Sunday, August 2
Plenary Stage
17 talks · 7 sessions
play_circle
Morning stream
play_circle
Afternoon stream
search
All
Talk
Panel
Fireside
Session
17 / 17 talks
AM
Opening Remarks
00:00
Opening Remarks (Day 2)
Jennifer Chayes; Dawn Song
· Jennifer Chayes — Dean of CDSS, UC Berkeley;Dawn Song — Professor, UC Berkeley; Co-Director, Berkeley RDI; VP of AI Research, Meta Superintelligence Labs
Chayes argues the future of AI has to be open — open source, open weights, open data where possible — for its benefits to be broadly shared; Song uses cybersecurity benchmark data to show frontier AI capability rising sharply, placing us at a critical point where action can't wait, and frames Berkeley RDI's three pillars (research, education, community & entrepreneurship) as its answer.
Session
AM
Session 1: Enterprise AI
00:26
Enterprise AI - Agent Governance
Rao Surapaneni
· VP/GM, AI Search & Specialized AI, Google Cloud
Enterprise agent adoption is stuck between two bookends — personal-productivity agents and high-code agents built by the CTO org — and the vast middle stays locked not because models are too weak but because governance is missing; the three pillars are visibility, control and security, and the way to deliver them is to give every agent a real identity and move permission checks from static configuration to task- and context-aware runtime enforcement.
Talk
00:41
Enterprise AI
Adarsh Hiremath
· Co-CEO, Mercor
Every company now wants to own its own intelligence, but whether you train your own model or engineer a harness around a frontier one, the common prerequisite is a good eval — and most enterprise AI projects never reach production precisely because there is no way to measure ground truth, leaving teams guessing and iterating on anecdotes.
Talk
00:50
Off-the-Shelf AI Hit a Wall. Here's What HubSpot Did to Solve It.
Duncan Lennox
· Chief Product & Technology Officer, HubSpot
Three years ago HubSpot found off-the-shelf co-pilots couldn't hold up at their scale, so they built their own harness and backfilled their own context into it — engineering adoption went 80% → 100%, velocity gains 51% → 60%, and reliability went *up*, not down; the unexpected part was that everything they learned about building *with* AI turned into what they needed to build *for* customers.
Talk
01:04
Panel: Enterprise AI
Aaron Jacobson (moderator); Rao Surapaneni; Adarsh Hiremath; Duncan Lennox; Anahita Tafvizi; Surojit Chatterjee
· Aaron Jacobson — GP, NEA;Rao Surapaneni — VP/GM AI Search & Specialized AI, Google Cloud;Adarsh Hiremath — Co-CEO, Mercor;Duncan Lennox — Chief Product & Technology Officer, HubSpot;Anahita Tafvizi — Chief Data & AI Officer, Snowflake;Surojit Chatterjee — Founder/CEO, Ema
The five agree that enterprise agent adoption is still around 10%, but converge on the view that the bottleneck is now organizational rather than technical — ROI has to be measured as workflow-level cycle-time compression rather than token consumption, trust is earned through verifiability rather than promises, and since the model layer is commoditizing, the architectural investment that matters is the layer that lets you swap models at will.
Panel
AM
Fireside Chat
01:36
Fireside Chat: Ali Ghodsi × Andy Konwinski
Ali Ghodsi; Andy Konwinski
· Ali Ghodsi — Co-Founder & CEO, Databricks;Andy Konwinski — Co-Founder, Databricks; Perplexity; Laude Ventures
Two Databricks co-founders trace a line from a 2009 Berkeley research group to the 2026 benchmark crisis — concluding that models and harnesses will increasingly be hill-climbed by optimizers, so the place humans should spend their brainpower is writing good evals; that benchmarks should be versioned like software instead of thrown away and rebuilt each round; and that open science needs a new "lab of labs" structure to stand a chance alongside the closed frontier labs.
Fireside
PM
Session 2: Frontier Research
00:08
The Eureka Machine: Recursive Superintelligence for Science
Richard Socher
· Founder/CEO, Recursive Superintelligence
If the ultimate invention is the one that automates all future inventions, the road there runs through AI automating AI research first — and their recursive-self-improvement system has already beaten years of community effort on three public benchmarks.
Talk
00:21
The Future of Personalized Universal Agents
Ed Chi
· VP of Research, Google DeepMind
Framing three decades as "only three ideas per era that really mattered" — indexing, vector space models, deep learning; then sequential transduction, chain of thought, post-training — he argues the next decade moves from ranking to personalized reasoning, and that training must shift from bottom-up RL toward top-down teaching, the way we teach children.
Talk
00:33
Combining Experiments, Large Language Models, and Theory to Discover Quantum Materials
Ekin Dogus Cubuk
· Co-Founder, Periodic Labs
No amount of intelligence can think its way to the next scientific breakthrough — the universe is too complex, and the entire history of superconductivity says so. The right question isn't whether AGI can zero-shot a discovery, but where agents belong inside the scientific method we've used for centuries.
Talk
00:42
Personal AI and Continual Learning: New Frontiers in Agentic AI
Igor Babuschkin
· Co-Founder/CEO, River AI
Coding agents worked because rewards were verifiable; personal AI has no such luxury. He breaks the gap into five unsolved problems — RL for the use case, personalization, memory, privacy/security, and cost — and argues the same shift is a rare chance to hand control of AI back to the individual.
Talk
00:53
Panel: Frontier Research
Richard Socher, Ed Chi, Ekin Dogus Cubuk（主持 / Moderator: Igor Babuschkin）
· Richard Socher — Founder/CEO, Recursive Superintelligence / Ed Chi — VP of Research, Google DeepMind / Ekin Dogus Cubuk — Co-Founder, Periodic Labs / Igor Babuschkin — Co-Founder/CEO, River AI
Every path past coding agents hits the same wall — rewards stop being binary. The three panelists answer from different directions (teach top-down, instrument everything, build environments that can't be reward-hacked), and the heaviest line of the session is Ed Chi's: model capability plus harness together have hit a plateau nobody can seem to get over, and that plateau is the industry's real bottleneck.
Panel
PM
Session 3: Agentic AI Developer Platforms
01:25
Panel: Agentic AI Developer Platforms
Matt White、Dmytro Dzhulgakov、Ivan Burazin、Mazin Gilbert(主持:Megan Morrone)
· Matt White — Former Global CTO of AI, Linux Foundation; CTO, PyTorch Foundation / Dmytro Dzhulgakov — Co-Founder & CTO, Fireworks AI / Ivan Burazin — Co-Founder & CEO, Daytona / Mazin Gilbert — Executive Director, Agentic AI Foundation & Linux Foundation(主持:Megan Morrone — Editor of Technology, Axios)
Four infrastructure and open-source leaders converge on the same verdict — the "open vs. closed weights" debate is asking the wrong question. The recent agent security incidents had nothing to do with weights and everything to do with what surrounds the model: the harness, the containment layer, the gateway, and the guardrails an enterprise is supposed to own itself.
Panel
PM
Session 4: Agentic AI in Finance & Legal
02:00
Advancing the State of the Art: The Frontier of Enterprise Agentic AI
Milind Naphade
· SVP, AI Foundations, Capital One
Capital One started on agentic AI more than two and a half years ago and had a multi-agent system in production by last January. Its differentiation isn't model choice — it's customizing the entire AI stack with proprietary data, and putting an independent evaluator agent in the loop to catch plans that violate policy.
Talk
02:09
Building Infrastructure for the Agentic Economy
Nikhil Chandhok
· Chief Product & Technology Officer, Circle
Every assumption baked into today's payment rails — finality, privacy, trust living at the application layer — presumes a human is one step away to govern. When one agent spawns a hundred sub-agents that each need value and rules, those assumptions all break. Programmable stablecoin money is the rail that survives at agent scale.
Talk
02:17
Reimagining Banking in the AI Era
Faraz Shafiq
· Head of AI, Wells Fargo
Every interface shift in banking — ATM, telephone, web, mobile app — redistributed value, and this time the interface itself becomes the agent. Wells Fargo's roadmap runs AI as a tool → AI as a teammate → autonomous AI, and the differentiator was never the model, since everyone has the same one. It's everything around the model.
Talk
02:32
Panel: Agentic AI in Finance & Legal
Nikhil Chandhok、Faraz Shafiq(主持:Matt Carbonara)
· Nikhil Chandhok — Chief Product & Technology Officer, Circle / Faraz Shafiq — Head of AI, Wells Fargo(主持:Matt Carbonara — Investor, Mayfield)
Finance and legal aren't just "harder agent problems" — they have a fundamentally different verification cost structure. Easily verified tasks (coding) automated first; a credit decision may not be verifiable for five years. That verification axis determines what agents take over first and what they take over last.
Panel
PM
Startup Spotlight
02:59
Startup Spotlight
Startup Spotlight
· Featured Startups: Narada AI, cognee, Nimblemind, AgntID, RELAI, Headroom, Founding Dev, ArmorIQ, Keenable AI, H Company, Ludo Robotics, Nimble
The summit's closing session — 12 agentic AI startups whose topic distribution is itself an ecosystem map: memory and context efficiency (cognee, Headroom), runtime governance (AgntID, ArmorIQ), continual learning and verification (RELAI, Nimble), and driving existing interfaces directly instead of waiting for APIs (Narada, H Company).
Session
中文版