What alphabell is, and what it isn't.

Origin and the bet

alphabell was founded in 2017 by a small group of researchers and engineers, mostly from Silicon Valley, who shared two unfashionable convictions. The first: that frontier AI research can be done in a meaningful way outside of large centralized labs, if the coordination problem can be solved well enough. The second: that the prevailing posture of treating AI capability advances as proprietary IP to be ring-fenced and announced as finished is a bad fit for the underlying technology, which is none of those things — it is unfinished, fluid, and shaped at scale by everyone who works on it.

The lab's name is a play on alpha and beta release cycles. A signal that the systems we build are never finished, and that the public posture of "ringing the bell" on a finished AI is one we regard as premature.

The node-cell model

alphabell organises its research into node-cells: small autonomous research pods of three to seven contributors, each centered on a shared problem. A cell sets its own cadence. It publishes internally to a common research index. It competes and collaborates with adjacent cells working on related threads. There is no central management hierarchy translating instructions down from a CEO. Cells are coordinated by a lightweight protocol layer — signed proposals, weighted consensus, asynchronous review — that has been refined over the lab's eight-year history.

For the deeper mechanics of the node-cell model — how cells are formed, how they fission, how they get and lose compute access — see /node-cells.

Four research axes

The lab's portfolio is organised along four primary axes. Three of them are research-producing in the traditional sense; the fourth is cross-cutting infrastructure.

• Agentic engineering. Agents as durable computational entities, not wrappers around language models. Long-horizon planning under partial observability, multi-agent negotiation protocols, sandboxed self-modification, and the development of agent substrates.
• World models. Predictive generative systems that learn the dynamics of physical, social, and symbolic environments. Compositional latent dynamics, counterfactual rollout for planning, embodied simulation for robotics pretraining, and the unification of perception and prediction.
• Recursive self-improvement. The lab's most closely held research line. Models that propose, evaluate, and incorporate modifications to their own training procedures, architectures, and evaluation criteria. Strict internal protocols on capability evaluations, isolated compute enclaves, and pre-registered stopping conditions.
• Interpretability & alignment infrastructure. A cross-cutting axis that supplies tooling and theory to the other three. Mechanistic circuit analysis at frontier scale, scalable oversight for agentic systems, and formal verification approaches for learned policies.

Two unusual structural choices

alphabell rejects two defaults simultaneously: secrecy-as-default and centralization-as-default. We treat decentralization as a safety property — no single jurisdiction, executive, or funder can unilaterally redirect the lab's trajectory. We treat openness as a capability multiplier — distributed peer review, with its associated coordination tax, compounds faster than internal review at scale.

Both rejections come with operational costs and we pay them on purpose. We do not move as quickly on flashy demonstrations as a centralized lab moves. We tax everyone with coordination overhead. We sometimes pause work — including work that would otherwise have shipped — when our paired interpretability cell is not satisfied. We have made those choices on purpose.

Funding and compute

alphabell does not raise conventional venture capital. The lab is funded through a mixture of philanthropic grants, sovereign research partnerships, and revenue from selective licensing of derivative tooling. Compute is pooled through a federated arrangement: contributing nodes commit GPU and TPU capacity into a shared scheduler. Access is allocated by a combination of research seniority, project priority signals, and a quadratic-voting mechanism among active contributors. The specifics live at /compute and /funding.

Publication policy and the charter

Publication is staged. Foundational results are released openly. Capability advances are released with delay and accompanying safety analyses. A small fraction of work — primarily within the recursive self-improvement axis — is held indefinitely behind internal review. Every contributor signs a research conduct charter on entry, and any cell working on dual-use capabilities is paired with an interpretability cell that has read-access to its checkpoints and training logs. The pairing is structural, not optional.

What you'll find here

If you want to understand what we research, start at /research. If you want to understand how the lab works, start at /node-cells and /governance. If you want to read the actual outputs, see /publications. If you want to contribute, /contribute describes the onboarding path, including how to be paired with a cell and the contributor cohort cadence.