AI that makes AI better
Alphabell supports researchers around the world, and we have high conviction in one idea: AI tools can be used to make AI tools better. This hub tracks, explains and funds the work where that loop is real.
The bet
Most AI progress is reported one artifact at a time: a model, a paper, a benchmark score. We care about something narrower and, we think, more consequential: work in which an AI system is an integral part of the loop that produces the next improvement in AI. When the tool doing the work is also the thing that gets better, each turn of the loop can make the next one faster, cheaper or more ambitious.
That loop does not have to be fully automated. A researcher in the loop is fine, and usually wise. The loop does have to be real, and its rationale has to be stated. "A great AI tool that did well on benchmarks" is not what this site is about.
The bar: what counts as a loop
An entry belongs here when you can fill in this sentence without hand-waving:
[An AI system] does [some work], which improves [a model, tool, dataset or piece of hardware], which makes [the same system or its successor] better at [that work].
Examples that pass:
- An AI research assistant that searches for the newest work on research assistants and folds what it finds back into its own retrieval and prompting.
- A coding agent used to write the next version of its own harness, with every version scored on the same task suite.
- A model trained to design better GPU kernels or chips, where the speedup is used to train or serve that same model, which then designs the next round.
- Models that filter, label or synthesize the training data for their successors.
- Models that act as the judge, critic or opponent in their own training.
- Models used to write the next version of themselves. This is the classic case, and only one of many.
Examples that do not pass on their own:
- A strong model, product or paper with no path back into AI development.
- A benchmark of general capability, unless it measures the ability to do AI research or engineering.
- AI applied to another science, however impressive, unless the result feeds back into AI.
- Commentary about recursive self-improvement without a project, dataset or result behind it.
Highly specific projects are welcome, and the loop can be indirect. A model that designs better hardware, which later lets the same model run faster and design even better hardware, qualifies, as long as the road back is spelled out.
How far the loop has turned
Every entry carries one of three stages, so you can see how far the evidence goes:
- Closed loop. The improved output has been fed back and the gain measured.
- Partial loop. One turn of the loop is shown and the path back is clear.
- Prospective. The road to recursion is argued and plausible, not yet run.
Loop types
We group entries by the part of the AI stack the loop improves. The classic "model writes its next version" case lives under self-modifying agents; most of the interesting work is elsewhere.
| Loop type | What it covers | Entries |
|---|---|---|
| Self-modifying agents | Agents that edit their own code, prompts or scaffold, and keep the versions that score better. | 7 |
| Tools and harness | AI that writes or improves the tools, harnesses and infrastructure used to build and run AI. | 4 |
| Training data | Models that generate, filter, label or curate the data their successors learn from. | 8 |
| Self-reward and self-play | Models that supply their own training signal: judges, critics, verifiers and self-play opponents. | 7 |
| Kernels and compilers | AI that writes GPU kernels, compiler passes and low-level code that train or serve AI faster. | 6 |
| Hardware and chips | AI used to design the chips, boards and systems that AI then runs on. | 3 |
| Inference efficiency | AI that makes AI cheaper to run: distillation, quantization search, draft models and routing. | 5 |
| Architecture and optimizer search | AI searching over architectures, optimizers, losses and training recipes for the next model. | 5 |
| Research automation | AI that proposes hypotheses, runs experiments and writes up machine learning research. | 5 |
| AI R&D evals | Benchmarks that measure how well AI does AI research and engineering: the scoreboard for the loop. | 6 |
| Interpretability and oversight | AI used to explain, audit and supervise other AI, with the findings fed back into training. | 5 |
How the hub works
Every day an editor pipeline reads new papers, repositories, benchmark results, lab blogs and community posts from public sources. A language model drafts a loop statement for each candidate and scores how directly the work feeds back into AI. Items without a clear loop are dropped. The rest are ranked, and the strongest are published on the radar with the source credited and linked. Entries that hold up are added to the loop index.
The editor is a small instance of the idea: an AI tool that reads about AI tools so that it, and the people using it, find better ones. Its write-ups are machine drafted and can be wrong. Titles, results and links belong to their authors; we link rather than republish. If you spot a mistake or a missing project, tell us.
How we support the work
We fund and help people who are building loops, measuring them, or replicating claims about them.
- Grants: low paperwork funding, sized to the work, decided fast.
- Fellowships: funded time for researchers who want to go deeper on one loop.
- Compute: GPU time and credits for experiments that have to run more than once.
- Competitions and hackathons: public, auto-scored challenges where the task is to make AI better with AI.
Where the money comes from
Alphabell is funded by philanthropic backers and independent donors. We do not run programs sponsored by individual companies.
What we do not do
- We do not try to out-scale a frontier lab.
- We do not fund capability work whose only lever is more compute or more data.
- We do not fund closed research with no public output.
- We do not issue degrees or certifications. The work, in public, is the credential.
Working on a loop?
Tell us what the AI improves and how it comes back. We read everything sent through the contact page.