1. The bulk of scientists and research output is normal science. They are the work of science. Civilisational scientific progress, output that moves fields forward is a small part of this – say a top 10%.

2. This is not about the type of output, it could be work on replication of other results. I guess it’s more about power laws, they are everywhere.

3. Top scientists want to work with the best in their field. This narrow body of work which retroactively is determined as most significant is often compounded if everyone doing it is placed in close contact with each other to allow rapid sharing of information.

4. The closedness of science is not a consequence of an individual actor’s motives and preferences. It is a result of how incentives are set up across the system. Knowledge is power, power is valuable, capital is limited, scientists are humans.

5. I am not confident that the future of science is open or public, largely private, as evident through various social currents.

6. I am unclear on what the end state would be but I believe we will see competing pushes from both ends – frontier labs and companies like Edison hiring the best scientists in house (as you can already see with Anthropic Fellows program and OpenAI’s spree of hiring scientists) and open science (as is showcased by say how quickly we got an OpenEvolve and ShinkaEvolve after AlphaEvolve).

7. That said, the specific nature of open science today, the infrastructure of academic labor markets and private/public research fellowships and grant seeking, the push to publish and the attempt to systematise the production of science with metadata and citation count is historically contingent.

8. This is presumably, an artefact of the last few decades of post Cold-War Neo-Liberal nation state and global trade driven world order.

9. Metadata for attribution and metadata for understanding of research / results sharing are different. The latter is determined by downstream use-case.

10. I am unsure if the age ahead of us includes authorship attribution in any meaningful way considering much of open science centred around reproducibility and low-hanging fruit extensions will be done by AI agents, and frontier, human science will be kept private.

11. Metadata for research understanding and results sharing is determined by the task and preference more than by any objective or principle driven by open science ethos. Ontology building is a part of an individual agent’s scientific method / process, and not in line with some god’s eye view definition of capital-S science.

12. Would argue that for much of science done today is already only partially public. This is domain specific but it is not just a consequence of institutionalised publishing being rent seeking.

13. Instead, institutionalised publishing maintains the signals of credibility / credentials which feed into academic labor markets and serve the function of curating the large volumes of research produced in the pursuit of ‘public’ and ‘open’ science contributions.

14. Builders of AI agents and systems can work directly with scientists and these private publishers instead of only relying upon an infinite resource of humanity’s knowledge, or open science.

15. Diversity in science is less about volume and more about perspectives and good science is always epistemically diverse, so the top science set which is necessary for understanding and moving the field forward is already compositionally diverse. This is also evident through the interdisciplinary teams at frontier AI labs and research teams doing frontier research.