Dario Amodei describes a possible future of powerful AI as a "country of geniuses in a datacenter": many highly capable instances working independently or together. It is a description of capabilities he anticipates, rather than a claim about what today's models already achieve. Machines of Loving Grace.
The image makes me think about the country of people outside the datacenter. They are spread across the United States and the wider world, doing research, writing explanations, solving problems, and leaving records from which AI can learn. Their intelligence supplies much of the material through which machine intelligence develops.
That connection raises a question about credit. I think people should receive accurate attribution for what they contribute. I also think we can become so preoccupied with securing recognition that we prevent the contribution from doing anyone else any good.
The intelligence outside the datacenter
The work that feeds AI comes from a much larger community than the people who build a model. Consider the Pile, a language-model training corpus assembled from 22 collections of text, many drawn from academic and professional sources. Even this single dataset brings many communities' work into one training process. The Pile.
And the people who produced that work were already drawing on others. A researcher inherits methods, a programmer uses concepts developed over decades, and a teacher explains ideas made possible by generations of inquiry. Individual contributions are real, but they happen inside a shared intellectual history.
My point about the country outside the datacenter concerns this inheritance. It does not require reducing an AI model to a store of copied answers. Learning from other people's work can support new combinations and new capabilities. The important fact is that the resulting system has an ancestry.
I want us to take that ancestry seriously without making every use of knowledge wait for a final accounting of everyone who helped create it.
When a secret formula protected a career
There is a historical parallel in the Renaissance disputes over solving cubic equations. A method could give its holder an advantage in mathematical contests, making secrecy valuable alongside the mathematics itself.
In 1535, Niccolò Tartaglia defeated Antonio Maria Fior in a contest involving cubic equations. In a later exchange recounted by Tartaglia, he declined an offer to publish his method under his name in Cardano's book, saying he intended to publish it in his own work. He also resisted providing solved examples because Cardano might infer the method from them. Tartaglia's account of the dispute.
The stakes included employment. In 1548, Tartaglia was asked to compete against Lodovico Ferrari to establish his credentials for a lectureship in Brescia. His mathematical standing had consequences for his livelihood. MacTutor biography of Tartaglia.
The dispute also involved promises of secrecy and competing claims of discovery. It would be too simple to turn it into a story about one selfish mathematician. What interests me is the incentive: if knowing something that others cannot do secures my position, making that knowledge widely usable can threaten me.
The solution then has two kinds of value. It helps people solve a problem, and it protects the standing of the person who controls it. Those values can pull in opposite directions. Everyone else benefits from being able to build on the method, while its holder may benefit from keeping them dependent.
Better intelligence could mean better attribution
My expectation is that sufficiently advanced AI will make the history of ideas much easier to investigate. If we reach the superintelligence associated with the singularity, I expect it to trace connections among papers, drafts, code, experiments, and discussions far more thoroughly than people can today.
In that future, I expect accurate credit to become a routine consequence of understanding how a result came about. A prominent name on the final paper would be only one piece of evidence. A system could investigate who supplied a crucial observation, who developed a method, and who enabled others to finish the work. It could also estimate how much particular contributions mattered to a defined outcome.
Restricted versions of this problem already exist in machine learning. Data Shapley assigns value to training data according to its contribution to a model's performance. The value depends on the dataset, learning algorithm, and performance measure; it is not a universal measure of worth. Data Shapley.
That distinction matters when imagining percentages of credit. Two people can both be indispensable to a result. Knowing that fact does not, by itself, tell us how to divide recognition between them. We still have to specify what we are measuring, while a better system could do much more of the difficult historical and causal analysis.
This is a forecast about what intelligence may make possible. Reliable attribution would still depend on evidence of the work, and future capability alone would not ensure that an institution uses it fairly. Keeping dated drafts, contribution records, and research histories gives future investigators something to examine.
The prospect nevertheless changes how I think about the urgency of recognition. We may be delaying the development of systems that could recognize contributions more accurately because we insist on settling credit with the limited tools available now.
The dragon keeps eating while we negotiate
CGP Grey's adaptation of Nick Bostrom's The Fable of the Dragon-Tyrant gives this problem a more human scale. A society becomes accustomed to a dragon consuming its people, until a technological effort offers a way to kill it. Near the end, a man discovers that his father is on the last train to the dragon. The possibility of finishing sooner becomes painfully concrete. CGP Grey's adaptation; Bostrom's original fable.
Now imagine adding an obstacle of our own: people with useful knowledge withhold it while negotiating whose name will be attached to the solution. The weapon takes longer to build because its contributors are still settling recognition.
That is my extension of the fable. The dragon does not wait for the argument to end. The people who pay for the delay are not limited to the people arguing.
This is what troubles me about holding useful research back to protect a name on a paper or an exclusive position. If sharing a result would let others solve an important problem sooner, withholding it for recognition has a cost beyond the researcher's career. The desire for a legacy is being weighed against things other people might gain from the work.
The analogy does not establish that AI will defeat aging or that any particular attribution dispute has delayed a cure. It makes the ethical question harder to avoid: when useful progress is available, what justifies making others wait for it?
Preserve the contribution and let the work proceed
Accurate citations and contribution records already help people understand, check, and extend research. Often, recording credit takes very little time and can happen alongside the work. That is worth doing now.
The problem arises when recognition becomes a condition that must be fully negotiated before others can benefit. I object to making knowledge scarce for the purpose of preserving the discoverer's status. A solution can become more valuable to society precisely when it becomes less exclusive to its author.
There is a real difficulty here: researchers need jobs, funding, and a way to continue working. Someone protecting their position may be responding to a system that makes sharing costly. Understanding that incentive helps explain the behavior; it also shows why changing incentives matters. Institutions should make prompt, useful contributions a route to security and recognition.
The same standard should apply to the organizations using the knowledge. A company cannot persuasively invoke shared human intelligence only when asking other people to contribute, while treating the benefits as exclusively its own. The argument concerns what helps knowledge become useful, whoever is in a position to withhold it.
My view is that we should preserve the evidence, attribute what we can, and allow the work to continue while improving the accounting. Future intelligence may let us see contributions that today's authorship conventions miss. We should help create the conditions for that recognition rather than obstruct them to protect the recognition we already know how to obtain.
A legacy can consist of what others were able to do because a person shared a solution. If the work can help us kill the dragon, keeping it to ourselves until our names are secure gives the dragon more time.