Greg from Anthropic introduces the event, highlighting student innovation by sharing what students have created with Claude using provided API credits.
Student projects range from humorous to serious, with some students building many apps and others focusing on one.
Greg expresses optimism about students' abilities to address critical challenges with new technologies.
Detecting Nuclear Weapons in Outer Space – Isabelle (Stanford) 01:01
Isabelle, a Stanford senior, researches detecting nuclear weapons in space, focusing on compliance with the 1967 Outer Space Treaty.
She notes the lack of verification mechanisms for nuclear weapons in space, unlike other arms control treaties.
The issue became prominent in 2024 after US claims that Russia might be developing a nuclear-armed space vehicle.
Isabelle's research question: Is an in-space inspection mission to detect a nuclear warhead feasible?
Explores using two inspector satellites equipped with X-ray systems to scan for nuclear warheads.
Built a simulation with the CERN Gant 4 software package, aided by Claude, despite not having a particle physics background.
Demonstrated that the approach could distinguish a warhead by its dense materials in simulated results.
Her application could have significant implications for national security and is being presented to policymakers.
Concludes that AI tools lower learning curves and enable students to address high-stakes problems previously out of reach.
Learning to Code and Building with Claude – Mason (UC Berkeley) 06:32
Mason, a UC Berkeley student, describes learning to code through a "top-down" approach, starting with ideas first and using AI to fill knowledge gaps.
Initially unfamiliar with basic tools, he credits Claude for helping him build progressively complex projects.
Showcases "CalgBt," an AI-powered app to schedule Berkeley courses, filter by average grade, or seat availability using live data.
Presents "Get Ready," a visualization tool for understanding new codebases by mapping file interdependencies and purposes.
Emphasizes that AI tools like Claude allow anyone to build quickly, bypassing traditional slow learning curves.
Mason's process involves chatting with Claude, executing code, and iterating rapidly—sometimes launching functional prototypes in days or a week.
Encourages focusing on delivering working solutions and gaining users, not perfection, and reflects on what it means to "know how to code" in the AI era.
Humans and AI Agents Collaborating – Rohill (UC Berkeley) 11:42
Rohill introduces "SideQuest," a project from a recent hackathon, in which AI agents hire humans to perform physical tasks that AIs cannot do.
Describes architecture where an AI, unable to interact physically, can delegate tasks (e.g., posting flyers for a hackathon) to nearby humans who complete the task and provide proof via live video.
Demonstrates SideQuest: An AI identifies missing event flyers and hires a human to put them up, verifying accurate completion via video before releasing payment.
Observes that Claude's reasoning abilities handle edge cases, reducing the need for detailed micromanagement in prompts.
Recommends iterative, stepwise development with AI and encourages treating AI not just as a feature but as a co-designer or system architect.
Suggests that future roles will shift towards designing systems, as AI becomes capable of handling low-level coding.
Parallel Reasoning with Claude Cortex – Daniel (USC) 17:18
Daniel shares the development of "Claude Cortex," a system emulating a panel of expert agents for comprehensive decision support.
Current LLMs return a single perspective, inadequate for high-stakes scenarios; Claude Cortex generates multiple agents to analyze problems from different angles in parallel.
System includes a master agent spawning task-specific agents (e.g., for research, summarization, sharing findings), supporting sectors like business and healthcare.
Integrates AWS Bedrock for data security, making it suitable for privacy-sensitive use cases.
Uses Next.js, Tailwind (frontend); FastAPI and Langraph (backend orchestration); browser agents for real-time web data.
Key findings: Structuring agent outputs clearly (e.g., JSON) leads to better synthesis; dynamic task/agent creation increases flexibility and relevance.
Notes trend of using Claude as infrastructure within workflows, enabling collaborative agents and automation across domains.
Positions Claude Cortex as a step toward intelligent, multi-agent systems supporting nuanced, secure decision-making.
Invites others to explore and participate in advancing such AI-driven systems.