α · Research

Four axes. Twenty-one active cells. One internal index.

Every cell at alphabell is organised under one of four research axes. The interpretability axis cross-cuts the other three: any cell working on a dual-use capability is paired with an interpretability cell with rolling read-access to its checkpoints, training logs, and proposal commits.

Active cells 21

In review / paused 2

2025 published reports 24 (and counting)

Delayed-release queue 3

01 · α

Agentic engineering

Agents as durable computational entities with persistent state, learned tool affordances, and verifiable execution traces. Long-horizon planning under partial observability, multi-agent negotiation, sandboxed self-modification, and agent substrates that expose memory, identity, and resource budgets as first-class primitives.

Explore the axis

02 · β

World models

Predictive generative systems that learn the dynamics of physical, social, and symbolic environments from heterogeneous data. Compositional latent dynamics, counterfactual rollout, embodied simulation, unification of perception and prediction.

Explore the axis

03 · γ

Recursive self-improvement

Models that propose, evaluate, and incorporate modifications to their own training procedures, architectures, and evaluation criteria. Capability evaluations, isolated compute enclaves, pre-registered stopping conditions. The lab's most closely held research line — and the one with the strictest paired-cell requirements.

Explore the axis

04 · δ

Interpretability & alignment infrastructure

Cross-cutting axis supplying tooling and theory to the other three. Mechanistic circuit analysis at frontier scale, scalable oversight for agentic systems, formal verification of learned policies, and the trust protocols that make cell pairing functional.

Explore the axis

How an axis works, briefly

An axis at alphabell is not an org-chart unit. It is a shared problem statement and an internal review pool. Cells choose their axis. Axes have a steward — a long-tenured contributor who maintains the axis's published research agenda and convenes the axis's quarterly cross-cell review. Stewards rotate every two years. The current stewards are listed at /people.

Cells can and do span axes. A cell working on agent negotiation might consume tooling from the interpretability axis, train its agents in environments developed by a world-models cell, and contribute findings back into both. About a third of currently active cells are formally cross-axis.