AI Safety for Builders
Micro Fellowship
4 weeks. Small teams (2–5). Real AI safety tools, shipped.
- 4 weeks
- Online
- Part-time
- Small teams (2–5)
- Free
Why
Plenty of research. Few tools in real use.
There is a lot of research and empirical work happening in AI safety, but much less focus on building products and tools.
If one takes both AGI capabilities and risks seriously, a future where AI goes well will require scalable, deployed AI safety solutions.
| Research | Building | |
|---|---|---|
| Capability | AbundantAcademic labs, papers | AbundantStartups, model labs |
| Safety | MATS · SPAR · MARSStrong, research-first | The gapNo tools-first programmeThis fellowship |
Others see the gap too
- ApolloThe case for AGI safety products
- HalcyonCall for safety-tooling founders
- ForethoughtDesign sketches for a more sensible world
- ApartAI safety needs to scale
- SurplusAI safety software incubator
- GoodfireAI safety still needs great engineers
Open globally, with a deliberate effort to bring in builders from underrepresented regions such as India.
How it works
Four weeks, one shipped artifact.
- Week 1Scope
- Week 2Build
- Week 3Test
- Week 4Demo Day, Awards
Mentors
What will mentors do?
You give
- A project idea you want built (or pick one from the list)
- Weekly check-ins and feedback
You get
- A live artifact that helps ensure AI goes well
- Strong prospective collaborators
We support
- Awards and compute
- Ecosystem connections
Fellows
What will fellows do?
- Join a small team on a project a mentor has scoped
- Ship something concrete and open-source
- Put in 4 weeks (8–12 hours per week), then present at Demo Day
What you can build
- Toolse.g. simplified red-teaming tools
- Sandboxese.g. agent boundary tests on Digital Public Infrastructure
- Eval harnessese.g. interactive evals for interaction harms
- Instrumentse.g. an Indic bias leakage probe
- Guardrailse.g. a post-deployment legal answer audit layer
- Protocolse.g. machine-checkable logic for rule-bound domains
- Pipelinese.g. an incident-to-benchmark-to-law gap mapper
- Pluginse.g. rhetoric highlighting wherever you read and write
- Platformse.g. an India AI incident and near-miss reporting dashboard
Example AI safety tools
Call for mentors
Call for mentors
Have a safety tool you wish existed? Pitch it. We'll match it with a small team of builders who ship it in four weeks.
Mentors may come from either side:
- Industry, product, or buildingExperienced builders, domain experts, practitioners, or prospective users of a safety tool. You know where a tool is missing, who needs it, and what “good” looks like.
- ResearchAI safety researchers who want to do more operational work. Bring research that needs engineering support: an evaluation harness, data or ML pipeline, experiment dashboard, monitoring system, interpretability tool, benchmark, or research prototype. We’ll help refine the scope and match it with builders who can implement it.
You do not need to be an engineer: what matters is bringing a valuable problem, relevant context, and timely feedback.
What we'll ask
- Name and affiliation
- Area of expertise
- Project pitch: what kind of build, for whom, and what ships
- Available hours
- Time zone