mission

Frontier models, trained and owned by the many

The most powerful AI models today are developed secretively and are not released to the general public. This is driven by safety and economic concerns; however, because of the cultural, economic, and geopolitical impact the models have, there is potential for massive power asymmetry to arise between those who shape their development and those who don’t. This dynamic becomes especially pronounced as the models begin to accelerate the development of their own successors.

There are two ways to address this problem: 1) drastic reformation of the institutions that will exert control over the development of this technology 2) formation of a neutral layer for model creation, governance and release. Pluralis focuses on the latter. Open weights are not sufficient – a sustainable ecosystem must be capable of producing the models rather than consuming them or applying modifications on top of expensive pretrains. Historically, neutral, open layers with these properties have resulted in better capabilities, broader adoption and better global equilibria.

However, AI is not like previous tech. It is hugely expensive to develop, some economic aspects look far more like physical infrastructure than they do software, and the models have the ability to do damage in a way no previous technology could. Yet unlike physical infrastructure, the model artifacts and training process are purely digital. This is a strange combination – something of unprecedented real world utility, with huge cost to create and a novel safety profile, that retains many properties of software. As a result, the open ecosystem around these artifacts will also look fundamentally new; half traditional open source, and half real world economics with capital formation, programmatic value flow, and novel ownership and governance mechanisms.

This third path for model development; plural, multi-party construction of strong models, where builders are owners has many technical challenges; one being that all training today is done in a single-party manner which renders large amounts of the current ML stack unusable. Training libraries, the communication layer, kernels, and even model architectures, are all designed under the assumption that one party carries out the design, assembles highly networked compute in a datacenter, and trains the model independently. Our goal is to create an alternative stack that can train collaboratively, where many participants with different hardware and software can be combined into one large pool of compute, and where even a large minority of actors cannot disrupt the overall system.

Until recently this was widely considered to be infeasible. Among other problems, training a large model means moving enormous volumes of data between GPUs at every step. Our view was that this is surmountable: architectures have co-evolved in the presence of the datacenter substrate, and so model ↔ substrate co-design was likely to yield methods that could train in this way. Our Subspace Networks work was the first of these. It compresses what passes between the parts of a model by around a hundredfold and allows the model itself to be sharded over devices during training. The research demanded here is broad, covering novel optimizers, data-parallel compression, heterogeneous training and other sub-areas. We have published this work in NeurIPS, ICML and ICLR across nine papers so far. Agora, our training system, has pretrained models of up to 13 billion parameters in open runs with hundreds of participants.

This research has a critical side effect; because the model itself must be sharded over participants, the full weights never exist in any one place. Participants can help create the model without any individual being able to obtain it. This retains the benefits of a closed lab (the ability to monetize and have exclusivity over the model), while removing the major risks of open weights (abliteration, misuse, de-alignment). It also introduces a novel property; permissionless collaboration on development of the models themselves. This enables broad dispersion of the economic benefits and governance, while maintaining a safety profile that is arguably stronger than closed-model labs1 since there is no single copy of the weights to steal, leak, or unilaterally modify.

Much work remains; however our view is these approaches will scale without major changes. A third path for model development appears to be opening.

1We view this not just as having better economic and safety profiles, but as the correct technological approach – models are not like software; they are grown not programmed – they are stochastic not deterministic, the best learning algorithms are designed to be robust not fragile. A noisy, dynamic hardware substrate, to us, feels more natural than a monolithic, static block of compute.