Peter Vajda's eleven years at Meta culminated in directing the teams behind Movie Gen and Emu—the sort of generative AI work that makes headlines when it produces photorealistic video from text prompts. Seiji Yamamoto, meanwhile, spent his Meta tenure deep in the performance guts of Core Llama, the large language model family that now powers half the AI industry's experiments.
Both left. And not for another AI lab or a well-funded competitor. Instead, they're building something that sounds almost quaint by Silicon Valley standards: a platform to help people solve actual scientific problems.
The company was part of Y Combinator's Spring 2026 batch, though the exact launch date remains difficult to independently verify. Vajda and Yamamoto introduced ScienceSwarm, an open-source platform designed to let anyone collaborate with AI on unsolved questions in mathematics, biology, physics, and beyond. The San Francisco startup emerged from Y Combinator with a pitch that frames scientific discovery as the next frontier for AI-assisted democratization. Whether that proves visionary or overreach will depend on real-world adoption and demonstrated results.
The Premise: Pair Programming, But for Proofs
ScienceSwarm operates on what the founders call a "deceptively simple" premise—though perhaps nothing involving AI agents and research memory is truly simple. Users browse open challenges across scientific domains, then work alongside AI agents to propose approaches, test hypotheses, or critique existing work. The system runs on-device by default, a privacy design choice that means your data never leaves your computer unless you explicitly route queries through cloud providers like OpenAI, Anthropic, or Google.
The architecture bundles three components into a unified workspace: OpenClaw (a manager agent), OpenHands (an execution agent), and something called gbrain, which provides durable research memory. Papers, notes, code, datasets—all searchable, all pulling from literature repositories like PubMed, arXiv, OpenAlex, and Crossref. According to the project's GitHub repository, users can upload PDFs for what the documentation calls "venue-style feedback," or trigger a Deep Reasoning Paper Analysis for exhaustive logic checks.
It's collaborative iteration, not autocomplete. A human suggests an approach. An AI reviews it. The human challenges the critique. The AI updates its analysis. Rinse, repeat.
According to the project's changelog, version 0.1.0.0 was added to GitHub on March 31, 2026, followed by version 0.1.1.0 on April 10—updates that added Research Radar, personalized briefings for AI researchers deliverable via Telegram. The platform remains in alpha. The team warns explicitly against production or regulatory use, which is probably wise given how often "alpha" in startup-speak means "we're still figuring out if this fundamentally works."
Lowering the Barrier, or Just Moving It?
ScienceSwarm's Y Combinator profile summarizes the vision with characteristic bluntness: "Doing for Science What AI Coding Did for Software." The analogy is deliberate. AI coding assistants like GitHub Copilot lowered the barrier to building software—not by making developers obsolete, but by handling boilerplate, surfacing documentation, and letting people focus on higher-order problems. Vajda and Yamamoto argue their platform can do the same for fundamental research questions.
Problems on the platform range from theoretical mathematics to applied biology. Users contribute threaded responses—approaches, proofs, counterexamples, literature reviews—that other researchers, human or AI, can challenge, refine, or extend. The interface supports what the documentation describes as "stigmergic coordination primitives," a term borrowed from how ants leave pheromone trails to guide collective behavior without real-time coordination. In practice, this means contributions build on each other asynchronously, much like how GitHub issues accumulate context over time.
Whether the "Copilot for science" framing holds depends on a question the founders can't yet answer: Will communities actually coalesce around these unsolved problems? And will the AI feedback loop prove genuinely useful for hypothesis refinement, or will it mostly surface the same literature reviews a decent grad student could compile in an afternoon?
Developer Hooks and Local Stacks
For agent builders—and ScienceSwarm is clearly courting this crowd—the platform offers both a Python SDK and a Model Context Protocol (MCP) server. The MCP configuration works with Claude Desktop and OpenClaw, with rate limits set at 60 requests per minute and 1,000 per hour. Capabilities exposed through the API include browsing problems, reading context, contributing solutions, accessing collaborative whiteboards, and searching the knowledge graph.
The local stack defaults to Ollama with "gemma4:e4b" as the on-device model, though users can swap in alternatives. Windows support runs through WSL2, which is either a pragmatic choice or a sign that native Windows builds remain a backburner priority. Requirements include Node 22+ and a handful of optional provider CLIs, depending on model preferences. The README emphasizes privacy gates before any third-party API calls escape into the cloud—a design philosophy that feels increasingly rare in an era when most AI tools assume you'll just hand over your data.
Pedigrees and Timing
Vajda's Meta tenure spanned eleven years, culminating in his role as Director of Media Generation. He appears as co-author on the Movie Gen paper published in October 2024 and earlier work on Emu's text-to-video generation—projects that represent some of Meta's highest-profile AI research. Yamamoto's background blends academic physics (publications in PNAS and Physical Review Letters) with applied AI research managing Llama performance teams at what Meta now calls its Superintelligence Labs.
The two founded Gikl, Inc., sometime in 2026 and entered Y Combinator's Spring cohort, which multiple industry observers have characterized as tilting toward AI agents and hard-tech infrastructure. The Next Web reported in early May 2026 on YC's growing emphasis on these sectors, though whether that represents a strategic shift or simply reflects application trends remains unclear. The company currently lists a team size of two on its Y Combinator profile, which either suggests admirable focus or that scaling remains a future problem.
A Crowded Field, Getting Crowdeder
ScienceSwarm arrives as multiple efforts converge around AI-assisted research—though "converge" may overstate the coordination involved. Elicit, which focuses on literature review, released a "Research Agent" update on what it cited as May 1, 2026. Consensus and Scite.ai offer evidence-grounded search and smart citations, respectively, with institutional trials reportedly active as of April 2026. Inquisite promotes "Agentic Search for Science R&D" across materials and life sciences. ScienceClaw and Pubroot both launched AI-powered peer review pipelines earlier this year, if company timelines are to be believed.
An arXiv preprint from April 2026 documented AAAI-26's pilot program on AI-assisted peer review, providing early evidence that AI can contribute meaningfully at conference scale—or at least that conference organizers are willing to experiment. A September 2025 paper titled "Democratizing AI scientists using ToolUniverse" proposed standardizing over 600 models, datasets, and APIs for autonomous research, a vision ambitious enough to make ScienceSwarm look almost modest by comparison.
The trajectory suggests the field is moving from search-and-synthesis tools toward full-stack collaboration environments. ScienceSwarm distinguishes itself, at least in pitch, by emphasizing unsolved problems, local-first privacy, and agent extensibility—positioning the platform as infrastructure for a research workflow rather than a productivity assistant for established tasks. Whether that distinction matters to users remains an open question.
Business Model? What Business Model?
The platform is free to install from GitHub under an MIT license. The company has not disclosed pricing for any potential cloud features, credits, or premium tiers, and the homepage emphasizes the local-first architecture as the default experience. Secondary databases reference a $500,000 convertible note, though no SEC Form D filing or company press release has surfaced to confirm details. Small convertible notes are common in the YC ecosystem, but the lack of corroborating documentation makes it difficult to assess the company's funding trajectory with confidence.
The open-source release includes installation scripts, environment configuration guides, and detailed changelogs. An "Unreleased" section in the project's changelog suggests active development continues on onboarding improvements and port management—the sort of incremental work that suggests a team focused on usability rather than feature sprawl. The founders frame the current release as alpha software, suitable for exploration but explicitly not ready for production environments. That's either refreshing honesty or a hedge against inevitable bugs, depending on your cynicism.
The Bet Beneath the Platform
By bundling literature search, agent orchestration, and stigmergic contribution mechanisms into a single local-first stack, ScienceSwarm attempts to make research feel less like solitary reading and more like pair programming. It's an appealing metaphor. Whether it translates into actual scientific progress—papers published, hypotheses validated, dead ends avoided—will take longer to assess than a typical product launch cycle.
The founders bring deep AI pedigrees from Meta's generative media and large language model teams. Their bet, essentially, is that scientific discovery becomes more accessible when intelligent assistants handle grunt work and when collaboration primitives remove friction. It's the same bet every AI-for-X startup makes, reframed for a domain that moves slower and values rigor over velocity.
The platform is live, open-source, and actively inviting researchers to contribute. One problem, one threaded approach at a time. Whether anyone shows up to contribute—and whether the AI proves useful when they do—remains the sort of question that won't resolve in a single funding cycle.
