Today we're announcing Slingshots // THREE: 13 research projects across local inference, architectural innovation, continual learning, agentic systems, and the biological sciences. This is the third group of researchers we've brought into the Laude community through Slingshots, and with each batch the network of people building in the open gets deeper.
Some of what makes this batch interesting is what it says about how open research compounds. OpenJarvis picks up directly where Intelligence-Per-Watt left off in Batch Two, moving from measuring capability per unit of energy to building the full local-first stack. TBench Science extends the Terminal-Bench and Harbor family that started in Batches 1 and 2.
The rest of the batch is varied by design. Token costs rising makes the case for edge inference more urgent, and two projects are working on that problem from different angles. Two researchers are independently working on architectural contributions — one on energy-based training, the other on persistent memory in recurrent state. There are new benchmarks, new execution engines for agentic workflows, a world model framework for embodied agents, and a microfluidic platform applying automation to protein interaction research.
We are proud to back the work of these researchers and scientists working and building in the open. Support for each project is bespoke. For some teams that means compute; for others it means product strategy, introductions, or help standing up the infrastructure around the research itself. The only constant is that the work ships open source.
Now, let's meet Slingshots // THREE.
An open-source framework for local-first personal AI, which runs agents on-device by default, calls the cloud only when necessary, and treats energy, cost, and latency as first-class metrics alongside accuracy. Stanford University
A low-storage RAG system that enables fast, accurate, and fully private retrieval directly on personal devices. UC BerkeleyPrinceton University
A new generative modeling paradigm for end-to-end training or hybrid training of diffusion models, energy-based transformers, LLMs, and other generative models. UIUCHarvard University
A harness and user interface for AI-assisted writing. Carnegie Mellon UniversityUC Berkeley
A sequence modeling architecture based on Omega learning rule that builds persistent associative memory into the recurrent state, enabling long-horizon context retention without quadratic attention costs. Cornell University
Frontier-CS is an unsolved, open-ended, verifiable, and diverse benchmark for evaluating AI on challenging computer science problems. UC BerkeleyPrinceton UniversityUCSD
A microfluidic chip that generates protein interaction datasets at the scale required to train tomorrow's protein language models. Stanford University
A framework that unifies LLM continual learning methods, showing that the data and task conditions determine whether continual learning really requires learning. UC Berkeley
A runtime adaptive execution engine for agentic workflows. Harvard University
World models for planning, scientific simulation, and verification across physical and industrial systems. MilaMcGill University
A framework for teaching LLMs to codify human expertise into formal logic. Stanford University
A neural operating system that generates screen frames end-to-end from user actions. University of WaterlooNational Research Council Canada
A benchmark for evaluating AI agents on real computational workflows across the natural sciences, with tasks authored and verified by scientific domain experts. Stanford University
Researchers: Tell us what you're working on and what you need to get it into the world. We accept projects year-round, deploy resources immediately, and announce new batches every few months. Apply here.