Open Sourcing Intelligence

In the push for who gets to define intelligence, there is a 3rd way.

We spent centuries believing intelligence would be defined through philosophy, but it’s now being hardcoded by engineers who trained their models on stolen and scraped data. This is one of the reasons for the hidden architecture behind the world’s largest AI frontier models. Those few companies built on documentable data have been able to be transparent, letting their technical choices spread as the global default.

Boringly, this isn’t about ideology as much as about how copyright violations have almost accidentally determined the definitions of intelligence in the AI era.

1. The Battle for AI’s Soul

For the last century, the question of what intelligence actually is belonged to a loose alliance of philosophers, psychologists, scientists and what we might call logicians. In thinking about technology, they have spent decades debating the “ghost in the machine,” trying to measure IQ or map the logic of the human soul.

Today, that territory has been almost entirely occupied by computer scientists and engineers.

In real time, we are watching tech companies interrogate the deepest mysteries of human existence: what is a mind, and how do we perceive truth?

When an engineer structures a machine’s thought process, they are doing more than laying out and training computer code; they are hard-coding a specific version of the world. They are deciding in advance which patterns matter and which truths are chosen or discarded.

If you look closely, the current frontier ecosystem presented as a race for better software is in fact a sort of strategic dance around the idea of the future of 'general' intelligence. It is an attempt to plant a flag on the very definition of how a mind should work, ensuring that whoever wins the architecture wins the right to define reality itself.

If this is the window of time in which decisions on our shared perception of the mind are shaped, it is then a fundamental debate over the source of truth.

Competing blueprints for AGI

At the heart of the modern AI race is the pursuit of Artificial General Intelligence (AGI). This is the hypothetical threshold at which a machine can understand or learn any intellectual task a human can.

While this definition provides a common goal, the path to reaching it has mostly split into two distinct architectural philosophies.

On one side stands the perspective of Demis Hassabis and Google DeepMind. This is a vision of Artificial General Intelligence designed to understand the fundamental nature of reality. It is a bet that truth is something discovered through scientific inquiry. And a high-stakes breakthrough intended to accelerate human discovery for the benefit of all society.

Opposite them is the “World Model” perspective championed by Yann LeCun, formerly at Meta and a pioneer in AI. This vision prizes physical intuition and common sense, building an AI that understands the world by observing and predicting it. For LeCun, truth isn’t found in a probabilistic language model; it is found in the physical reality we all inhabit, learned through self-supervised observation.

He dismisses the pursuit of LLM-based AGI as a Silicon Valley obsession. He argues that language alone is too narrow a data source; true human-level intelligence requires physical world grounding and an understanding of causality that text alone cannot provide. The real world is the only place in which intelligent thinking can be truly designed.

Philosophy used to be settled through argument. Now it is being shaped by system architecture, which is the structural blueprint that determines how a machine reasons.

Both represent a different path toward the same hopeful goal: a world where artificial intelligence enhances our own. AGI, a once-terrifying worldview, is now part of 2026 business planning.

Huge breakthroughs have already changed how we live and work. Science has accelerated in previously unthinkable ways. And possibilities seem awe-inspiring and sort of infinite.

But perhaps what is most troubling is that I am referencing two very wealthy men working for the biggest corporations on earth as having a blueprint for the future - of being responsible and indeed capable of setting the standard for how we define our shared human truths.

There is a third way.

What is not up for grabs in either debate above is the boring business of frontier model transparency. And this is critical. Every AI model is essentially a technical bet on how thinking works: it embeds specific assumptions about whether intelligence emerges from probabilistic language patterns or from structured world models that understand physical reality.

Not knowing what is happening in the AI black box has become a de facto truth that no one seems to be able to force - whether they’re the European Union or the Abu Dhabi government. That is, until the Chinese company DeepSeek split the debate open.

DeepSeek AI is a Chinese artificial intelligence company that develops advanced, cost-efficient large language models (LLMs) with a core philosophy centered on open-source accessibility, reasoning transparency, and computational efficiency. I think that Open Source in AI is the spanner in the works. Let’s see what happens.

There is an argument here:

1. Philosophy has been hijacked by engineering - The question of what intelligence is has moved from philosophers to engineers who are embedding their answers in code.

2. Copyright law created an opening - Western companies (OpenAI, Meta, Google) trained on stolen copyrighted data, so they can’t be transparent without facing lawsuits. DeepSeek trained on data they can document, so they can be fully open.

3. Openness determines adoption: developers use what they can actually access and modify. DeepSeek’s architecture spreads because it’s genuinely available, while others hide behind APIs.

4. Architecture becomes reality - When DeepSeek’s specific technical choices (attention mechanisms, optimisation strategies) get embedded in thousands of applications, those choices become the global definition of how intelligence works.

5. Scientific language makes it permanent - Once these definitions are framed as “optimisation problems” and “loss functions,” they acquire scientific authority that’s hard to challenge.

2. Background: Knowledge as Open Source

But first, we need to back up. To understand why openness matters in AI, we need to look at how intelligence has actually been produced in the past.

Some of the earliest surviving literary works were produced through collective processes rather than single authorship.

The Epic of Gilgamesh dates to roughly 2100–1200 BCE in ancient Mesopotamia. It centres on Gilgamesh, king of Uruk, and follows his friendship with Enkidu, his confrontation with loss, and his search for immortality. The stories existed first as oral narratives, later recorded on clay tablets in Sumerian and Akkadian. Different versions were copied and rearranged by scribes across regions over several centuries. There is no single authoritative text.

The Kalevala was compiled in 1835 by Elias Lönnrot from oral poems collected across Finland and Karelia. These poems described creation myths and magical practices. The material varied between singers and locations and changed with repeated performance. Lönnrot assembled these variants into a written epic, but the content itself came from many anonymous contributors over time.

In both cases, the material predated its written form. The texts were shaped through repeated transmission rather than a single act of composition.

Across cultures, durable knowledge has often been produced without a single author or fixed form.

• In South Asia, the Vedas were transmitted orally for centuries before being written down, preserved through collective memorisation and ritual repetition rather than textual authority.

• In West African societies, griot traditions maintained historical, legal, and moral knowledge through performance, adaptation, and communal correction.

• Indigenous legal systems, from Māori tikanga to Andean ayllu practices, encoded social and ecological knowledge in shared customs that could be revised through use rather than decreed from above.

• In China, classical texts such as the I Ching evolved through layers of commentary added over generations, with meaning emerging from accumulation rather than origin.

The I Ching is a foundational text in Chinese philosophy that describes the universe as a dynamic interplay of complementary forces: yin (broken lines) and yang (solid lines). Divination using the I Ching involves a form of cleromancy (divination by casting lots) to generate one of 64 possible hexagrams, which are six-line figures.

Divination in the context of the I Ching (or Book of Changes) is a method of seeking guidance and insight into current situations and their potential transformations, rather than mere fortune-telling. It is an ancient practice, with roots extending back over 3,000 years to the Western Zhou period (1000–750 BCE) in China.

In each case, knowledge was sustained through participation. And in the late twentieth century, similar collective dynamics re-emerged through computation.

Computational poetry developed as a field in which writers and researchers use algorithms and digital tools to generate and analyse text. Rather than producing finished poems, practitioners design systems that shape how language is generated and recombined.

Much of this work has been built through open-source tools.

Projects such as open computational poetry platforms and applications distributed through repositories like F-Droid allow users to write poetry using shared codebases. These tools expose their underlying logic, making the generative process inspectable and adaptable rather than fixed.

“When people first started programming computers, one of the first things they were trying to do was to get them to write poems, So there’s this long history, especially with poems and computers, that’s actually out there and really fascinating.” Lillian-Yvonne Bertram

What changes with contemporary AI is the dismantling of collective revision. These systems do not “think” in any meaningful human sense, but they increasingly mediate how thinking is organised. When those mediating structures are sealed inside proprietary architectures, the capacity to contest or correct them disappears.

And poetry has played a far more significant role in AI than most people even know.

The claim that poetry has become an active site of AI safety research rests on the fact that poetic language reliably breaks assumptions built into language models. This has been demonstrated in multiple ways.

Techniques described as “adversarial poetry” use verse, metaphor, and structural ambiguity to bypass model safeguards, revealing vulnerabilities in how language models interpret intent and meaning.

Adversarial attacks exploit a simple reality — machine learning systems do not “understand” language the way humans do. They detect patterns, predict likely continuations and follow instructions based on what their training and safety layers interpret as intent.

Recent research (“Poetry Breaks AI Safety: 62% of Language Models Fail When Prompts Rhyme”) has shown that poetic and constrained language can also interact unpredictably with large language models.

Researchers at Icaro Lab, Sapienza University of Rome, and DEXAI found that Large Language Models (LLMs) often provide restricted or dangerous information when the request is presented as a poem, even after refusing the same request in plain prose.

This has made poetry an active site of research in AI safety and language modelling.

In this context, computation does not replace authorship. It redistributes it. It provides the opportunity for non-linear and culturally disruptive input. And meaning emerges from the interaction between human input, formal rules, and shared tools rather than from a single, closed system.

Intelligence, in this context, is no longer something a community shapes through use. It becomes a technical layer delivered to users who can interact with its outputs but have no access to its assumptions.

Source: Beyond Narrative Description: Generating Poetry from Images by Multi-Adversarial Training April 2018. The framework of poetry generation with multi-adversarial training. We first use image and poem pairs (a) from humanannotated paired image and poem dataset (MultiM-Poem) to train a deep coupled visual-poetic embedding mode.

3. AI & Open Source

a. Software & Open Source

Who even remembers Linux? But when you talk about open source software, it was one of the first.

In traditional software, “open source” is simple: anyone can see or modify code. Platforms like Linux, Mozilla’s Firefox, and Google’s Android ecosystem are genuinely open source as we define it today.

In other words, anyone can download their complete source code and modify it to create their own version.

Linux runs most of the world’s servers because companies can customise it exactly for their needs without paying Microsoft or Apple. Firefox’s code is completely public, which is why developers have created hundreds of variants from Tor Browser to specialised forks that Mozilla doesn’t control. Android lets phone manufacturers like Samsung or Xiaomi take Google’s code and modify it however they want.

That’s why Android phones from different companies look and work differently—they’re all modifications of the same open source base.

But AI is different. And open source in AI is another paradigm.

To have a working AI model, you need the training code (the recipe for how to build the AI), the training data (millions of documents, images, and conversations the AI learns from), and the weights (the actual “brain” that results, which costs millions to create).

In late 2024, the Open Source Initiative (OSI) released OSAID 1.0, the official definition for open-source AI. To qualify, you must release complete training code and model weights with no restrictive licences. But there’s a massive catch: you don’t have to release the actual data, just the “data information.”

Why? Because AI training data is a legal minefield of copyrighted books and news articles. Requiring full data release would mean no major AI could ever be open source, then they’d face endless lawsuits.

b. Why is AI not “Open”?

Perhaps the most blunt way to think about this is in terms of the data and who owns it. The big AI platforms do not have the copyright for a vast amount of the data they have used to train their models.

OpenAI, Meta, and Google have all trained their flagship models on vast collections of text scraped from the web, including books, articles, songs, images and websites protected by copyright.

Much of this material was not licensed directly from rights-holders.

That choice now sits at the centre of a growing wave of lawsuits from authors, publishers, trade union bodies and media organisations, who argue their work was used without permission or compensation. The companies, in turn, claim that training constitutes fair use because models learn statistical patterns rather than reproducing original texts.

This legal uncertainty is one reason training data remains opaque. Full disclosure would expose the scale and sources of those datasets, sharply increasing legal risk. Until the courts settle how copyright applies to AI training, most frontier models are likely to remain closed or only partially open.

c. How does this matter in AI today?

Meta built their entire AI strategy around claiming Llama is “open source”—it gets them free development work from thousands of contributors who presumably think they’re part of an open project. But under OSAID 1.0 rules, they actually fail the test.

Their training code is incomplete, and their weights are subject to restrictive licences that prohibit commercial use without permission. Meta loses the “open source” label and the developer goodwill that comes with it.

d. The DeepSeek Moment

A third approach is emerging through DeepSeek, a Chinese AI lab founded by Liang Wenfeng, who built his fortune in high-frequency quantitative hedge funds. DeepSeek views AI as infrastructure to be commoditised, not a service to be sold via monthly subscription.

This represents a fundamental shift in how AI is valued and distributed. Traditionally, AI has been treated like Netflix (a premium service you pay for monthly); DeepSeek wants it to be treated like Electricity (a cheap, basic utility that powers everything). To make AI a commodity, it has to be cheap.

Instead of selling intelligence or the “brain,” they provide the “infrastructure” for others to build tools on top of.

In doing this, they’ve also downgraded the high-minded Silicon Valley chit chat about the future of humanity, which presents many questions about how disruptive or god-like we think AI might be.

In December 2024, they released DeepSeek-V3, matching GPT-4’s performance at 95% less cost: 671B parameters with only 37B active through Mixture-of-Experts design, trained for $5.5M versus OpenAI’s estimated $100M+. They released complete technical papers, training code, and weights with zero restrictions.

The real breakthrough was both a better model than OpenAi’s - at the time; it was a different kind of capitalist logic. By releasing detailed technical blueprints - which Silcon Valley companies had never done! - e.g. the rather dense Multi-head Latent Attention (MLA), DeepSeek proved that elite intelligence could be achieved without endless scaling.

The argument, in a sense, is that the future belongs to those who can build the most intelligence with the fewest resources.

There is a strong argument that DeepSeek can be transparent because they were primarily trained on Chinese internet data, where copyright enforcement differs, or on data they have clearer rights to. They face different legal constraints than companies that scraped Western copyrighted content.

Once DeepSeek showed this was possible, openness stopped being a moral argument and became a tactical one. By releasing their weights and code without Western legal constraints, they shifted the centre of gravity.

In the last year, Western labs quietly adopted DeepSeek’s efficiency gains. They redesigned systems to reduce costs—while maintaining tight control over distribution.

They took the technical lesson but suppressed its implications: that intelligence could now be built outside protected institutions. Meanwhile, DeepSeek’s architecture has spread rapidly globally because developers can actually use it.

What does this mean for the third way?

There is unlikely to be a singular “winner” in the short run. But the scale of this shift, backed by tens of trillions of dollars and not to mention the full weight of geopolitics, means that whichever AI model gets adopted becomes the de facto answer to what intelligence is.

4. Quantum AI & Open Source

If you think the battle for AI is messy, it is important to look at what’s happening in Quantum. The quantum computing field faces similar questions about openness, but with different constraints.

For context, quantum computers will likely power the next generation of AI.

They’re not necessarily going to be faster at matrix multiplication, but despite this, they can explore solution spaces that classical computers can’t reach. Quantum neural networks could find patterns in data that are fundamentally invisible to classical AI. And madly! - Quantum samplers could generate training data from probability distributions that would take classical computers the age of the universe to explore.

Quantum computers require extraordinary physical infrastructure. The quantum bits (qubits) only maintain their quantum properties at temperatures near absolute zero, about -273°C. To achieve this, companies use dilution refrigerators, multi-stage cooling systems the size of large wardrobes that use liquid helium and sophisticated thermodynamic processes to create the coldest environments on Earth.

On top of this, quantum programming is exceptionlly complicated. You’re manipulating quantum states through gates that rotate and entangle qubits in precise patterns. Every operation introduces errors that compound exponentially. The qubits have limited connectivity—not every qubit can talk to every other qubit. The quantum information constantly decays. You need hundreds of physical qubits to represent one logical qubit that works reliably.

The qubits themselves sit in vacuum chambers to prevent any air molecules from disrupting their quantum states. Even a single photon of thermal radiation can cause decoherence—destroying the quantum information. The control systems use precisely calibrated microwave pulses to manipulate the qubits.

Each pulse must be timed to nanosecond precision with exact frequencies and phases. Any error compounds exponentially.

IBM, Google, Quantinuum, and IonQ spent billions building these systems. Their business model was straightforward: you write your quantum algorithm, submit it to their cloud service, they run it on their hardware, and bill you for the computing time.

In a way, it is just classic cloud computing economics iterated and applied to commercialising frontier physics.

But programming a quantum computer means choreographing quantum gates—operations that rotate and entangle qubits in precise patterns. Researchers will have to account for the noise in the system and the constant decay of quantum information. Without higher-level software tools, only a small group of physicists and quantum specialists could realistically programme these machines.

So IBM created Qiskit, a Python tool that lets you build quantum programmes without working directly at the hardware level. Google released Cirq for its own processors. These tools take a high-level description of a quantum algorithm and convert it into the exact sequence of operations that each machine can run.

Then Xanadu changed everything. They released PennyLane in 2018—a framework that lets you write your quantum programme once and run it anywhere. IBM’s superconducting qubits, IonQ’s trapped ions, Xanadu’s photonic chips—the same code works across all platforms. Your quantum algorithm becomes portable knowledge rather than something tied to a single vendor’s hardware.

The Open Quantum Movement

The Unitary Foundation, founded by quantum researcher Will Zeng in 2018, is ensuring quantum computing doesn’t become another walled garden. They maintain Mitiq, which enables today’s noisy quantum computers to operate despite their errors. Without error mitigation, these machines would be costly number generators. Mitiq lets startups achieve real results now, rather than wait decades for perfect quantum computers. They fund developers worldwide, including in Nigeria, Brazil, and India.

Chinese research institutions are making significant contributions to open-source quantum tools. They’re publishing implementations of variational quantum eigensolvers (VQE). These are algorithms that could revolutionise drug discovery by simulating protein folding and interactions. They’re releasing quantum approximate optimisation algorithms (QAOA) that solve the routing problems.

In an extraordinary move, the Institute of Computing Technology in Beijing released its entire quantum control stack. This software layer translates code into the actual microwave pulses that manipulate qubits.

This is the layer companies usually keep secret because it’s where the real engineering happens. By open-sourcing it, they’re letting anyone build quantum computer control systems without starting from scratch.

Arts, Philosophy & Open Quantum

This is the same pattern as AI: whoever’s architecture becomes the standard wins. But with quantum, the stakes are different.

We’re not just defining how to process information—we’re defining how to process information that exists in superposition, that can be entangled, that operates on fundamentally different mathematical principles.

The communities forming around open quantum tools include cryptographers preparing for quantum computers that will break current encryption, climate scientists who need quantum simulations to model complex systems, and materials scientists designing room-temperature superconductors. They’re building a shared language between domains that never had to talk before.

I’m involved in one such effort, working with researchers across quantum physics, AI, design, and the humanities at a point where standards and assumptions are still being formed. The aim is not to build faster systems or proprietary advantage, but to shape the intellectual and governance frameworks alongside the technology as a baseline, and before the infrastructure hardens and those decisions become more difficult to revisit.

The technical standards being established now, how quantum and classical computing interface, will determine who participates in these breakthroughs.

Taking it one step outside of science, what this looks like in practice is not theoretical. When artists work with quantum systems, they develop visual languages for phenomena that have no classical analogue. When philosophers engage with superposition, they’re developing frameworks for systems that exist in multiple states simultaneously until observed.

5. Frontier Science & Open Source

The shift triggered by open architectures is no longer confined to AI models. It is beginning to reshape the infrastructure used for scientific simulation, including quantum research.

Until recently, advanced simulation depended on tightly coupled stacks of software and hardware owned by a small number of organisations. Access was limited not just by expertise but by cost, exclusivity, and institutional control. Running meaningful experiments often meant committing to a single vendor’s tools and machines.

What has begun to change is the software layer. Over the past five years, open-source frameworks have enabled the development of scientific and quantum programs that run across different systems. The same code can now be moved between quantum processors and cloud providers with relatively minor modification. The bottleneck has shifted from rewriting software to gaining access to hardware time.

As these tools become standard, their assumptions travel with them. Decisions about how problems are framed, which variables are optimised, and what counts as a valid result are embedded in the software itself. Once widely adopted, those choices shape which questions are easy to ask and which answers are recognised as legitimate.

• When intelligence is framed in scientific terms, those definitions acquire a durability that is difficult to challenge.

• Once intelligence is rendered reproducible and machine-readable, other ways of knowing begin to fall outside the frame.

Power, in this context, no longer comes only from owning the most advanced machines. It comes from shaping the infrastructure others rely on to do their work.

This matters because science and mathematics carry a particular kind of authority. For centuries, their language has claimed neutrality and universality, presenting its outputs as descriptions of reality rather than interpretations. Equations, benchmarks, models, and simulations all operationalise the makers’ assumptions and intentions.

In this sense, the power of AI lies beyond computation in the scientific grammar through which it speaks.

Data Quality & Frontier Science

In 2026, a crucial middle path is occupied by the science-led AI startups such as Axiom Math and Periodic Labs. These firms have moved away from general-purpose chatbots that often “guess” answers based on internet data. Instead, they focus on absolute accuracy in specialised fields like robotics, mathematics and chemistry.

The industry is witnessing a shift from probabilistic token prediction toward deterministic precision. So no more hallucinations.

While general-purpose LLMs rely on statistical likelihood—often leading to hallucinations in rigid domains—specialised startups utilise symbolic reasoning and reinforcement learning to ensure accuracy. By treating disciplines like mathematics as code that must “compile” against logical axioms, these firms replace the “best-guess” nature of chatbots with verifiable, zero-error architectures.

Axiom Math embeds formal mathematical logic directly into AI architecture, moving from probabilistic guessing to absolute proof. When their AI solves a problem, it provides a verifiable proof with 100% certainty, establishing an open-source standard for how machines process objective truth.

Periodic Labs trains AI on the laws of physics and chemistry at a molecular level, creating 'Digital Twins' of matter that simulate reactions with perfect accuracy. Their AI-controlled robots conduct actual chemical experiments, discovering materials and medicines that never existed in digital form.

Conceptually, they curate or generate high-quality, verified data to build uniquely rigorous reasoning engines. They’re generating proprietary data through physical experiments rather than scraping what exists. You could say that they are manufacturing knowledge rather than mining it.

6. The Human Role & Open Source

There is a synthesis for intelligence. We spent centuries believing that the deepest questions about consciousness required the deepest thinking. That intelligence would be defined by rigorous debate, scientific breakthroughs, or philosophical revelations.

Instead, it’s being defined by who broke the right data.

Ultimately, the battle for the soul of AI is not a zero-sum game. What we are witnessing is the birth of two essential, complementary infrastructures for machine intelligence.

On one side, the proprietary path led by figures like Demis Hassabis offers a glimpse of High Science at a scale previously unimagined. His vision for AGI is already decoding the biological and mathematical secrets that have baffled us for millennia. And LeCunn’s model of physcial intelligence will likely also supercharge robotics.

On the other side, the open-source movement is ensuring that the foundation of this new era remains a public utility. By making intelligence an open standard, they ensure the technology remains in the hands of the many rather than being locked away behind corporate paywalls.

The fundamental tension of this very human debate on intelligence could be a defining story of our times. Or it could simply end in a whimper, to massive corporations sparring over industry dominance.

As you read this, an inflection point in the philosophy of intelligence could be devoured by code.

In an ideal world, that code would be written to serve us all. We should no longer be just searching for the nature of thought; we are building a collaborative future where the mysteries of the universe and the common sense of the street finally meet.

7. Conclusion: The futures of intelligence

What we actually need right now is a different understanding of sovereignty. If intelligence is the new infrastructure, then letting private companies define it isn’t a neutral move.

When the defaults for how we think and discover are locked inside closed models, our own judgment starts to erode. The quality of what we think is good enough becomes just that - good enough.

A public-utility approach to AI isn’t about fighting high science. It’s about making sure we get to participate, and that the human is not just in the loop but continues to define the goals and outcomes.

It’s also about keeping the foundations open so that people and institutions can decide how knowledge is made and used in their own backyards. This may seem unrealistic, but that is why the debate over sovereign AI needs to stop being just a conversation.

If the stakes weren’t so high, it would be laughable that Google and Meta should be “in charge” of the debate on AI futures. Sovereignty isn’t about building walls; it’s about the right to stay the author of the systems that are starting to think for us.

There are ways of laying out new blueprints:

• “Treat intelligence like the roads or the power grid” → Policymakers who still think AI is just another tech vertical instead of foundational infrastructure. The EU commissioners, UK’s AI Council, and national security advisors who haven’t connected the dots between cognitive infrastructure and actual sovereignty

• “Fund the open tools” → Governments with budgets, but also foundations, universities, and sovereign wealth funds who keep pouring money into OpenAI rounds instead of building alternatives. UKRI, NSF, Wellcome Trust—the institutions with enough capital to create genuine counterweights.

• “Demand to see under the hood” → Enterprises and institutions currently signing contracts with Anthropic, Microsoft, Google. The NHS trusts adopting AI diagnostics, the law firms using Claude for research, the universities embedding ChatGPT into curricula. Every procurement officer who’s accepting black-box systems for critical infrastructure.

• “Build things that can actually be challenged and changed” → Developers and researchers who’ve given up on open alternatives because “GPT-4 is just better.” The technical class who could be building on open architectures but choose convenience. Every startup that defaults to OpenAI’s API instead of harder open paths.

• “Don’t quietly outsource your own brain” → All of us. Every individual who’s replacing their own judgment with model outputs. The students, writers, analysts, decision-makers who are letting convenience erode their own capacity to think.

The power currently sits with a handful of massive corporate entities - not even governments. It is OpenAI, Anthropic, and Google DeepMind, to name a few. More precisely, it is their controlling investors and boards who’ve decided intelligence can be a private asset.

It should be mindboggling that so much of the weight of our intellectual futures sits with whoever controls the defaults that billions of people will use to think. There is a risk that this intelligence is hardened into infrastructure. There is a chance, however distant, that it will become a source of reality and no longer a set of tools. And there is a power in having a window into this future truth.

As always, I welcome comments & feedback. Please send me a link if you like it.