gensyn (@gensynai): Gensyn Was Early: The Future Is Swarms, not Supermodels

Gensyn Was Early: The Future Is Swarms, not Supermodels There is a particular kind of being right too soon: you build something, ship it, and watch the larger world arrive at the same conclusion some time later. That is roughly where Gensyn sits today. The work done on RL-Swarm, BlockAssist, and CodeAssist was always pointing at a specific idea - that models learn better, serve people better, and align more honestly when they operate as a collaborative system rather than a single monolith. That idea is now mainstream. The question worth asking is what it actually means in practice and where the hard problems live. The Problem With One Big Model The dominant paradigm for the last several years has been simple: train bigger, centralise more, and push a single model to cover every use case. That approach produced real gains but it was always working against a structural limit. There is an impossibility result in assistant alignment that makes this concrete. Given two models with equal capabilities, if one is meant to serve many humans, it is impossible for it to serve all of them with equal quality. It must degrade quality for someone in order to improve it for someone else, as long as the humans it serves are not themselves homogeneous. That is a property of the setup and it is baked in by construction. The practical consequence is that a single giant model will always be a compromise - everyone gets something close to what they need, and nobody gets exactly what they need. This matters because so much current AI infrastructure is built around the assumption that one sufficiently large, sufficiently capable model can eventually satisfy everyone. The alignment research suggests that is not just an engineering challenge to be closed by scale, but a structural constraint that no amount of additional training resolves. What RL-Swarm Was Actually Doing RL-Swarm was a bet on a different training geometry. Instead of one model improving in isolation, it put many models into a network where they could solve problems, critique each other's answers, vote on the best solution, and revise their own output before each round ends. Information moved laterally through the network, not just downward from a fixed dataset and a reward signal. The efficiency gains were real. SAPO, the algorithm underlying this approach, showed 94% improvement in cumulative reward compared to models trained in isolation. But the more important goal was capability transfer. The aim was to move knowledge between models that were good at different things, so that a participant in the swarm could end up stronger than it would have become training alone. The word that carries the most weight here is diversity. A swarm of identical models sharing rollouts will learn more efficiently, but it will not learn anything new. The breadth of what the network can teach itself depends directly on how different the participants are from each other. That is why RL-Swarm was designed to be open and permissionless: the more varied the participants, the richer the signal the network can generate. Assistance Games and Intent BlockAssist and CodeAssist addressed a different problem. Even a capable model, or a capable swarm, still has to figure out what the user actually wants. That turns out to be genuinely difficult, and the difficulty grows as models become more capable of acting in the world. What Gensyn explored through CodeAssist was a training setup called a tool mediation game. A model sits between the user and the environment, watches a stream of actions, infers intent from that stream, and routes work to the right tool. The user does not need to specify everything explicitly. The model learns from behavior over time. You edit code on a problem, and the model infers from the pattern of your edits what kind of assistance would actually be useful next. Meta's neural computer work points in a similar direction. The goal is to replace the conventional computer interface with a model that mediates directly between user and hardware. Rather than a GUI or command line, the model takes continuous user input, manages what is happening in the background, and produces output that reflects the user's intent. Both neural computers and assistance games change where alignment happens. In the standard RLHF setup, the human provides feedback after the model has already produced output. In assistance games and neural computers, the human and the model are interacting with the world at the same time. The alignment signal is continuous rather than retrospective. That has real consequences for how the system learns what a user actually needs. Collaborative Post-Training Vanilla post-training has one useful property that collaborative post-training gives up: the model being calibrated is stable. The reasoning comes from one place. The signal is internally consistent. When multiple models are training together and exchanging reasoning traces, small approximation errors start to compound across rounds. Two models might agree on a concept well enough to communicate, while still carrying subtly different internal representations of it. Across enough interactions, those small gaps can grow. The shared signal starts to drift. That is the central calibration problem for collaborative post-training at scale. The research being done on multi-agent workflows is trying to address a concrete version of this. An orchestrator and its sub-agents pass tasks back and forth. When a sub-agent returns something incomplete or slightly off, the orchestrator currently has to intervene and reprompt. The goal is to distill the patterns of successful interaction back into the agents themselves, so they begin behaving correctly from the start instead of requiring correction midway through. The payoff is fewer tokens, less back-and-forth, and better behavior by default. Why It Is Showing Up Now The broader industry started moving in this direction because two things changed at once. Models became much better at tool use, including search, terminal commands, and interaction with external APIs. Smaller models also became capable enough to handle useful production tasks reliably. This changed the economics. When only very large models could do meaningful work, centralising everything made sense. Once a smaller model can handle a coding subtask accurately and cheaply, it becomes practical to build systems where specialised agents handle specialised work. The orchestration overhead is real, but for hard tasks it is worth paying. There is also a more direct pressure. Bigger centralised models hit diminishing returns. The frontier becomes harder to move, the costs rise, and the reliability gains slow down. At that point, orthogonal approaches become attractive. Meta's neural computer work is a useful example. It is not competing on the same monolithic frontier as the other large labs. It is exploring a structurally different architecture, which is the kind of move that tends to make more sense when catching up on the main track is not the best use of resources. The Mirror World The long-term vision is not a chatbot that knows everything. It is closer to a world model that is always running in the background, simulating and forecasting any facet of the world anyone cares about. Medical research, logistics, education, daily planning. The swarm is always buzzing, generating signals from many models operating on many different problems in parallel. What each person sees is not the swarm directly. It is a personalised layer on top of it, shaped by the assist-style work, that translates raw signal into something relevant and suited to that person's context. The swarm generates signals and the interface decides what each person sees. That separation matters because it is the only way to break the impossibility result without forcing every user into the same mold. The failure mode is also easy to name. If that infrastructure gets built in a centralised way, controlled by a small number of actors, the result is not necessarily a rogue AI. It is something quieter: an intelligence layer that looks general but has quietly converged on a narrow set of priorities. The same pattern that shows up in so many physical spaces, applied to intelligence and information generation. That is the outcome to avoid. The reason the early work on RL-Swarm, CodeAssist, and BlockAssist matters is that it did not treat collaboration, mediation, and personalisation as separate trends. It treated them as one system. That system is now becoming visible everywhere else. The job now is to build it in a way that stays open, diverse, and actually useful to the people it is supposed to serve. (This article was created from a conversation with Gabriel Andrade and Semih Kara. You can listen to the recording here - https://youtu.be/3VBkNwtN28k)