Delayed or unequal access. In cases where designated mathematicians see material before everyone else, they get a head start on problems others may have worked on for years. Other mathematicians may have valuable contributions to make based on the material but may not have access to it, and may be dissuaded from working on it.
The remaining reasons apply even if all the material is released publicly at once.
The Advisory Group on Mathematics and Artificial Intelligence (AGMAI) recommends that AI companies stop designating mathematicians to receive their output, “often before anyone else”, and lists four more reasons: denied agency, biased selection, borrowed reputation, and that it is unnecessary.
This is the opposite of progressive disclosure. There, the circle widens from the authors to “mathematicians known to be working on the problem”, and only then to the public. Only yesterday I said that, had it been done right for OpenAI’s release, OpenAI would reach these people first. Those are designated mathematicians.
Their emphasis is on openness, the same end of the spectrum as math wanting the whole chain open: nobody should see the material before everyone else. But there’s always someone having access to the SOTA internally, the lab itself, and its internal model is superior to what’s publicly available by construction, since the frontier is always a model still in training, one generation ahead of the public one, as I argued in the economy post.
And the recommendation keeps that. Pathway 2A, for scholarly articles produced by AI lab employees, means the lab’s own staff see the material first and write it up. So the way I read it, if you want privileged access to the SOTA AI math breakthroughs, you should work for an AI lab. If you aren’t their employee, you haven’t earned that privilege, and should be equal to everyone else.
Their ideal scenario is that the labs don’t test on open math at all: their September 29 guidelines “ask them to stop testing advanced mathematical problems on proprietary models”. In the economy post I explained why that ain’t gonna happen. The same guidelines ask that significant results be released “as soon as possible”. So without designated mathematicians, what is left for an AI lab is basically to release everything, once found, as soon as possible, publicly and at once.
Would a progressive disclosure with only two layers work? That is, before the public release, the results go to a private network that any mathematician can join. Would that make both parties happier?
The record in progressive disclosure includes the mathematicians who were originally working on the problem, such as Buckmaster on Navier–Stokes. To record the progress of each, and know who worked on it before OpenAI in order to assign credit, the network has to include them, and at the same time it can’t discriminate against other mathematicians. So anyone who wishes to join should be able to. The problem then is that it is not really different from being public on day 1. My original idea was that only the top experts known to be working on the problem are invited, to make sure the release is coordinated, as I borrowed from coordinated vulnerability disclosure in cybersecurity.