Daniel Beutel on Flower SuperGrid and Collaborative AI

Flower Labs’ Daniel Beutel presents his company’s vision of collaborative AI: organizations can run computation where data resides and coordinate work across multiple sites. The talk is a product presentation, so its adoption and performance claims are the speaker’s.

Beutel opens with a collaboration with Starcloud on training a vision transformer aboard a satellite. He presents this as an example of distributed compute; the demonstration does not establish that space-based training is generally economical.

He traces Flower to work at Cambridge on decentralized AI. Rather than collecting all raw data into one place, participants can retain local datasets and coordinate computations across them.

Beutel describes Flower as an open-source framework and cites community and organizational use. His characterization of it as an industry standard is a company claim; the talk does not compare adoption across the whole field.

His motivation is to make more distributed data useful for training and inference. He offers estimates of public and private token volumes, but those figures depend on definitions and data quality and are not independently established in the talk.

Organizations often obtain additional data through collection or licensing. Beutel proposes collaboration among data holders as another path, while acknowledging that the data remains divided across many separate systems.

He compares a collaborative network to the development of electricity grids and the internet. The analogy explains his vision of shared infrastructure; it does not by itself establish the economics or governance of any proposed network.

Flower’s proposed stack has three parts: a distributed platform, agents designed to collaborate and a training pipeline that can learn from data at participating sites.

The first component, SuperGrid, is intended to simplify setting up a federation. Beutel demonstrates creating a federation, inviting participants and attaching individual nodes at data sites.

A developer can initiate a distributed run and inspect its status and logs. Beutel contrasts the demonstration with a much longer manual setup, though the time saved in a real deployment depends on integration and governance requirements.

He lists isolation, auditable communication, model transfer and governance as platform features, along with confidential-computing integrations. The actual security properties depend on configuration, threat model and deployment.

Beutel cites a primary-care project as an example of working across patient data sites without centralizing raw records. This is a reported customer example, not independent evidence of a particular clinical outcome.

For agents, he distinguishes access to public web information from access to an organization’s own data or another participant’s data. A network could make additional context available when participants agree to collaborate.

He describes retrieval, reasoning and action as core agent capabilities, then adds collaboration between agents. Production collaboration also needs discovery, permissions, security and auditability, not just a message protocol.

In his architecture, a coordinator sends tasks to agents at individual sites. Each site can decide whether to participate and whether a proposed response may leave its boundary; the coordinator then combines permitted responses.

Beutel summarizes the design as keeping raw data in place while agents exchange governed messages. Local data residency alone does not prevent sensitive information from appearing in responses; policies and enforcement must address what is actually shared.

Project Kaya is his team’s early-access collaborative agent. In the demonstration, a coordinating agent contacts several geographically distributed nodes and prepares an answer from the responses it receives.

The third component is a decentralized training pipeline. Beutel says it can use information from participating sites to improve a model without first collecting all raw training data centrally.

He discusses Flower’s UK-focused open-weight model Lizzy 7B; the automatic captions misrecognize the name. [Flower’s model page](https://flower.ai/models) identifies it as a UK-oriented model. The performance comparisons in the talk are the company’s claims.

Beutel reports up to a thousandfold reduction in communication costs in a particular decentralized-training result. “Up to” depends on the baseline and experimental conditions, so it should not be assumed for every network or model.

In the illustrated training loop, a coordinator sends a model to participating nodes. Each node trains locally, sends an update back, and the coordinator aggregates updates; the cycle repeats.

Keeping raw records local can reduce direct data movement, but it does not automatically guarantee privacy. [Federated-learning research](https://proceedings.mlr.press/v234/zhao24b.html) shows that shared model information can leak training data under some conditions; defenses require separate evaluation.

He describes a collaboration involving the Department of Energy and Sandia National Laboratories on decentralized training across three sites. The talk presents this as work in progress rather than a published, independently verified outcome.

Beutel closes by reiterating the three layers: SuperGrid for the network, Frontier for distributed training and agents such as Kaya for collaboration. His broader forecast is that these networks will become more important to AI development.