Emergence of superintelligence from AI hiveminds: how to make it human-friendly?

post by Mitchell_Porter · 2025-04-27T04:51:50.020Z · LW · GW · 0 comments

Contents

No comments

"Hive mind" may not be the ideal term. What I'm really envisioning is something modelled on the existing relationship between human users and frontier LLMs, but with the "users" actually just being LLMs too, or simple programs that mediate between LLMs. 

But first, consider OpenAI's o-series models, and their human user base. For any particular model, OpenAI presumably has a number of copies running on their servers, constantly fielding requests from hundreds of thousands of users. 

Now consider something like a "Deep Research" request. Let's say one copy of o3 can perform an average of 100 Deep Researches per day, and that there are a thousand copies of o3 on the servers. (Realistic estimates of these numbers would be very welcome.) 

For now, those requests are mostly coming from human beings, and if they are put to use at all, it is by private human individuals or by human organizations. 

However, given the scenario above (1000 copies of o3, each doing 100 Deep Researches per day, for a total capacity of 100,000 Deep Researches per day), it should not be at all difficult to have all those copies of o3 interacting via Deep Research reports. 

Let's say half the copies of o3 are set aside for human users; that's still 50,000 Deep Research reports per day, or over a million per month, generated in a dialogue among AIs. 

This is, more or less, the scenario in "AI 2027". It is a scenario at least as old as the idea that you could have swarms of intelligent agents interacting through the Internet. What is striking is that in the current o3 era [LW · GW], this scenario can actually be realized, with agents that are human-level in many respects. 

In fact, it's logical to suppose that OpenAI, Google, and Anthropic (at least) have already been running experiments of this nature for some time. When I think about the scale of what is already possible, the AI 2027 timeline seems unnecessarily slow. 

As an outsider to all the organizations with a chance of creating superintelligence, my idea of how to tilt the probabilities, in however small a way, towards a good outcome, is to look for potentially helpful ideas and insights that can be stated publicly, which might be noticed and assimilated by the private organizations that are actually building superintelligence. (Publishing those ideas and insights via the arxiv seems to be the best way to communicate them, since everyone keeps an eye on the arxiv.) 

Given that research swarms or research hiveminds look to be part of the pathway to superintelligence, I feel that insiders and outsiders alike need models of multi-agent safety [LW · GW], multi-agent capabilities research, and multi-agent safety/alignment research [LW · GW]. 

In fact, the situation is later than that. Any of these big companies, that possesses a frontier reasoning model and ample server space, could right now create a research swarm with the explicit goal of creating a human-friendly superintelligence. 

Just to demonstrate to myself that this is possible, I just had a short talk with Gemini 2.5 on the subject of what the exact goals of a human-friendly superintelligence should be. At first it gave me a number of reasons why that would be immensely difficult, but with some elementary cajoling, it quickly produced a hypothetical mission statement for such an entity. Once that was done, it was easy to ask it to think of remaining problems, solutions to the problems, and so on. 

I'm sure many of us have had conversations of that kind with an AI (e.g. here's the output of one I had, two years ago [LW · GW]). What's new is that the technology now exists for whole swarms of AIs to have these conversations with each other, on all aspects of the process, from high theory, to redesigning themselves and the architecture of their conversations. 

The idea that superintelligence was to be obtained via self-modifying and self-enhancing AI has also been around for decades. It may be that a self-modifying research swarm of reasoning models is an architecture for social cognition that is good enough to create a superintelligence via this kind of bootstrap. So it seems pretty urgent to think about how that kind of ecosystem of mind could evolve into something that is not just superintelligent, but truly human-friendly as well. 

0 comments

Comments sorted by top scores.