One billion
agents.
One experiment.

Most LLM-based social simulations top out around a million agents, and reaching even that scale usually needs dozens of GPUs running for weeks. A team spanning Zhongguancun Academy, the Zhongguancun Institute of Artificial Intelligence, Shenzhen University, Shanghai University of Finance and Economics, and Tsinghua University built a framework called Light Society to push three orders of magnitude past that ceiling, then used it to study two classic problems in social science: trust and reciprocity, and how opinions spread through a network.

The paper, posted to arXiv in June 2025 and revised in June 2026, argues the real bottleneck in LLM-based social simulation was never behavioral realism. Language models already generate agents with personality, memory, and context-sensitive responses. The bottleneck is computational: every added agent means more model calls, more storage, more latency. Light Society’s contribution is a set of engineering choices that collapse that cost curve without collapsing the fidelity of what gets simulated.

Each agent carries a static profile, an internal status that evolves over time such as memory, beliefs, and emotion, and an external status reflecting observable conditions. Everything that happens, a message sent, a relationship formed, gets logged as a timestamped event in a shared queue. The efficiency comes from a separate stack layered on top: semantic prompt caching that reuses answers to structurally similar queries, knowledge distillation that trains compact surrogate models to approximate routine decisions, and a mixture-of-models router that sends each operation to whichever backend fits the moment’s fidelity and resource tradeoff.

The first demonstration built 96,125 agent profiles from the World Values Survey and dropped them into a classic trust game: a trustor sends part of a $10 endowment, the amount triples in transit, and the trustee decides how much to send back. Agents assigned upper-class or higher-education identities sent noticeably more than lower-class or less-educated ones, and trustees returned roughly proportional to what they received. The more striking result showed up when the team varied population size directly: the trust gap between younger and older agents widened, and the confidence intervals around it narrowed, as the simulated population grew larger. That is close to the opposite of what usually happens in statistical sampling, where a bigger sample just narrows the error bars around an effect that was already visible.

The second study generated a synthetic scale-free network of one billion nodes and seeded a single statement, that AI automation will lead to mass unemployment, across the population, with the most connected 20 percent of nodes designated influencers. When influencer opinion leaned toward agreement, the wider population shifted toward agreement over successive rounds. When influencers leaned toward disagreement, the population shifted the other way. When influencer opinion already matched the general population, the shift was minimal and trended toward neutral. Higher-education, higher-income agents proved both more persuasive and more resistant to being persuaded themselves, with influence success peaking among agents holding postgraduate degrees and high income.

The team is explicit that this is a demonstration of feasibility, not a validated model of any real society. The agent profiles come from survey data rather than tracking of real individuals, and none of the efficiency or reproducibility figures have been checked outside the authors’ own reporting. Richer memory, longitudinal behavior modeling, and integration with real-world data are the stated next steps toward simulations that could inform policy questions rather than just demonstrate scale.

A team at Stanford University and the Arc Institute has done something no one had documented before: they asked an AI model to write a complete viral genome, letter by letter, and watched it produce a virus that worked. The study, led by Stanford bioengineering researcher Samuel King and chemical engineering professor Brian Hie, was published Thursday in the journal Science. Of 302 AI-designed genomes the team synthesized and tested, 16 turned out to be viable bacteriophages, viruses that infect bacteria rather than humans, capable of infecting and killing E. coli.

The tool is Evo, a genome language model built the way large language models are built, except its vocabulary is the four letters of DNA instead of words. Evo 1 trained on 2.7 million genomes; its successor, Evo 2, scaled that training to trillions of nucleotides pulled from across the tree of life. For this study, the team fine-tuned both models on 14,266 genomes from the Microviridae family and gave them a starting point: PhiX174, a bacteriophage that has been a standard laboratory tool since the 1970s with a genome of under 6,000 base pairs, small enough to test whether a model could generate a full genome end to end rather than stitch together fragments of existing ones.

Sixteen of the 302 synthesized genomes worked, diverging substantially from anything found in nature, and several outperformed PhiX174 itself. One, designated Evo-Phi36, contained a functional truncated protein that had failed when researchers tried to engineer the same truncation into wild-type PhiX174 through conventional methods. When the team combined several AI-designed phages into a cocktail, it overcame E. coli strains that had evolved resistance to natural PhiX174, a result aimed directly at phage therapy’s persistent bottleneck: bacterial resistance to any single phage develops quickly.

Evo 2 is released open source, including its training code, parameters, and training data, a deliberate choice the team says lets researchers study dangerous natural pathogens and build safety checks directly into the design process. The safeguard against misuse is a filter applied at the training stage: Evo 2’s training data excludes human-infecting viruses. That is a data curation choice, not an architectural limit on what the model could learn if retrained. An outside analysis of the earlier preprint noted a sufficiently motivated party could in principle fine-tune an open version of Evo on human viral sequences. HIV’s genome runs to about 10,000 bases and a coronavirus genome to about 30,000, neither dramatically larger than the phage genomes Evo was already generating. What still separates a phage result from a human pathogen is data availability and the physical scale of synthesizing and assembling a genome three orders of magnitude larger than PhiX174’s.

The paper’s authors addressed the gap directly rather than leaving it to outside commentary, writing that genome design with AI raises biosafety and biosecurity considerations requiring active governance work, and recommending that any group doing whole-genome design consult safety professionals from the earliest design stage through publication. Proposals now circulating include a legal requirement for DNA-synthesis companies to screen every order against known threat sequences, and detection methods tuned specifically to catch AI-generated genetic designs, which differ in subtle statistical ways from natural sequences. No AI model in public release has been trained to design a virus that infects people, and that exclusion still held as of publication.