Cosmos Institute opens applications for grants of up to $10,000 for AI prototypes focused on human autonomy

To give you some inspiration, here are what some of our current grantees are working on.

Priori, built by @smolotnikov, surfaces the silent choices behind an AI's answer, like its tone, what it emphasises, and the values it assumes, and turns them into controls you can adjust.

Steven has since taken a Cosmos Fellowship to study how that shapes autonomy, along with @thecatfangs.

Anyone can apply, no credentials needed, just an idea you can turn into a working prototype. Applications close on Sunday July 26.

https://blog.cosmos-institute.org/p/cosmos-grants-are-back

We run two tracks.

The first funds work that helps people keep exercising judgment rather than handing it to AI.

The second, in partnership with @TheFIREorg, and with compute available from @PrimeIntellect, backs projects that keep ideas open to challenge, instead of settling them with a single confident answer.

Campus, built by @kaseyklimes, is a project-based learning tool that adapts hard projects to your level and keeps asking for your input, building the habit of self-directed growth.

Kasey has since started a company, Primitive, focused on coordinated intelligence. https://newsletter.getprimitive.ai/p/what-i-learned-building-whole-earth

MoSaIC, built by Caitlin Morris, pairs AI conversation analysis with expert human evaluation to capture the social side of learning that solitary AI tutoring tends to flatten.

Caitlin's work was published at ISLS CSCL and IEEE FIE, and became part of her PhD at the MIT Media Lab.

TLM-1, built by Brandon Duderstadt at @calco_io, is a language model that accounts for time. It predicts both what a document says and the year it was written, surfacing how words shift meaning across decades.

Brandon's paper was accepted to the ACL 2026 Main Conference.

Thought Refractor, built by @HcwXd, turns any text into several visual forms, like breakdowns, comparison frames, flowcharts, and concept maps, so you can see how an argument is built from more than one angle.

It's the basis for @WonderingApp, Cheng-Wei's new startup.

Prompt Injection as Role Confusion, produced by Jasmine Cui and team, showed that prompt injections exploit a flaw in how models handle roles, leaning on memorised rules rather than reasoning through a situation the way a person would.

Jasmine won OpenAI's red-teaming contest, and her paper was accepted to ICML.