https://github.com/paradigms-of-intelligence google team pushing social intelligence, collective intelligence, continual learning, etc. against singleton. the links to publications here are very good. super relevant.
- “none of us individually know how to put a person on the moon”
- collective intelligence > individual intelligence -> sociality, coordination architectures, political system, economic system, interaction models, etc
- permissionless economics/finance?
- https://arxiv.org/pdf/2602.16301 Feb 2026. first example ive seen of MARL applied to LLMs. havent well understood it yet. https://gemini.google.com/app/3162efb27df1393b this is NOT on LLMs. its on sequence models, but not sequence models previously trained on next token prediction https://gemini.google.com/app/524e92b44456b5f7. still, its the practical implementation of MUPI, showing that training with MARL on a diverse set of agents produces emergent cooperative behavior from mutual extortion ^b98596
- would love to discuss this paper with relevant people but hard to access them. don’t know where they are. relates to potential network effect value prop of PDEV
- “Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs” (Kim et al., Mar 2026) The core finding is that tracking someone else’s mind and asserting your own are structurally decoupled within the model.
- https://arxiv.org/pdf/2603.02960 Architecting Trust in Artificial Epistemic Agents Mar 2026. experiment idea 1 https://gemini.google.com/app/7e192ccff8ca369a relates almost 1:1 with issues with old Slate multi-agents. basically quantifying the epistemic drift of an emergent network over time Experiments 2 and 3 are also good Interaction > ^2bff57 ^b203be
- also seems like these concepts would be useful for PDEV since a lot of the product is meant to legibilize process while also interacting with AI heavily to bounce ideas around
- does this paper explain the difference between context as a message vs context as a program? i read this somewhere but forget where