The Spark: AI Is Not a Race. It Is the Largest Collaboration in Human History.

Something is happening to time itself. I see it in the work I do every day, and I cannot look away from it.

Not as a metaphor. As a measurable, verifiable, documented reality. The distance between a question and its answer is collapsing. The distance between an idea and its execution is shrinking. The years it once took to move from discovery to deployment are becoming months. The months are becoming weeks. In some domains, the weeks are already becoming days.

I have studied technological revolutions closely, and I believe this one is different in one specific way: the pace of its own improvement is itself accelerating. Every other revolution had a ceiling defined by physics or economics. AI's ceiling, at this moment, is not yet visible. And that changes everything about how we should think about what we are building and for whom we are building it.

The instinct, when something this powerful emerges, is to treat it as a weapon in a competition. Nations are doing this. Companies are doing this. The headlines are written in the language of race and rivalry. Who will build the most capable model. Which country will win the AI era. Whether open source will beat closed systems or closed systems will absorb open source.

That framing is understandable. It is also wrong. And understanding why it is wrong is, I believe, the most important intellectual task of this decade.

The Compression of Time

In 1972, Christian Anfinsen won the Nobel Prize in Chemistry for work on protein folding. He showed that a protein's three-dimensional structure, which determines its biological function, is encoded in its amino acid sequence. The implication was profound: if we could predict that structure from the sequence, we could understand biology at its most fundamental level.

For 50 years, the world's best structural biologists could not solve this problem computationally for most proteins. It was considered one of the hardest problems in science.

In 2020, AlphaFold solved it. Not approximately. With accuracy comparable to laboratory experiments. Within 18 months, AlphaFold 2 had predicted the structure of over 200 million proteins, covering nearly every known protein in existence, and released the entire database for free to the global scientific community.

Fifty years of effort. Solved in months. Given to the world at no cost. That is what AI does to time.

Consider what this compression means across every domain simultaneously. Drug candidates that took 5 to 7 years to identify in the laboratory are being discovered in 18 months with AI-assisted molecular design. Insilico Medicine brought a drug candidate from AI-generated hypothesis to Phase II clinical trials in under 4 years, a pace that would have been considered impossible a decade ago, as documented in Chemical Science in 2023 (Ren et al., doi:10.1039/D2SC05709C).

NVIDIA's FourCastNet produces a 10-day global weather forecast in approximately 2 seconds on a single GPU. Traditional numerical weather prediction on supercomputers takes hours. Climate scientists can now run thousands of model scenarios, exploring the full envelope of possible futures, in the time it once took to run one.

GitHub's research with Accenture found that developers using AI coding assistance complete tasks 55% faster than those working without it. That is not a productivity statistic. That is a statement about how much human creative energy is being freed from mechanical execution.

The compression is not happening in one field. It is happening everywhere, simultaneously, at a pace that compounds on itself. Each breakthrough enables the next. Each model improvement accelerates the tools used to build the next model. This is why the next decade will not feel like a continuation of the last. It will feel like a different kind of time entirely.

Open Source and Frontier: Not a War, a Rising Tide

The narrative of open versus closed AI is the most seductive false binary in technology today. On one side: Meta's Llama family, Mistral's models, Falcon, DeepSeek, the thousands of fine-tuned derivatives published to Hugging Face every week. On the other: the frontier labs whose most capable models run behind APIs and paywalls.

The narrative says one must win and the other must lose. The data says something entirely different.

In July 2024, Meta released Llama 3.1 405B. On the MMLU benchmark, a test of reasoning across 57 subjects, it scored higher than GPT-4 as originally released in March 2023. An open-weight model, freely downloadable, running on commodity hardware, surpassed what was considered the frontier just 16 months earlier.

But GPT-4 did not stand still. The frontier moved too. And the movement of the open models forced the frontier to move faster than it otherwise would have. The competition between open and closed is not a war. It is a ratchet. Each side pulls the other forward.

What does this mean for the world? It means that the most capable AI is no longer exclusively the property of the three or four organizations with the largest compute budgets. A researcher in Nairobi, a startup in Jakarta, a university lab in São Paulo can access models today that match what only the wealthiest technology companies could build two years ago. The cost of capability is falling at a rate that has no historical parallel in the history of technology.

The cost curve is equally dramatic. In 2020, accessing GPT-3 via API cost approximately $60 per million tokens. By 2024, frontier model access through Claude 3 Haiku and GPT-4o Mini had dropped below $1 per million tokens for many use cases. Open source models, self-hosted, run at effectively zero marginal cost per token. The price of intelligence, as a computational service, has fallen by orders of magnitude in four years. It will continue to fall.

The Industries Being Rebuilt

When people say AI will disrupt industries, they reach for the vocabulary of destruction. Disruption. Displacement. Replacement. That vocabulary misses what is actually happening. AI is not destroying industries. It is rebuilding them from the inside, expanding what they can do, reaching people they could never reach, and solving problems that were previously considered unsolvable at scale.

These are not speculative futures. They are current deployments, documented and measurable. The McKinsey Global Survey on AI in 2023 found that 55% of organizations globally had adopted AI in at least one business function, up from 50% in 2022 and a fraction of that in the years before systematic tracking began. The adoption is accelerating in every sector, every region, every size of organization.

The most significant thing is not the adoption rate. It is the distribution. The sectors where AI is creating the most profound impact are not the already-wealthy sectors of finance and technology. They are healthcare in low-resource settings, education in underserved communities, agriculture for small farmers, financial services for the unbanked. AI is following a pattern unlike any previous technology: it is reaching the most underserved populations faster than the most privileged ones, because the marginal cost of deploying a model to a new user approaches zero.

The Individual: What One Person Can Now Do

The framing of AI as an industrial force, reshaping sectors and nations, obscures what I consider the most radical thing it is actually doing. It is changing what a single human being can accomplish.

Not a team. Not a company. One person.

A first-generation university student, the first in her family to go to college, studying medicine in a city where the professors are overwhelmed and the library is inadequate, now has a research assistant who has read every paper ever published in her field and can explain any of them at any level of depth she needs. She has a study partner available at 3am. She has a diagnostic reasoning tool she can practice with a thousand times before her first clinical placement.

A small business owner in a town of 5,000 people who cannot afford a lawyer, an accountant, or a marketing firm now has access to document review, financial modeling, and campaign creation that previously existed only for companies with six-figure professional services budgets.

A researcher in a field where only 12 people in the world have deep expertise can now synthesize 500 papers in an afternoon, identify the five most important open questions, and generate hypotheses that would have taken months of literature review to surface manually.

GPT-4, tested on the Uniform Bar Examination, scored at the 90th percentile, according to OpenAI's technical report published in 2023. The same model passed the US Medical Licensing Examination at a passing threshold. These are not demonstrations of novelty. They are demonstrations of access. Legal reasoning and medical knowledge, once gated behind decades of education and geography, are now accessible to anyone with a smartphone and a connection.

This does not replace lawyers or doctors. It means that for the first time in human history, a person without access to a lawyer or a doctor has something better than nothing. In a world where 1.4 billion adults are unbanked, 3.5 billion people lack access to safe surgery within two hours of travel (The Lancet Commission, 2015), and 300 million children attend schools without qualified teachers (UNESCO, 2022), "something better than nothing" is one of the most important sentences in the history of technology.

The Global Compact: Why the Only Winning Move Is Shared Progress

Fifty countries have national AI strategies as of 2024, according to the OECD AI Policy Observatory. The language of many of these strategies is competitive: leadership, dominance, strategic advantage. This language is understandable. It is also a misreading of how AI actually creates value.

AI models improve with data. The more diverse the data, the more robust the model. A medical AI trained only on data from one country, one demographic, one healthcare system will fail in ways that are invisible until they are catastrophic, when it encounters a patient from a different context. The AI systems that will be genuinely capable across the full diversity of human experience are the ones trained on the full diversity of human data. That requires collaboration, not competition.

AI safety requires global coordination. The risks that come from advanced AI systems, misalignment, misuse, and unintended consequences at scale, do not respect national borders. A misaligned system deployed in one country affects every country. The frameworks for responsible AI, the standards for evaluation, the norms for deployment, the governance structures for accountability: these require the same kind of international coordination that the world built for nuclear technology, aviation safety, and financial regulation. No single country can build these frameworks alone. No single company can be trusted to build them for everyone.

The countries and institutions that understand this earliest will have the greatest influence on how the global AI ecosystem develops. Not because they are the most powerful, but because they are the most collaborative. The standards and norms being written right now, in research labs, in policy offices, in international bodies, will determine what AI systems are required to do for a generation. The voices in those conversations matter.

India, with the world's largest democratic population and one of the most sophisticated digital governance experiences, has standing in these conversations that it has not yet fully claimed. Africa, with its 1.4 billion people and the fastest-growing AI research community on the continent, has perspectives on what intelligence should do for underserved populations that the field desperately needs. Europe, with its leadership on data rights and algorithmic accountability, is writing norms that the rest of the world is watching and beginning to adopt.

This is not idealism. It is a strategic description of how global infrastructure gets built. The internet's protocols were written by a relatively small group in the 1970s and 1980s. The world has been operating on those decisions for 50 years. AI's foundational protocols, its evaluation frameworks, its safety standards, its deployment norms, are being written right now. The people writing them will have influence that lasts generations.

Consciousness, Empathy, and the Question That Matters Most

There is a question that the technical community often defers and the policy community rarely asks precisely enough: what is AI for?

Not in the abstract. Specifically. Each system, each deployment, each design decision is an implicit answer to this question. An AI system designed to maximize engagement on a social platform has a different answer than one designed to help a child learn to read. An AI system designed to optimize advertising revenue has a different answer than one designed to help a clinician make a better diagnosis. The technical choices are downstream of the values. Always.

The AI systems that will endure, that will genuinely serve humanity across the breadth of human experience, are the ones built with depth of purpose. Not just intelligence. Consciousness about what they are doing and for whom. Empathy baked into design, not added as a feature after the fact. Diverse builders who bring the perspectives of the people who have been most failed by previous technological systems, because those perspectives surface the failure modes that the historically privileged builders never encounter.

This is why the global embrace of AI is not just a nice idea. It is a technical requirement. The models that reflect only the data, values, and assumptions of a narrow group of builders will fail in ways their builders cannot predict, for populations their builders did not consider. The models built with global input, global data, global perspectives on what human flourishing actually means in different contexts will be more capable, more robust, and more genuinely useful.

Open source serves this purpose. When a research team in Lagos can download a model, fine-tune it on local data, evaluate it against local needs, and publish their findings back to the global community, the global community learns something it could not have learned any other way. The flow of knowledge is not one-directional. It is a network, and networks become more valuable with every node added.

The Moment We Are In

In 1440, Gutenberg printed the first book with movable type. Historians of print culture estimate that Europe held roughly 30,000 manuscript books at that time. By 1500, printed books numbered in the millions (Elizabeth Eisenstein, The Printing Press as an Agent of Change, Cambridge University Press, 1980). The acceleration of knowledge that followed changed science, religion, politics, and the structure of human society in ways that took centuries to fully unfold.

We are at a moment of comparable magnitude. Not because AI is like the printing press. Because the compression of time AI enables is at least as profound, and it is happening across every domain simultaneously rather than sequentially.

The people alive right now are the ones who get to decide what this moment becomes. Not in the aggregate, through the drift of market forces and geopolitical competition. In the specific: through the choices made in research labs, in policy offices, in investment decisions, in the problems that individual engineers and entrepreneurs choose to work on. I have made my choice. And I believe the choice in front of all of us right now is the most consequential one our generation will make.

The choice to build AI that serves the full spectrum of humanity rather than the already-served top of it. The choice to share models, datasets, and benchmarks rather than hoard them as competitive assets. The choice to include the voices of the most affected communities in the design of the systems that will affect them. The choice to treat AI governance as a global coordination problem rather than a national advantage problem.

These are not altruistic choices in tension with strategic ones. They are, increasingly, the same choice. The AI systems that will be most capable are those trained on the most diverse data. The AI governance frameworks that will be most legitimate are those built with the broadest participation. The AI companies that will be most trusted, and therefore most valuable over the long term, are those that can demonstrate genuine alignment with human flourishing rather than just the flourishing of their shareholders.

The race framing will persist for a while. It is too convenient and too emotionally legible to disappear quickly. But the underlying dynamic is already visible to those who look closely enough.

The countries that are winning in AI are not the ones that are hoarding their models. They are the ones producing the most researchers who publish openly, attracting the most global talent, and contributing the most to the shared infrastructure of tools, benchmarks, and frameworks that the entire field runs on.

The companies that are winning are not the ones with the most closed systems. They are the ones building the most trust with the most users, which requires the most genuine alignment with user needs, which requires the most genuine humility about what they do not yet know.

This is the spark. Not the technology itself, extraordinary as it is. The spark is what happens when the most powerful cognitive tool ever built meets the full range of human creativity, curiosity, and need. That meeting is not complete. It has barely begun.

The question is not whether AI will transform the world. It already is. The question is whether the transformation will be shaped by the people who understand that the only version of this story that ends well is the one where everyone is writing it.