A lab built around a simple bet.

The bet

The dominant paradigm in AI today is straightforward: more data, more parameters, more GPUs. The recipe has carried the field a long way, and it isn't going to stop working tomorrow. But we think it has been quietly absorbing two different problems into one. Sometimes a model needs more compute. And sometimes the researchers need a better idea — and "we don't yet have a good idea" keeps getting mistaken for "we don't yet have enough compute".

Our wager is that the next round of real breakthroughs will come from the first kind of problem: from sharper hypotheses, careful theory, elegant constructions, and the kind of focused thinking that doesn't get cheaper just because you spent another billion dollars. The bottleneck is the idea, not the cluster.

If that's right, the strategically correct response is not to try and out-scale the frontier labs. It's to find as many great thinkers as possible, give them runway to think, and put the results into the open. So that's what we do.

Why this rather than another frontier lab

Two-thirds of the global AI research budget today goes to ingesting more of the internet and building larger data centers. Plenty of work has the budget, the cluster, and the headcount to push that frontier. Almost no-one is funded to pursue the careful, ideas-first work that historically produced the leaps — backprop, attention, RL — none of which arrived because someone added another wing of GPUs.

Alphabell exists to make that ideas-first track a little more legible and a little better funded. We are small on purpose. We don't try to win at scale. We coordinate the people who, in another era, would have been the corresponding members of a small institute or a Bell-Labs-style group — except now they live in 41 countries and they get paid to think out loud.

What we work on

The agenda is the set of problems where the bottleneck is the idea rather than the cluster.

• New architectures and training methods — small, principled, often theory-led.
• Mechanistic interpretability and evaluation tooling — sparse-feature atlases, causal scrubbing, refusal-direction studies. Work that gets unblocked by a clean hypothesis, not a larger GPU pool.
• Agent systems and long-horizon evaluation — schema-aware tool use, trajectory datasets, harnesses that reveal real failure modes.
• Dataset experiments — new training data, new ways to study existing data, and underserved-language eval sets built with native speakers.
• Replications, audits, and reference implementations — slow, careful work that the field's incentives systematically under-fund.
• Weirder bets at the edge of the field — the long-tail of ideas that almost no-one is funding, often because the hypothesis is too specific to fit any single lab's roadmap.

Projects are proposed and led by members rather than handed down. Most live in public from day one, with code, weights, evals, and write-ups released as they mature. Reproducibility is treated as a first-class output, not as a footnote.

How members participate

There is no single "alphabell role". A high-school student writing a clean implementation, a PhD on sabbatical, and a senior engineer chipping in evenings all participate on equal footing — judged by the work. We've built four overlapping ways to engage:

• Mini-grants — small, low-paperwork awards ($500 to $25,000) for promising ideas, decided fast. We borrowed the spirit from Emergent Ventures: a single short application, light overhead, decisions in weeks rather than months.
• Fellowships — three to twelve months of funded time for researchers who want to commit deeper to an alphabell project, or pursue their own agenda with mentorship and compute.
• Competitions and hackathons — recurring sprints with cash prizes, leaderboards, and public scoring. We borrowed the spirit from Foldit: standings are visible, top performers are credited by name, and a strong run is itself a credential.
• Compute and infrastructure — GPU clusters, datasets, evaluation harnesses, and engineering support, available to contributors with active projects. Enough compute to test an idea, not enough to brute-force a problem.

A portable research reputation

Co-authorship, citations, prize records, and leaderboard standings accumulate over time. Members carry that reputation with them — into academia, into industry, into a more senior role inside alphabell on a bigger project, or back out into the world. We don't issue degrees. The work, in public, is the credential.

How we make decisions

Mini-grants are reviewed by two members from the L7+ pool, drawn at random. Reviewers are paid per accepted review, not per decision. No CVs in the review packet — only the proposal. Median time to decision: 14 days.

Competitions are auto-graded against held-out evals. The leaderboard is the decision.

Fellowships go through a slower three-step process: a written round, a short video conversation, and a cohort-fit check. Two windows a year, eight fellowships per cohort.

Where the money comes from

Alphabell is funded by a small set of philanthropic backers and a rotating roster of independent donors. We publish a yearly transparency report — total budget, breakdown by program, average grant size, review-cycle times. The 2025 report lives on GitHub.

As a matter of policy we don't accept money from frontier AI labs and don't run programs sponsored by individual companies. That's a hard line, not a guideline.

What we don't do

• We don't try to out-scale a frontier lab. That isn't the game.
• We don't fund capability-only work whose only differentiator is more compute or more data.
• We don't issue degrees or certifications. The work, in public, is the credential.
• We don't fund closed-source research with no public output.
• We don't fund work whose stated purpose is to win a particular publication race.

How to get involved

Three doors:

1. Apply for a mini-grant or a fellowship.
2. Enter a competition or a hackathon.
3. Join an open-call project as a collaborator, annotator, or reviewer.

Or just write: hello@alphabell.com.