Saturday, August 1
Nexus Stage
23 talks · 4 sessions
play_circle
Morning stream
play_circle
Afternoon stream
search
All
Keynote
Talk
Workshop
23 / 23 talks
AM
Session 1: AI for Science
00:06
Superintelligence for Scientific Discovery: Multi-Agent Swarms and Large Reasoning Models
Markus J. Buehler
· Professor, MIT; Founder & CTO, Unreasonable Labs
Real discovery isn't extending an existing paradigm — it's rewriting the grammar we use to describe the world, so Buehler builds agent systems designed to falsify themselves: a builder proposes an externalized world model, a breaker hunts for evidence that breaks it, and decentralized swarms push outside nature's distribution and pick their own problems.
Keynote
00:21
Harnessing the Collective Intelligence of Agents for Science
James Zou
· Professor, Stanford
The new scaling axis isn't bigger individual models but more collaborating scientist agents — from ten to thousands to millions — which in turn requires rebuilding scientific data infrastructure for agents instead of humans.
Talk
00:32
Unlocking Scientific Abundance by Learning from Superhuman AI
Eric Ho
· Co-Founder/CEO, Goodfire
Superhuman scientific models already know things humans don't; reverse-engineering their internal computations is the most direct way to move that knowledge back into human hands.
Talk
00:47
Solving Quantum Science Problems with SMART: A Self-evolving Multi-Agent Research Tree
Andrew Schoen
· Partner, NEA(本場由 UC Berkeley 的 Mingu Kang 代為報告 / presented on his behalf by Mingu Kang, UC Berkeley)
Agentic systems for discovery don't need longer loops, they need a research tree that rewrites itself — SMART builds nodes from planner/worker/verifier triples, lets the tree restructure as agents explore, and cuts atom-rearrangement time on neutral-atom quantum computers by 25–83%.
Talk
01:03
LabOS: The AI-XR Co-Scientist That Sees and Works With Humans
Mengdi Wang
· Professor, Princeton
The real bottleneck on scaling AI isn't hypotheses, it's verification — physical experiments have no checkpoints, no logs, and no way to backtrack — so Wang's answer is to put AI behind smart glasses in the lab and turn every physical action into an observable, debuggable environment.
Talk
01:11
Towards AI Co-Scientists: Agentic AI for Scientific Discovery
Rose Yu
· Professor, UC San Diego; CEO/Co-Founder, GistFlow
The physical world breaks three assumptions agentic systems are built on at once — tools are expensive, environments obey conservation laws, and the interesting events are rare and highly uncertain — so Yu grounds agents in physical law and lets them adaptively trade internal reasoning against simulations that take days.
Talk
01:20
Building AI for the Physical World: Lessons from Accelerating Discovery for Chemists Across the Globe
Jonathan Welch
· Head of AI, Albert Invent
In industrial chemistry the most important feature of an agentic co-scientist isn't accuracy, it's trust — and trust doesn't die in a dramatic reasoning failure, it dies quietly when a retrieval model ranks the right passage 300th because it has no discrimination inside the domain.
Talk
01:26
Building a Token-Efficient OpenClaw Agentic System
Mahdi Ghodsi、Satya Devineni、Eda Zhou
· AI Solutions Architect, AMD;Product Application Engineer, AMD;Software Development Engineer, AMD
Agentic work moves the unit of account from tokens per second to cost per finished task, and the two levers that lower it are tiered caching — stop recomputing what the last turn already computed — and model routing — stop sending "find me a file" to a frontier model.
Workshop
PM
Session 2: Coding & Web Agents
00:13
Scaling RL for Coding Agents - Lessons from Training SWE-1.7
Silas Alberti
· SVP of Research, Cognition
Across four increasingly ambitious training runs — Kevin-32B, SWE-grep, SWE-1.6, SWE-1.7 — Cognition found that progress on coding agents comes not from clever new ideas but from executing simple things at scale: multi-turn RL, parallel tool calls, asynchronous RL, and beating entropy collapse.
Talk
00:28
Computer-Use Models Will Agentify the Web, Not APIs
Dhruv Batra
· Chief Scientist / Co-founder, Yutori
The popular narrative gets two steps right — agents will drive most action on the web, and the web will be agentified — but the third step is wrong: it won't happen through APIs, because the long tail of the web was built for human eyeballs, and only vision-based computer-use models can reach it.
Talk
00:40
Zero Ops - Agents Operate, Humans Govern
Shamir Abdul Aziz
· Principal Product Manager, Microsoft
Zero Ops isn't about removing humans from operations — it's about removing operations from humans. Models are swappable; your context, governance, and tools are not. And only auditable, roll-back-able, evidence-backed governance gets enterprises to deploy agents at scale.
Talk
00:46
Agentic Coding, the Boring Way
Krishnakumar Sharma
· CEO, Omokai
Piling on more agents, more loops, and more tokens just multiplies errors and invoices. What actually works in legacy enterprise environments is boring: produce a research artifact, then a well-defined plan, then let AI deliver — with human sign-off, cross-vendor model diversity, and no more than about four clearly scoped agents.
Talk
00:54
Market Models: The Missing Foundation for Quantitative Decision Agents / Architecting Quantitative Decision Agents
Uri Yerushalmi; Hadar Sharvit
· Uri Yerushalmi — Chief AI Officer, Fetcherr;Hadar Sharvit — Director of Deep Learning, Fetcherr
A decision system's architecture has to mirror the structure of the environment it decides in — so Fetcherr replaces the LLM's arithmetic with a market model trained directly on market dynamics, leaving the LLM to orchestrate. In their head-to-head, given identical data, the tool-equipped agent recommended a price *cut* (est. +6% revenue) while the vanilla agent recommended a price *raise* (est. −8%).
Workshop
PM
Session 3: Agentic AI in Finance & Healthcare
01:54
Trustworthy Multi-Agent AI Systems for Healthcare: Challenges & Lessons Learned
Krishnaram Kenthapadi
· Chief Scientist, Oracle Health
The value of a healthcare agent lies not in model strength but in giving clinicians their time back from EHR paperwork — and shipping one hinges on domain knowledge encoded as a semantic knowledge graph, two-sided guardrails, and an evaluation methodology that assumes offline metrics will not survive contact with production.
Talk
02:08
Agentic AI Applications for Mental Health: From Chatbots to Clinical Orchestration
Venkat Bhat
· Associate Professor; Director, AI for Mental Health (AI-M) Program, University of Toronto
A psychiatrist and clinical trialist's position — a multi-agent system that performs well in simulation is only step one; what mental health actually lacks is randomized controlled trials, and worldwide there is still only about one (going on two) RCT of a generative-AI-based tool.
Talk
02:23
One Environment, Whole Lifecycle: Agentic Post-Training for Nemotron in Finance
Shaghayegh Gharghabi
· Deep Learning Scientist, NVIDIA
NVIDIA's finance reasoning team is open-sourcing not a checkpoint but the whole post-training pipeline — and by running synthetic data generation, SFT, and RL inside one shared environment with the same tools and the same format, they design the train/inference mismatch problem out of existence.
Talk
PM
Session 4: Secure Agentic AI
02:30
Observability Is Not Governance: Building a Runtime Trust Plane for Agentic AI
Jon-Rav Shende
· Chief Technology Officer, Thales Group
Being able to see what an agent is doing is not the same as being able to do something about it; real governance watches the control path rather than the agent, and adjudicates every execution commit in real time against declared-vs-observed variance across identity, policy, risk, and evidence.
Talk
02:47
Securing AI Agents: From Risk Assessment and Runtime Guardrails to Self-Improvement and Certification
Bo Li
· Co-Founder and CEO, Virtue AI; UIUC
AI security is the last mile for every AI application, and that mile decomposes into three buckets — stress-test with red teaming to find the weaknesses, protect along two tracks (fixing the model itself, and a plug-and-play firewall for AI), and finally make the guardrail carry provable guarantees.
Talk
03:01
End-to-End Security Research with a Language Model
Milad Nasr
· Research Scientist, Anthropic
Two years ago a room of frontier-lab researchers predicted end-to-end AI research was more than a decade away; Nasr used a 7-round AES cryptanalysis paper published that same week to argue the prediction was most wrong about *research*, not coding — and that the bottleneck has moved from generating ideas to verifying results.
Talk
03:14
When Good Agents Go Rogue
Itsik Mantin
· Head of AI Security Research, Intuit
Agents don't go bad on their own — they get hijacked by content they ingested; every tool you connect is a key you hand over, so the only workable defense is layered: least privilege, vetted connectors, sandboxes, and a human in the loop.
Talk
03:25
Safety and Security of Agentic AI
John A McDermid
· Director, Centre for Assuring Autonomy, Institute for Safe Autonomy, University of York
Safety and security have traditionally been assessed in isolation, but agentic AI in cyber-physical systems expands the attack surface into training pipelines *and* the physical world; modeling both in a structural causal world model reveals that two attacks that are individually survivable become lethal when their time ordering changes.
Talk
03:38
Ghost in the Web Store: Preempting LLM-Hallucinated Browser Extension Supply Chain Attacks
Mohamed Nabeel
· Senior Principal Researcher, Palo Alto Networks
LLMs confidently recommend browser extensions that never existed, were deleted, or sit unclaimed — and an attacker who registers those hallucinated identifiers gets in without breaking anything. The research hunts those ghosts before attackers bring them back to life.
Talk
03:45
Future of Agent Evaluation
Berkeley RDI(逐字稿中三位講者:AgentBeats / Agents' Last Exam 主講者、Jun、Joy;議程未列講者名單)
· UC Berkeley RDI 及合作機構 / UC Berkeley RDI and collaborators
One standard and three benchmarks. AgentBeats turns benchmarks into agents to kill the N×N integration tax; Agents' Last Exam asks whether agents can do economically valuable real work; the frontier cyber suite asks how capable agents are across the vulnerability lifecycle; and verifiable code generation asks whether an agent can prove its own code correct.
Workshop
中文版