The “shepherd” group of theorists is still keeping secret for their own use the proofs involving Yang-Mills, QCD and a “completion of the Standard Model” given to them by OpenAI. I’ve been hearing some odd things about the recent invitation-only KITP workshop organized by OpenAI-associated theorists. One of these is that the workshop led to the formation of a small secret group of US QG/QFT theorists that has a Slack channel and is in the process of getting preferred access to models and compute for themselves from OpenAI and the two other frontier labs (Anthropic/Google DeepMind). At the moment this group is US only, but they have formed a committee to consider possible non-US theorists to let in on the deal. While the long-term plan is something open to a much larger group, for now that’s definitely not the case.
I also finally figured out why some of the KITP workshop panels were listed as “academics only”: many of the participants felt that if they discussed their ideas in the presence of people connected to the frontier labs, these ideas would be stolen and exploited using compute/better models they had no access to.
There’s a major issue for scientific research here: who gets access to the best AI tools and who doesn’t? OpenAI has ruthlessly exploited its access to these tools to promote its financial interests. Others who want to compete are now stuck trying to figure out how they can get on the inside track.
Update: The membership of the group referred to above is still secret, but as of today they have a name, mission statmment and website. They’re the Alignment Institute for Fundamental Science, and they
are building a collaborative platform to offer researchers the resources and support to explore new styles of collaboration, and to explore new ways of doing science in a rapidly changing world.
At the moment they’re a pilot program, an unidentified small group focused on “the nature of quantum gravity and the mysteries of strongly interacting quantum field theory” but they promise to soon broaden this and ultimately to provide “broad access to AI resources.”


https://hep-open-release.org/
This was mentioned in an update to the previous posting. It’s now signed by 405 theorists. As far as I can tell, essentially no one competent to try and understand the Yang-Mills result thinks it’s acceptable for this group to keep the writeup secret for their own use. The AI gold rush is causing some weird violations of community norms, with the secret Slack group trying to set up a system where they control access to frontier lab resources another example.
The grant system would need a quick change: replace the now too big grants (1-2 M€ for ERC) with many more small grants.
@Peter Woit : I am not happy about these important results being kept secret. But is this seemingly in-your-face practice clearly beyond the established norms?
My understanding is this is pretty much everyday practice in big science: exclusive access period is enforced for data of JWST (https://jwst-docs.stsci.edu/accessing-jwst-data/exclusive-access-period#gsc.tab=0 ), CERN Atlas (https://opendata.cern.ch/record/413/files/ATLAS-Data-Policy-2021.pdf – embargo periods are mentioned) and many other facilities/organizations.
Alessandro Strumia,
Yes the grant systems of the world need to change to deal with the financing of AI compute for researchers. I’m curious right now how this is being financed: grants, personal funds, university resources, AI company programs, special deals with the AI companies…
Akhmeteli,
If you want to conceptualize these AI proofs as experimental results generated by the AI companies, I suppose you can. In that case the AI companies can analyze them however they want, including giving them to David Gross to work on. What’s wrong here is not them deciding to give the results to someone else to deal with, but the way Gross and others have handled this responsibility.
Even if you think restricting access to Standard Model and QG results to a small group is a good idea, isn’t it clearly a really, really bad idea to restrict results from all of physics to this same small group? Isn’t that what the mission statement seems to imply that this group wants? This is a huge resource grab, and it seems to me that you should let (say) condensed matter researchers have access to the condensed matter AI results, since they would understand them a lot better.
Peter Shor,
Sorry to not be clear. What the new group is trying to get access to is the use of latest models/compute to generate new results themselves. They’re a group of QG/QFT theorists, so would generate new results in this area.
The issue of what to do with results generated internally by the AI companies themselves (as part of testing, publicity stunts, or whatever) is something quite different. The reason I’ve written about these together is that they both involve groups of theorists trying to get in bed with the AI companies, but what they’re doing there is quite different.
The rationale of ensuring “broad access” doesn’t make sense: the AI companies already provide free compute to tens of thousands of scientists, and for everybody else, their most expensive offerings are a small fraction of a graduate student’s stipend. I think it’s clear the real motivation is to ensure certain people get credit. AI will do everything behind the scenes, but its papers will be released ceremoniously with the appropriate senior people (and a few favored junior people) stamped on the author list.
What’s less clear is why the labs are putting up with this. I suppose one reason is that they have a bunch of ex-hep-th people who still desire the admiration of their former community. Another reason is that hep-th is pretty good at presenting a unified message, so there’s the implicit threat that they could all say lab A’s models are much better if lab B doesn’t play along. However, neither of these factors really matters in the long-term.
Fluxonium: No, there is a big difference between the paid subscription rates for “ordinary people” and the corporate compute plans. I know a couple of physicists who have access to pools of tokens in the mid 5-digit (dollar) range, and some may have even vastly more, especially those who are on the payrolls of these companies (something many people don’t self-advertise). I personally know for a fact that this kind of resource can solve problems that are absolutely out of the range of any of us paying even $200/month. There is a big gap between the haves and the have-nots.
Fluxonium,
Those in the group are the one’s who
Have nothing to say that is really new
And then expect to get feeded by the beast.
When others refused to participate to the feast,
Who had something that could be of some interest,
Prefer not to participate to avoid getting their ideas best
Eaten bite by bite, and bit by bit by the beast.