Reposting our paper from a few months ago as recent events again underscore its relevance:
"a theory of appropriateness with applications to generative AI"
We are currently hiring for an open research scientist position working on my team in multi-agent artificial general intelligence: boards.greenhouse.io/deepmind/jobs/…
In our latest paper we studied how even silly rules can benefit a society. They provide additional opportunities for all to practice the generic skills of enforcement and compliance upon which all social norms depend.
How can deep reinforcement-learning contribute to our understanding of complex social interactions like social norms and the evolution of culture? Learn more: dpmd.ai/spurious-norma… 1/
Very happy to launch Concordia, a library for generative agent-based modeling where you build simulations like tabletop roleplaying games! The game master (storyteller) is an agent who controls the environment where player agents interact, tracks state and can call external tools
Extremely excited to share our work on Generative Agent-Based Modeling. We are open sourcing Concordia, a platform for creating social simulations with language models.
Code: github.com/google-deepmin…
Paper: arxiv.org/abs/2312.03664
Social interactions are key to intelligence, but do artificial agents understand this? Introducing Melting Pot, an evaluation suite for reinforcement learning agents that tests their socio-cognitive skills: bit.ly/dm-meltingpot
OS: bit.ly/meltingpot-os#ICML2021
Very happy to announce that the Melting Pot challenge, a NeurIPS 2023 contest, is now open for submissions!
Test how well your agents cooperate with familiar and unfamiliar co-players in complex social scenarios!
"Don't eat fish on a Friday" seems like a silly rule, yet we are surrounded by many social norms that have no direct impact on welfare. Why? This paper uses multi-agent RL and shows how such "silly rules" can benefit society. Review video & interview here:
youtu.be/6dvcYx9hcbE
We are very happy to announce the Melting Pot contest! Is your RL agent good at cooperating in groups with unknown numbers of familiar and unfamiliar individuals? Enter the contest to find out!
Excited to announce the Melting Pot Contest has been accepted for NeurIPS Competition Track!
Sign up below for early access
iryc1zmdzmb.typeform.com/to/GrsAjGAW
A very cool paper on social norm learning here! Among lots of interesting things about this paper and model, the authors also made a new Melting Pot environment by bringing together elements from Clean Up, Commons Harvest, and Territory in a single environment. Great idea!
How can we ensure cooperation between (natural & artificial) agents? Humans do this via social norms that constrain uncooperative actions. In this new paper, @xuanalogue and I show how artificial agents can *learn* these norms from observation!
Link: arxiv.org/abs/2402.13399
The Concordia Contest is opening very soon now! The task is to design an agent decision-making process that operates in unconstrained natural language environments structured like table-top role-playing games! Agents have to cooperate in groups containing strangers (zero-shot).
We're excited to be partnering again with @apartresearch for a hackathon next weekend in the run-up to the Concordia Contest at @NeurIPSConf! The challenge: advancing the cooperative intelligence of language model agents. Sign up here: apartresearch.com/event/the-conc….
The Concordia NeurIPS Contest is now live!!
Find out if your agent architecture can discover cooperative resolutions to novel multi-player social dilemma situations in open-ended text-based worlds!
Submit your solution here: codabench.org/competitions/3…
The contest challenges participants to advance the cooperative intelligence of LLM agents in rich, text-based environments, based on the Concordia framework, which uses language models to create open-ended worlds similar to tabletop role-playing games: github.com/google-deepmin….
Since several articles arguing that scaling is slowing down are appearing,
OpenAI, Google and Anthropic Are Struggling to Build More Advanced AI bloomberg.com/news/articles/…
it's a good time to post our paper on this again:
A social path to human-like ai
arxiv.org/abs/2405.15815