around it:

developing a clear taxonomy of tasks and use cases;
transitioning from geometric to parametric models;
creating tools aligned with actual workflows;
standardizing print profiles by object type rather than hardware;
advancing accessible methods for geometry acquisition;
identifying a limited number of scalable application domains.

Until such developments occur, 3D printing will remain an effective but localized tool—widely used in professional and semi-professional contexts, yet lacking a mechanism for broader systemic adoption.

Rising on the Shoulders of Giants: Top 10 Technologies That Succeeded by "Hijacking" Infrastructure

oleg kholin — Fri, 10 Apr 2026 04:58:47 +0000

Introduction: Why Do Some Technologies "Fly" While Others Stagnate?
According to the "Four Trees" framework, the most expensive and time-consuming stage of progress is growing the "Roots"—Trees 3 & 4 (the auxiliary tools and the infrastructure for mass production). Most brilliant ideas perish because they attempt to grow these roots from scratch.
However, there exists a strategy of "Engineering Refactoring." Instead of growing its own tree, a Strategic Subject performs a scan of the global market and identifies a mature Tree 4 in a completely different industry. By "hijacking" this existing infrastructure and adapting it to a new task, they achieve a lightning-fast leap to Tree 2 (the mass product). This approach allows them to collapse the cascade of costs by 80–90% and seize the market.
Below are 10 examples of how identifying and utilizing "foreign" auxiliary trees created new technological empires.

Consumer Drones Tree 1 (The Idea): An autonomous flying robot. The Donor (Tree 4): The smartphone industry. The Refactoring: DJI and others performed a "reconnaissance from the small," assembling drones from "spare parts" intended for phones. This collapsed the cost of such devices from $100,000 to $1,000.
Artificial Intelligence (AI/Neural Networks) Tree 1 (The Idea): Machine learning via neural networks. The Donor (Tree 4): The gaming industry (GPUs). The Refactoring: AI stagnated for decades until developers "recognized" in NVIDIA’s gaming video cards the perfect tool for matrix calculations. AI "parasitized" the global demand for video games, gaining a ready-made computational base.
Electric Vehicles (Tesla) Tree 1 (The Idea): A mass-market, high-performance electric car. The Donor (Tree 4): The consumer electronics industry (laptops). The Refactoring: Elon Musk used standard 18650 lithium-ion cells already mass-produced for laptops. Tesla rose on the "roots" planted by Panasonic and Sony, which were already mature and cheap.
Uber and Modern Logistics Tree 1 (The Idea): Real-time global management of movement. The Donor (Tree 4): Military satellite navigation (GPS) + 4G networks. The Refactoring: Uber utilized the existing Tree 4 created by the Pentagon and telecommunication giants. By changing the mechanism—using an app on a "borrowed" Tree 4—they annihilated the old taxi industry.
Medical Express Tests (PCR/COVID Tests) Tree 1 (The Idea): Instant molecular diagnostics at a scale of billions. The Donor (Tree 4): The food and beverage industry (PET bottle production). The Refactoring: The industry "hijacked" the production of PET preforms (the blanks used for soda bottles). The same mass-production lines of Tree 4 allowed the world to be flooded with cheap tests.
"Digital Twins" in Architecture Tree 1 (The Idea): Photo-realistic modeling of cities and factories. The Donor (Tree 4): Game Engines (Unreal Engine / Unity). The Refactoring: Hijacking game-engine technology allowed architects to design factories with a speed and fidelity that traditional CAD systems could never match.
Warehouse and Domestic Robots (LiDARs and Sensors) Tree 1 (The Idea): Robots that navigate space autonomously. The Donor (Tree 4): The automotive industry (ADAS systems). The Refactoring: Robotics "harvested the fruit" from trees planted by Toyota and Mercedes, which collapsed the price of LiDARs and cameras by implementing them in mass-market cars.
Cloud Computing (AWS) Tree 1 (The Idea): Selling computational power as a utility. The Donor (Tree 4): Amazon’s retail infrastructure. The Refactoring: Today, half the internet runs on the "roots" Amazon originally grew to support its own online bookstore.
Vertical Farming (AgroTech) Tree 1 (The Idea): Year-round food production within urban centers. The Donor (Tree 4): The LED display industry (TVs and Smartphones). The Refactoring: Vertical farming became viable only when mass production of screens collapsed the price of LEDs. Agriculture "plugged into" the auxiliary tree of the electronics industry.
3D Bioprinting Tree 1 (The Idea): Printing human organs and tissues. The Donor (Tree 4): Office inkjet printing. The Refactoring: The first bioprinters were created by hacking standard HP and Epson printers. The mature Tree 4 mechanics of droplet delivery were hijacked to print living cells.

Investment Radar: Identifying Future Unicorns through Tree Analysis
For an investor, the "Four Trees" model serves as a precision instrument to identify hidden market leaders at an early stage. A true "unicorn" is often born not in a lab, but at the intersection of a mature, external infrastructure and a novel business objective.

"Donor Signal" Detection The beginning of an explosive growth phase for a new player can often be predicted by analyzing the financial reporting of entirely different companies—component suppliers. The Core Concept: A sudden, unexplained spike in sales of specific components among "shoulder companies" (producers of chips, sensors, batteries) often signals the emergence of a hidden giant in an adjacent industry. Example: An abnormal growth in demand for miniature gyroscopes and Li-Po batteries between 2008 and 2010 was the signal for the birth of the consumer drone industry, well before brands like DJI became household names. Investors should watch those who suddenly begin buying "smartphone parts" for non-smartphone purposes.
Investment Arbitrage on the "Broken Cascade" The most profitable companies are those that reach the mass-market product stage (Tree 2) without bearing the colossal costs of building their own fundamental base (Trees 3–4). The Strategy: Look for companies with abnormally high margins in their early stages and a low level of Capital Expenditure (CapEx) compared to industry leaders. The Indicator: If a company shows results comparable to industry giants while having a 10x smaller R&D budget, it is a sure sign they have successfully performed an Engineering Refactoring and are using someone else's infrastructure as a free resource.
Searching for "Universal Roots" (Predicting the Next Wave) An investor can predict the next wave of technological breakthroughs simply by evaluating the maturity of specific "donor markets." The Forecast: As soon as a technology in one industry (e.g., satellite internet, laser scanners, or solid-state batteries) reaches the Tree 4 stage—meaning it has become cheap, reliable, and mass-produced—it automatically opens a "window of opportunity" for new Tree 1 ideas in adjacent fields. The Action: Analyze which mature "roots" are currently sitting on the market, unused. The entity that is first to "graft" these roots onto a new market niche will become the next global leader.

Conclusion: The Lesson of the Strategic Subject
All these examples share a single pattern: Success came not through the invention of new "roots," but through the ability to recognize them in other industries.
The winner is not the one who tries to build everything from scratch, but the one who performs Engineering Refactoring and "grafts" their idea onto the strongest and most massive Tree 4 available on the market. Do not just follow science journals; watch the supply chains of mass-market components. Real business subjectivity is the ability to recognize someone else's mature Tree 4 faster than the competition and rebuild your architecture to exploit it, collapsing the cost structures of everyone else in the process!

The Invisible Roots of Progress: Top 10 Supermaterials Stuck in the Laboratory

oleg kholin — Wed, 08 Apr 2026 20:20:26 +0000

The popular essay "The Four Trees" offers an original lens through which to view technological progress. According to this concept, the development of any technology rests upon four metaphorical "trees":

Tree 1 (The Idea): The fundamental concept or laboratory proof-of-concept. The principle is proven; the physics works.
Tree 2 (The Mass Product): The stage of mass production and widespread infrastructure. What we produce at scale and use in daily life.
Trees 3 & 4 (The Auxiliary Roots): Auxiliary tools and the secondary technologies used to produce them. These are the "hidden" roots—the lithography machines, the specialized furnaces, the methods of purification, and the precise manipulation of matter.

The Great Barrier: From Assembly to Integration

The reason many supermaterials fail to go mainstream is deeper than mere "cost." We are currently stuck in the trap of Miniaturization. This is the stage where we simply shrink individual components and attempt to connect them (similar to how vacuum tubes were replaced by discrete transistors).

The true revolutionary leap is Micro-miniaturization (Integration). This is the transition from "assembling discrete parts" to "forming a structure." In microelectronics, we don't solder millions of transistors together; we grow them simultaneously as a single, integrated structure on a silicon wafer through deposition and etching.

The tragedy of modern supermaterials: We still treat them as "discrete parts" (we try to "cut" graphene or "glue" a nanotube). We are still thinking in the category of "assembly," whereas we desperately need a "lithography for materials." Until we learn to form the structure of a device directly out of the material itself, we will remain in the era of "expensive transistors," never reaching the era of "cheap integrated circuits."

Below are the Top 10 Tree 1 materials waiting for their "integrated revolution."

1. Graphene: Two-Dimensional Carbon

Tree 1 Status: Proven in 2004. A single-atom-thick layer of carbon. The strongest and most conductive material in the universe.
The Trees 3–4 Bottleneck: We are still trying to "transfer" it like a delicate film. This is the era of assembly. For graphene to reach Tree 2, it must be grown directly into the specific regions of a chip as part of a unified integrated circuit, rather than being "pasted" onto existing ones.

2. Nitinol and Shape Memory Alloys

Tree 1 Status: Alloys (e.g., Titanium-Nickel) that return to a complex original shape when heated.
The Trees 3–4 Bottleneck: We currently produce them as discrete "parts" (stents, wires). We lack the technology to integrate "shape memory" directly into the 3D structure of a product during the fabrication stage, allowing the material itself to act as the mechanism.

3. Carbon Nanotubes (CNTs)

Tree 1 Status: Cylindrical carbon structures, 100 times stronger than steel and lighter than aluminum.
The Trees 3–4 Bottleneck: We can produce "nanopowder" (a discrete additive), but we cannot yet form a continuous macro-structure (like a thread or a sheet) without losing their unique properties at the molecular boundaries. We need a method of "weaving" the structure at the moment of formation, rather than attempting to assemble billions of individual fibers.

4. Metallic Glasses (Amorphous Metals)

Tree 1 Status: Metals with a disordered, liquid-like atomic structure. Extremely strong and immune to corrosion.
The Trees 3–4 Bottleneck: We are limited by the requirement of "extreme cooling rates," which restricts us to making only thin ribbons or small parts. We lack the Tree 4 technology to form bulk structures where the amorphous state is preserved during the casting of large masses.

5. Aerogels: "Frozen Smoke"

Tree 1 Status: A crystalline lattice consisting of 99% air. The lightest solid and the world’s best thermal insulator.
The Trees 3–4 Bottleneck: Production requires supercritical drying in high-pressure autoclaves—a boutique, "batch-assembly" method. To become Tree 2, aerogels must evolve into a material that can be deposited like a spray-on foam directly at a construction site, forming its structure in-situ.

6. MXenes: 2D Metals

Tree 1 Status: Two-dimensional crystals made of metals and carbon, capable of ultra-fast battery charging.
The Trees 3–4 Bottleneck: Currently obtained through "etching" (a subtractive and dirty chemical process). This is a discrete, wasteful method. We need the technology to "grow" MXenes directly as electrodes within the pre-integrated structure of a battery.

7. Borophene: Single-Atom Boron

Tree 1 Status: A 2D layer of boron, even stronger and more flexible than graphene.
The Trees 3–4 Bottleneck: It can only survive in an ultra-high vacuum. We lack the technology for "integrated encapsulation"—where the material is grown and immediately sealed with a protective atomic layer in a single, continuous process.

8. Perovskites: "Printable" Solar Power

Tree 1 Status: Crystals that convert light to electricity more efficiently than silicon.
The Trees 3–4 Bottleneck: They degrade rapidly in the presence of moisture. The solution is not just "better chemistry" but a breakthrough in "integrated sandwich-structure" fabrication, where the active perovskite and its transparent protection are formed as a unified, airtight structure during the printing process.

9. High-Entropy Alloys (HEAs)

Tree 1 Status: Alloys made of 5 or more metals in equal proportions, offering extreme heat and radiation resistance.
The Trees 3–4 Bottleneck: The problem of homogeneity. We need "atomic mixing" tools (such as high-speed laser deposition) so that the alloy is formed directly as the final part, rather than being smelted into a bulk ingot that requires further, less precise processing.

10. Synthetic Spider Silk: Bio-Kevlar

Tree 1 Status: Stronger than Kevlar and more elastic than Nylon.
The Trees 3–4 Bottleneck: We can produce the protein (the raw material), but we cannot yet "form the thread" (the structure) with the same molecular grace as a spider. This is the transition from "brewing a soup" in a bioreactor to "molecular-level weaving."

Final Conclusion

Today’s supermaterials are stuck at the transistor level of the 1950s. We have learned how to create them, but we have not yet learned how to unite them into "Integrated Circuits of Matter."

The problem is not that these technologies are inherently too expensive; the problem is that we are still trying to assemble the future by hand, piece by piece, instead of forming its structure as a unified whole. The entity that first creates a "lithography for materials"—moving from the assembly of parts to the integrated growth of systems—will become the new technological leader, controlling real progress!

Decontextualization of Anthropomorphism in Robotics: An Epistemological and Physical Analysis of a Paradigm Shift

oleg kholin — Sat, 28 Mar 2026 19:38:24 +0000

Abstract
Contemporary robotics relies heavily on anthropomorphic morphology as a universal standard for interaction with the human environment. This paper proposes a reconsideration of this assumption. It is shown that anthropomorphism is often not an engineering necessity, but a historical and cognitive legacy that emerged within a specific infrastructural context. A method of decontextualizing anthropomorphic assumptions is proposed through physical stress-tests of the environment, demonstrating the limitations of humanoid architectures. The concept of separating social and physical compatibility of agents is introduced, and a program of adaptive tests is proposed aimed at identifying optimal morphologies for extreme and unstable environments. The paper formulates a framework for transitioning from anthropocentric design to physically conditioned morphological optimization.

Introduction The majority of contemporary robotics is developed within an implicit assumption: a robot must be compatible with infrastructure created for humans. From this assumption, a second is often derived: a robot must be anthropomorphic. However, this assumption is rarely subjected to systematic analysis at the level of the physics of interaction with the real environment. In many scenarios — rescue operations, unstable surfaces, chemically active zones, destroyed infrastructures — humanoid morphology proves to be not optimal, and sometimes physically unstable. This leads to the key research question: can anthropomorphism be not a universal engineering solution, but a contextually limited historical design strategy? This paper proposes viewing the current situation as an epistemological shift in robotics, in which a change in the physical context of tasks naturally leads to the loss of applicability of old paradigms — without the need for their direct refutation.
The Physical Impossibility of Old Paradigms Scientific paradigms sometimes cease to function not because they are logically refuted, but because the physical context in which they were effective disappears. In engineering systems this manifests particularly clearly: a change in environment automatically changes the optimal morphology of the agent. When a robot must function under conditions of:

unstable surfaces,
degrading objects,
chemically active environments,
phase transitions of materials,

anthropomorphic architecture begins to encounter fundamental limitations of mechanics and stability.
In this sense, the paradigm shift occurs not through argumentation, but through the physics of tasks.

The Historical Contextuality of Infrastructure Human infrastructure is often perceived as a universal environment to which all artificial agents must adapt. However, infrastructures were formed historically and adapted to human biomechanics. The history of technology shows that: environment and agents evolve jointly. Examples:

automobiles led to the appearance of asphalt roads and traffic signals,
wheelchairs — to ramps and changes in architectural standards,
electric vehicles — to the infrastructure of charging stations.

Thus, the requirement for robots to fully adapt to existing infrastructure ignores the historical precedent of technological co-evolution.

Biomechanics as a Variable Human biomechanics is often regarded as a fixed model of interaction with the environment. However, real conditions demonstrate the opposite. When the environment changes, the movement strategy changes as well:

on slippery surfaces, additional points of support are used,
on steep inclines, the center of mass is shifted,
in unstable environments, the contact area is increased.

Consequently, the requirement to universally preserve anthropomorphic form means fixing one point in the space of possible morphologies, ignoring the adaptability characteristic of engineering systems.

Social Acceptance Does Not Require Anthropomorphism One of the common arguments in favor of humanoid robots is the convenience of interaction. However, empirical reality demonstrates the opposite. Many technological agents have been successfully integrated into the social environment without anthropomorphic form:

ATMs,
order terminals,
self-checkout machines,
information kiosks.

People interact with them without discomfort, because the key factor turns out to be not resemblance to a human, but:

predictability of behavior
and clarity of functional role.

This allows two independent parameters to be identified:

social compatibility
and physical compatibility.

Physical Limitations of Anthropomorphic Morphology Humanoid architecture possesses a number of characteristics that in certain environments become limitations:

high center of mass
limited contact area
discrete step kinematics
high sensitivity to loss of traction
complexity of load redistribution

A thought experiment involving a humanoid robot moving across thin ice with a load demonstrates these limitations.
Narrow contact points create high pressure on the surface, increasing the probability of structural failure of the support. A high center of mass reduces stability during sliding, and the discrete structure of the step limits adaptation to a continuously changing surface.
Under such conditions, anthropomorphism becomes not merely a suboptimal solution, but a potential engineering error.

The Method of Decontextualization of Anthropomorphism This paper proposes a method that can be described as the decontextualization of engineering myths. The idea of the method is as follows: an engineering hypothesis is tested not only in a standard environment, but also under conditions of changing physical context. If the model loses its operability when the environment changes, its universality proves to be illusory. The method includes three stages:

changing the physical context of the task
observing the degradation of the original morphology
searching for an alternative architecture

In this way, the paradigm is dismantled naturally — through incompatibility with new conditions.

The Evolution of Adaptivity Tests To identify the limitations of anthropomorphic systems, a sequence of tests of increasing complexity is proposed. Test 1 — Retention of a Degrading Object The system must accompany an object that is gradually losing its shape and structure. This tests the ability to adapt to continuous changes in geometry. Test 2 — Evacuation of a Human During Morphological Collapse A human in an extreme environment may lose the structural stability of the body. The robot must complete the rescue before the point of irreversible damage. This introduces a temporal and dynamic component to the task. Test 3 — Extraction of Materials Before Phase Transition The task includes chemical and physical instability of the object, which may become dangerous. Here the ability of the system to operate under conditions of changing matter is tested. These tests shift the focus from the form of the robot to the physics of interaction.
The Precedents Approach Instead of directly asserting a new paradigm theoretically, development through a chain of engineering precedents is possible. First, tasks appear in which old architectures systematically fail. Then different research groups independently find new solutions. As a result, a new practice is formed that gradually becomes the standard. This path is slow, but sustainable: it creates change through the accumulation of facts, not through declarations.
Separation of Social and Physical Architecture One of the key conclusions of the paper is the necessity of separating two functions of the robot:

the social interface
and the physical executive body.

Anthropomorphism may be useful as an interface:

for communication,
for recognition of intentions,
for reducing the cognitive load on the human.

However, physical morphology must be determined exclusively by the physics of the environment.
Separating these levels opens up new architectures of robotic systems.

Co-evolution of Environment and Agents The history of technology shows that environments change under the influence of new agents. In the long term, the emergence of new types of robots may lead to changes in infrastructure:

adaptive surfaces,
dynamic transport environments,
hybrid architectural systems.

Thus, the future of robotics may lie not in copying human form, but in the joint design of environment and agents.

Conclusion The paper proposes an analytical framework in which anthropomorphism is viewed not as a universal standard of robotics, but as a historically conditioned strategy, effective only in certain contexts. It is shown that a change in the physical environment of tasks naturally leads to the loss of applicability of old morphologies. Under these conditions, a transition becomes necessary — from anthropocentric design to physically conditioned morphological optimization. Such a transition is not the victory of one idea over another, but rather a change in the plane on which questions are posed in robotics. On the new plane, the key criterion is no longer the robot's resemblance to a human, but the correspondence of its form to the physics of the environment.

AI and Creativity: Making Sense of a Real Shift

oleg kholin — Wed, 25 Mar 2026 10:25:02 +0000

Our perspective on Tim Green's article "No Consent, No Credit, No Pay"

What Is Actually Happening

The public debate around generative AI and artists' rights is focused on legal details — datasets, lawsuits, licensing models. But behind all this noise, the main point escapes notice: we are witnessing not a technological improvement of existing tools, but the elimination of the very need for intermediaries between an idea and its realisation.

The chain used to look like this: idea → specialist → tool → result. Now it looks like this: idea → result. AI did not replace the designer with a better Photoshop. It made the designer an unnecessary link in the chain. And this is permanent.

Filters and plugins never cancelled the tools themselves. 3ds Max, After Effects, Photoshop — they remained necessary, which meant the people who knew how to use them remained necessary too. AI became a thin client that replaced both the tools and the specialists in one move: designers, layout artists, retouchers, illustrators, pattern makers. You no longer need them — and you no longer need what they used either.

The Analogy Everyone Misses

The authors of such articles and the participants in court proceedings compare what is happening to piracy or copyright infringement on the internet. That is an imprecise analogy.

The precise one is Gutenberg's printing press. It did not give scribes a better writing tool. It made the scribe an unnecessary link in the distribution of text. The profession did not disappear overnight — new niches emerged, new forms of craft. But the economic foundation was undermined irreversibly.

Even closer is the history of photography. The arrival of smartphones with quality cameras did not destroy photographers in direct competition. It simply turned out that the vast majority of consumers were satisfied with the level of photography available on their phones. Professionals survived in niches: artistic reportage, film photography, studio work. Enthusiasts of film still exist — just as enthusiasts of valve amplifiers do. But the monopoly on quality imagery collapsed forever.

AI is doing exactly the same thing to illustration and design.

What Really Worries Artists — and What They Are Confusing

The article lists grievances: lack of consent, attribution, compensation, style copying. Let us examine each honestly.

Style copying is painful, but it is not new. On DeviantArt, an interesting new style gets copied the very next day — without any AI involved. Style has never been a subject of legal protection in any jurisdiction. AI has merely accelerated and scaled what was already happening.

Legal lawsuits are not built on style but on fact: specific images were physically present in the LAION-5B dataset, downloaded and used for commercial purposes without the authors' permission. This is closer to a real violation — but even here the boundary is blurred. If you train a model on photographs of interiors where paintings hang on walls, purchased by the homeowners — who is the infringer? The law does not yet have an answer.

The demand for attribution from AI seems strange when we never demanded it from human artists inspired by each other's work. This is not an argument — it is a symptom of disorientation.

Compensation is the only genuinely strong argument. Companies earned billions not from specific images but from models for whose creation those images were indispensable. The Spotify analogy works here: the platform pays authors not because it reproduces a specific track, but because it uses the entire catalogue as the foundation of its business. The logic of Getty Images and Sweden's STIM is exactly this — and it is convincing.

But even winning every lawsuit and receiving royalties will not bring back the corporate illustration market. It is gone — just as photographers lost the family portrait market.

Where the Real Protection Lies

Practice shows that those who survive are not those who litigate, but those who keep moving.

Unique technique is real protection. An authorial system built on strict geometric curves, non-trivial mathematical foundations, rare combinations — this is opaque to AI. It averages what appeared frequently in the dataset. What is rare and original it cannot reproduce — it struggles even to describe it.

Fusion — combining two or more styles in non-trivial combinations — produces works that AI cannot decompose into base layers if that particular combination was absent from its training. Papercut art plus liquid art: AI sees the result but cannot see the structure.

However, there is a crucial practical nuance here. "Idea → result" is not yet a straight arrow. AI in its current mass form does not know your visual language. It averages. It drifts. It draws things you did not ask for. Ask it for a shadow from a fountain — and it will try to draw the fountain too. This is not a flaw that will be patched in the next update. It is a fundamental property of a model trained on averaged mass data: it does not understand local logic, a part without the whole.

This means the gap between intention and realisation is still real. And as long as that gap exists, craft does not disappear — it simply changes its form.

Selling process, not just result — the most sustainable model. An artist who sells not only paintings but brush profiles, techniques, and tools sells something AI does not produce. AI uses brushes but does not sell them.

Live contact with the audience creates attachment to the author as a person, not to the genre. This is what economists call switching cost — viewers may go to AI for an image, but they come back to the artist for the human being.

The Deficit That Was Always There

The deepest problem is neither technical nor legal. An idea is never protected by copyright anywhere — only its specific realisation. DeviantArt demonstrates this daily: a new idea survives one day before the first imitator appears. AI merely accelerates an already existing process.

The real deficit — of original ideas — existed long before generative AI. Where did the calligraphers go? They still exist, but the bulk of calligraphy is now produced by electronic and software means. This is not a tragedy — it is a civilisational shift.

What Comes Next — and Why It Cannot Be Stopped Either

There is one more dimension that changes the entire picture.

Right now AI is a mass tool with averaged models. This is like a typewriter: everyone gets the same font. But the moment is approaching when an artist will be able to train a model on materials they have personally selected — including their own paintings, their own visual language, their own logic of form. And this cannot be stopped either.

A personally trained model closes the loop. It knows your visual language. It understands your local logic. It realises your specific intention rather than an averaged one. The gap between intention and realisation — the shadow without the fountain — becomes your problem to solve, not a limitation imposed by someone else's model.

This is no longer "a thin client replaced the specialist." This is the specialist acquiring a personal thin client. A qualitatively different situation. And it destroys the linear picture of "AI displacing the artist." The real trajectory is more complex: mass AI displaces the mass market, but personal AI amplifies the unique author.

The Real Scale of What Is Happening

The authors of articles like this one and the participants in court proceedings are discussing compensation for the past. That is understandable and humanly just. But the future is already structured differently.

The real question is not legal. It is civilisational: what to do with the mass release of creative professions — just as the industrial revolution released manual labour, and digital photography released portrait photographers.

History gives no grounds for panic — every such shift generated new niches, new forms of mastery, new markets. But it gives no grounds for illusion either. What is gone is gone forever.

The artists filing lawsuits are fighting for compensation for the past. Those who will thrive are the ones who understand that the instrument has changed, build a personal relationship with their audience, develop techniques that cannot be averaged, and — when the moment comes — train their own model on their own material.

That is not the end of craft. That is craft in its new form.

And we have to live with that.

Four Trees of Technological Progress. What Usually Goes Unseen.

oleg kholin — Mon, 23 Mar 2026 09:24:31 +0000

Everyone knows the game Civilization. It has a technology tree — you research writing, mathematics unlocks, then astronomy, then navigation. Clean and logical.
But the real technological tree is more complex. Behind every node of the visible tree hide three invisible trees. And without them the main tree simply cannot grow.
Let us go through this with concrete examples.

Tree 1 — The Main Tree of Technologies
This is what everyone sees. A new idea or scientific discovery gives birth to a new technology.
Einstein in 1917 described the theory of stimulated emission of light. In 1960 Theodore Maiman built the first working laser. In 1903 the Wright brothers understood how to control the lift of a wing. In 1938 Otto Hahn split the uranium atom. These are all nodes of the main tree — ideas that change the world.

Tree 2 — The Tree of Instruments
Every new technology requires a new instrument for its realisation. This is the first hidden tree — the direct manifestation of the main tree.
The laser required a ruby crystal of special purity, mirrors accurate to the nanometre, a flash lamp of strictly defined power. The Wright brothers' aircraft required a light petrol engine — a steam engine was too heavy. The atomic bomb required centrifuges for separating uranium isotopes — without them uranium enrichment is physically impossible. The transistor required monocrystalline silicon of extraordinary purity — 99.9999999%.
It would seem that is all. There is an idea, there is an instrument for its realisation. But this is where things get truly interesting.

Tree 3 — The Shadow Tree of Instruments
The new instrument from Tree 2 cannot be created with old instruments. To produce it, entirely new auxiliary instruments are needed — not for the final product, but specifically for the creation of the main instrument.
To build the ASML lithographic machine for chip production — special machines for polishing lenses to atomic-level precision were needed. These machines did not exist before the task arose. To build the engine for the Wright brothers' aircraft — Charles Taylor, their mechanic, constructed a special lathe for turning aluminium cylinders. The existing machines of that time did not provide the necessary precision. To build the centrifuges for uranium enrichment in the Manhattan Project — special balancing machines had to be built, because rotor imbalance at 50,000 revolutions per minute destroyed the centrifuge within seconds. To grow monocrystalline silicon for transistors — Gordon Teal at Bell Labs in 1950 built a special Czochralski apparatus with temperature control accurate to fractions of a degree. Nothing like it existed in industry at that time.

Tree 4 — The Shadow Tree of Technologies
But that is still not the bottom. To create the auxiliary instruments from Tree 3 — new technologies are needed specifically for their production. This is the fourth tree, the most invisible of all.
To polish the lenses for ASML — a new technology of ion-beam etching of glass was required, which simply did not exist before. It was developed by the company Zeiss specifically for this task. Zeiss has existed since 1846 and today is the sole manufacturer of these lenses in the world. To balance the rotors of centrifuges in the Manhattan Project — a new technology of dynamic balancing of rotating bodies at ultra-high speeds was required. Before this, balancing had only been applied to steam turbines at speeds hundreds of times lower. To control the temperature in the monocrystalline silicon growing apparatus — new thermocouple and temperature regulator technology was required, with precision unachievable by industrial equipment of that era. To build Taylor's lathe for aluminium cylinders — a new technology of aluminium cutting was required. Aluminium was then a new metal and would "smear" ordinary cutting tools — special geometry and cutting speeds were needed.

From Electronics to AI — The Same Four Trees
Let us apply the same system to artificial intelligence — one of the most complex technological chains in the history of humanity.
Tree 1 — The Main Tree of AI Technologies
In 1943 mathematician Warren McCulloch and neurophysiologist Walter Pitts described a mathematical model of a neuron — a binary element that either fires or does not. This was pure theory, with no practical application.
In 1957 Frank Rosenblatt created the perceptron — the first learning neural network. But it could only solve linearly separable problems and hit a dead end.
In 1974 Paul Werbos described the backpropagation algorithm. The idea allowed a network to learn from its errors, adjusting weights at each layer. In 1986 Geoffrey Hinton, Rumelhart and Williams republished it and made it practical. This thawed AI research after years of "winter".
In 1997 LSTMs appeared — Long Short-Term Memory networks that could work with sequences and retain context. But they processed text strictly one word at a time — parallel processing was impossible.
In 2017 researchers at Google published the paper "Attention Is All You Need". The transformer architecture made it possible to process all text simultaneously rather than sequentially — each token is compared against all others through the attention mechanism. From this grew GPT, BERT, Claude and all of modern AI.
Tree 2 — The Tree of AI Instruments
The transformer architecture requires processing enormous matrices of numbers simultaneously — billions of matrix multiplication operations per second. A CPU is physically incapable of this — it is sequential. What was needed was a GPU — a graphics processing unit, architecturally designed for parallel computation.
Training modern large language models requires not just GPUs, but thousands of GPUs working simultaneously. By 2024 Meta had 600,000 H100 GPUs dedicated to AI research and development. Elon Musk's startup xAI built the Colossus supercomputer with 200,000 NVIDIA H100/H200 GPUs.
Beyond GPUs, special chips with high-speed HBM — High Bandwidth Memory — were needed, which ordinary GPUs did not have. And specialised server platforms with hundreds of gigabytes of RAM and high-speed inter-processor NVLink connections.
Tree 3 — The Shadow Tree of AI Instruments
GPUs for AI contain billions of transistors on an area the size of a fingernail. To manufacture them, lithographic machines with extreme ultraviolet radiation are needed — EUV lithography machines from ASML. These are machines the height of a double-decker bus, costing 150 to 200 million dollars each. Before the 2010s such machines simply did not exist — they had to be created specifically for this purpose.
To manufacture GPU chips, ISO Class 1 cleanrooms are required — spaces in which no more than 10 particles of size 0.1 micron are present per cubic metre of air. This is millions of times cleaner than ordinary air. Creating such spaces required new filtration systems, new construction materials, new working protocols.
To connect thousands of GPUs together into clusters for AI training, specialised high-speed InfiniBand and NVLink switches were needed — equipment that before the AI era existed only in narrowly specialised supercomputers and was not commercially available at the required scale.
Tree 4 — The Shadow Tree of AI Technologies
To build the ASML EUV lithography machine, a new technology for generating extreme ultraviolet light was required — a laser beam strikes a droplet of tin in a vacuum 50,000 times per second, creating a plasma at the temperature of the surface of the Sun, which emits EUV light. This technology was developed from scratch over more than 20 years through the joint efforts of ASML, Zeiss and dozens of other companies.
To create HBM memory for GPUs, a new technology of three-dimensional chip packaging was needed — TSV, Through-Silicon Vias. This is a method of vertically connecting silicon layers through microscopic holes. Before the AI era, such packaging density was needed by no one.
To train transformers on thousands of GPUs simultaneously, a new GPU programming technology was needed — CUDA. It was released by NVIDIA in 2007 and became central to NVIDIA's strategy of positioning the GPU as a universal tool for scientific applications. By 2015 CUDA development was increasingly focused on accelerating machine learning workloads. NVIDIA developed specialised libraries — cuDNN for deep learning and cuBLAS for linear algebra. CUDA is not simply a library. It is an entirely new technology for programming parallel computation, which did not exist before it.

The Conclusion — The Correct Order of Growth:
A step forward in the main tree of technologies actually looks like this:
First a new idea or technology emerges. To realise it, a new main instrument is needed. But that instrument cannot be created with old instruments — new auxiliary instruments are needed. And to create the auxiliary instruments, new auxiliary technologies are needed.
That is, the real direction of movement is reverse — first Tree 4 grows, then Tree 3, then Tree 2, and only then does Tree 1 become possible.
For AI this looked as follows: first in 2007 CUDA appeared — a new GPU programming technology (Tree 4). This made possible the creation of next-generation GPU clusters — H100 and A100 (Tree 3). Those in turn provided the instruments for practically training transformers at industrial scale (Tree 2). And only then did the theoretical idea from the 2017 paper "Attention Is All You Need" become GPT-4, Claude and all of modern AI (Tree 1).
Note: the transformer idea appeared in 2017. But the infrastructure for its realisation was being built from 2006 to 2020 — in parallel and independently. No one planned this as a unified programme. It simply happened that at the right moment all four trees were sufficiently grown for AI to become possible.

Why This Matters
This is precisely why genuine technological progress is so slow and expensive. Behind every visible step stand three invisible steps that had to be taken first.
And this is precisely why when countries want to stop someone else's progress — they block not the final products, but Trees 3 and 4. They do not ban rockets or aircraft. They ban ASML lithography machines, special machine tools, precision bearings, metal alloying technologies. Whoever owns Trees 3 and 4 controls Trees 1 and 2. Always.