The field published everything, and then some of it stopped

An openness that was unusual even among sciences

For most of the period in which the current technology was developed, the working assumption was that results were published, code was released and datasets were shared. Preprint servers meant a paper was public before it was reviewed, and often before it was submitted. The gap between a result existing and a result being available was measured in days.

This was not a moral position so much as a functioning system. The field organised itself around conferences rather than journals, which shortened cycles considerably, and around shared benchmarks, which made results comparable only if the details were disclosed. Openness was the mechanism by which the community could accumulate anything.

Industrial laboratories participated fully, publishing work with obvious commercial value. The justification given was recruitment and reputation, and it was rational: researchers wanted to publish, the best of them could choose their employer, and a laboratory that did not permit publication could not hire.

Release conventions grew more elaborate as stakes rose

A first shift came with the practice of staged release, in which a system was described and its weights withheld initially, then released after some period of assessment. This was contentious at the time, criticised by some as unnecessary caution and by others as a publicity exercise, and it established a template that others followed.

Alongside it developed a vocabulary of intermediate positions: releasing weights without training data, releasing a smaller variant, releasing under terms restricting certain uses, providing access through an interface rather than a file. Each of these is a different point on a spectrum that had previously had two ends.

Documentation practices developed in the same period, with structured descriptions of what a model was trained on and what it should not be used for. These were proposed by researchers concerned about deployment rather than imposed from outside, and adoption has been uneven.

The technical reports thinned out

As systems became products, the documents accompanying their release changed character. Where a paper would once have described architecture, data and training procedure in enough detail to be attempted, some now describe evaluation results and capabilities while explicitly declining to give the rest.

The reasons offered are competition and safety, in varying proportions and with varying credibility. The competitive reason is straightforward and needs no defence; substantial investment produces something worth protecting, which is how commercial research has always worked in every other industry.

The safety reason is genuinely contested. Some hold that detailed disclosure of capable systems assists misuse and that withholding is responsible. Others hold that it prevents independent scrutiny, concentrates the ability to evaluate in the hands of those with the most interest in favourable evaluations, and is difficult to distinguish from the commercial motive.

A counter-current ran the other way at the same time

The same period saw a large volume of weights released publicly, by organisations including some of the same ones publishing less about their most capable systems. Repositories hosting models and datasets became substantial infrastructure, and a considerable amount of work is done on openly available systems.

So the simple narrative of a field closing is wrong. What happened is a separation: broad access to capable-enough systems increased, and detailed access to the most capable ones decreased. Those are different axes and conflating them makes the situation harder to describe accurately.

The separation has consequences for who can study what. Researchers examining how these systems work internally need weights, and the systems they can examine are not the ones policy is written about. That mismatch is a real problem and is acknowledged across otherwise opposed positions.

Whether the older norm was sustainable

It’s worth asking whether the open culture could have survived contact with this much money regardless of anybody’s intentions. The history of other fields suggests not: pharmaceutical chemistry, semiconductor design and cryptography all had periods of relative openness that narrowed as commercial and security stakes rose.

If that pattern holds, the interesting question is not how to restore the previous arrangement but what institutions can supply the verification that openness used to supply incidentally. Independent evaluation bodies, structured access for researchers and disclosure requirements have all been proposed, and none is established.

The honest position is that a norm which produced a great deal of the current technology is weakening, that the reasons for it are not entirely cynical, and that nothing has yet replaced the function it served.