S⁶ is a research salon and fellowship for top-tier engineers transitioning into scientific roles. We build methods, systems, and benchmarks that make scientific discovery faster, safer, and more reliable.
We aim to raise the floor of human well‑being to a next‑century baseline by accelerating scientific capability with modern AI, especially large language models. We pursue open research, reproducible artifacts, and practical systems that compound across the scientific stack.
Evaluation, drift detection, unlearning, data pipelines, and scalable serving.
Agentic workflows, lab‑grade reasoning, and domain‑grounded modeling.
Open, fair, and durable tests that reflect real scientific tasks.
Transparent reporting, bias/robustness audits, and reproducibility standards.
Our projects explore the frontier between large language models, scientific reasoning, and new forms of machine understanding. Fellows and collaborators contribute to open, high-impact efforts spanning systems, modeling, and evaluation.
Integrating diagram and figure comprehension into LLM-based research agents. The goal is to enable retrieval and reasoning over scientific figures—so agents can interpret, reference, and cite visual evidence as naturally as text.
LLM Systems · Image Understanding · RAG
Developing compact diffusion-transformer (DiT) architectures for biological sequences. We study whether small-scale DiTs (4–20M parameters) can rival current hybrid transformer-SSM models for DNA representation learning.
Diffusion Models · Biological Foundation Models
Studying unlearning as a probe for knowledge regeneration—first removing a concept from a model, then testing whether it can rediscover it through reasoning. A position paper was presented at NeurIPS AI4Science 2025.
Evaluation · Knowledge Dynamics
Have an idea that aligns with our mission? We welcome proposals from fellows and collaborators. If the topic is beyond our core expertise, we’ll do our best to connect you with someone who can help.
Designed for elite engineers ready to transition into scientific research. Expect weekly paper/prototype cadence, strong mentorship, and public artifacts. Funding status: early stage; current cohorts may be unpaid while we build partnerships and donors.
Full‑time, 6–12 months. Co‑mentored projects targeting papers, benchmarks, and open‑source releases.
12 months at ~20+ hrs/week. Ideal for engineers exploring the research path while employed or transitioning.
Life today is incomparably better than in centuries past—thanks to technology rooted in basic science. Few of us would choose a world without modern medicine, climate control, or global communication. Our premise is simple: when science advances, daily life follows.
No amount of willpower or wealth can overcome missing scientific capability. History is full of examples—from quests for longevity to cures that society simply wasn’t yet able to make. Our work acknowledges this: to raise the ceiling of what’s possible, we must push the frontier of methods and tools.
Large language models can accelerate science by scaffolding reasoning, automating literature synthesis, orchestrating experiments, generating code and analyses, and stress‑testing hypotheses. We build systems and evaluations that make these capabilities reliable, safe, and truly helpful to working scientists.
S⁶ follows the spirit of early open AI research organizations: open by default, research‑first, safety‑conscious. We focus on LLM systems, AI for Science, and foundational infrastructure that improves the reliability and utility of modern models in scientific workflows.
Robert Yang (Stanford BS with Distinction, Terman Scholar; MS in Computer Science, AI focus) founded S⁶ after personally transitioning from engineering to research. His interests span data drift, AI for science, ML systems, foundation models, and unlearning. Previously, he worked as a Research Scientist at a biology foundation model startup (Rosa Bio), and he built LLM model serving and infrastructure at scale (ex‑AWS).
Interested in a fellowship or in partnering as a mentor, donor, or sponsor? Send a short note with your background and goals.
Prefer email? Reach Robert directly:
When funding solidifies, we may introduce paid cohorts. Until then, fellowships operate as mentorship and open research collaborations.