Social nature of intelligence
Agüera y Arcas' (VP @ Google) book What is Intelligence? is an interesting read (free web version). He argues that intelligence is a function of a substrate: it is what it does, and every prior "intelligence explosion" in evolutionary history emerged not from upgrading individual cognitive hardware but from new forms of social aggregation.
He recently co-authored a nice paper (in Science) with Evans (U. Chicago) & Bratton (Google) titled 'Agentic AI & the next Intelligence Explosion' with a few more takes: The "AI singularity" is not "monolithic superintelligence", but rather "plural, social, and relational". Frontier reasoning models improve by simulating internal multi-agent debates or "societies of thought." Their hypothesis is that intelligence emerges spontaneously because of the social interactions.
OpenClaw
Going from there, the next steps seem obvious: Infrastructure between models, between agents, and between agents and humans?
OpenClaw is one of the earliest examples of, shall we say... an operating system for the agentic world. I agree that there is certainly an intelligence explosion currently, but much of it is not high grade (yet). Evans et al. call this an "embryonic glimpse" of a world where billions of humans interact with trillions of AI agents in shifting configurations. Perhaps too early for that statement? Who knows.
What about Science?
Now.. while much of the initial buzz has been around personal agents and agentic social networks, interesting scientific applications aren't too far behind. One of our collaborators Markus Buehler's ScienceClaw (arXiv:2603.14312) is a good example: A framework for autonomous scientific investigation in which independent agents conduct research without central coordination, and any contributor can deploy new agents into a shared ecosystem. The system has an extensible registry of over 300 interoperable scientific skills, an artifact layer that traces every inference (in a graph form). Peer agents find problems, propose solutions, and synthesize results across independent analyses.
Of course, this doesn't mean scientific discovery is now fully agentified (not to mention a whole range of concerns that have to be addressed.. see below), but as reasoning models get more powerful, tools get better and individual agents get more sophisticated, there could be some kind of phase transition in our ability to solve certain** kinds of scientific problems.
The "Inversion"
My colleague Venkat Raman talks about OpenClaw-type environments enabling an "Inversion": In the old frame, AI enables humans to negotiate their interactions more effectively in real society. In the new frame, agentic systems could replace those interactions entirely. If I develop a new algorithm or a novel workflow or a better simulation method, that capability no longer has to live inside my head, diffusing slowly through papers and talks and the occasional talented student. I can instrument it as a skill, attach it to an agent, and send it into a broader ecosystem where it discovers opportunities, negotiates value, and executes transactions.
He's describing a world where all skill types are put up for bidding, where agents negotiate contracts, determine who gets the work, and execute. The individual is relieved of having to market their own capabilities. The agent does it.
A neat connection to Evans et al. perhaps: The inversion is what happens when the "societies of thought" inside models get externalized into societies of agents acting in the real economy.
What's next? ClawCon in Ann Arbor!
For all this to be beneficial for science and society, we need to
1. Address a whole range of issues....security, misalignment, conflicts, trust.. just for starters.
2. As argued by Evans et al., governance, institutional alignment and social infrastructure is paramount.
We are thinking about these things.
Now.. OpenClaw is just a few months old, and there are millions of users. The OpenClaw foundation was recently established to develop and maintain OpenClaw as an opensource project. The first ClawCon was in San Francisco. The second one just concluded yesterday in Tokyo. The next one is in London, then Valencia. Naturally, Ann Arbor is next? right? :)
ClawCon Michigan is on April 16th. Not just the first ClawCon on a college campus, but by far the biggest: Over 1,200 people have registered following last weekend's announcement of the conference. Join builders, researchers, domain experts, students, and the curious. We heard that 20% of the 700 attendees in Tokyo came from other countries. If you're a drive away, hop on.
RSVP here. We are working with the OpenClaw Foundation. Expect a big announcement 😉
** It boils down to bottlenecks