Saturday, August 1
Atlas Stage
17 talks · 3 sessions
play_circle
Morning stream
play_circle
Afternoon stream
search
All
Keynote
Talk
Workshop
17 / 17 talks
AM
Session 1: Foundational Capabilities
00:01
Agentic Modeling via Internalizing Agent Harnesses
Jianfeng Gao
· Technical Fellow & Corporate Vice President, Microsoft Research
A language model is a talker; the agent harness is what turns it into a doer — and the next step is distilling the harness's capabilities back into the model itself, because harnesses are to agentic modeling what internet data was to pre-training.
Keynote
00:17
World Models for Physical AI Simulation
Sanja Fidler
· Associate Professor, University of Toronto; Former VP of AI Research, Nvidia
The bottleneck in robotics isn't the policy, it's simulation — real robots run at physical time, while generative world models can *imagine* unlimited long-tail scenarios on GPUs faster than real time; in one year this went from "5 seconds of video per 5 minutes of GPU time" to real-time interactive, and now runs on a consumer GPU.
Keynote
00:32
When Language Models Learned to Reason
Giambattista Parascandolo
· Research Fellow, OpenAI
He dug out the actual four-to-five-year-old internal slides from OpenAI's earliest discussions of reasoning — back when giving a model more time produced the identical answer — and showed the humble evidence that convinced the team to scale it up: some ugly slides about counting bags of potatoes.
Talk
00:41
Redefining the Token Efficiency Frontier with Diffusion LLMs
Aditya Grover
· Co-Founder/CTO, Inception Labs
Every major leap in AI came from parallelizing something — GPUs parallelized matrix multiplication, Transformers parallelized training, and diffusion parallelizes *generation* itself; now that intelligence is abundant, the new currency is intelligence per watt.
Talk
01:08
Turning Chatbots into Agents: What Modern RL Looks Like
Lovre Pesut; Muhammad Hashmi
· AI Engineer, Daytona; DevRel, Daytona
Same harness, same tools — GPT-2 falls apart where Kimi K3 fixes your bug, and the difference is reinforcement learning; but modern RL's real bottleneck isn't the algorithm, it's who supplies the tens of millions of disposable sandboxes, and the fact that *environments* have become the new training data.
Workshop
PM
Session 2: Robotics & World Models
00:14
Outplaying Elite Table Tennis Players with an Autonomous Robot
Peter Stone
· Chief Scientist, Sony AI; Professor, UT Austin
Sony AI's Ace is the first autonomous robot to beat professionally ranked players head-to-head under official rules — proof that AI agents can now reach expert-level performance at fast, real-time decision-making in physical space, and it took perception, reinforcement learning, and custom hardware pushed to their limits together, not any single breakthrough.
Keynote
00:29
Trustworthy Agents in the Real World: Physical Autonomy Lessons for the Agentic AI Era
Vincent Vanhoucke
· Distinguished Engineer, Waymo
Waymo *is* an agentic AI company — its agents just happen to be big hunks of metal on four wheels — and the potholes robotics hit over twenty years (each "nine" of reliability costs a system redesign, Level 3 is the worst place to be, step-wise optimization always drifts) are now hitting agentic AI verbatim; the shared answer is that you need a world model first.
Keynote
00:44
Real-World Reasoning Agents
Trevor Darrell
· Professor, UC Berkeley
For agents to work in the physical world they need to *see* (catch the one small difference between two images that matters), *feel* (react to touch in real time rather than plan from vision alone), and *reason about motion abstractly* (predict SE(3) pose trajectories instead of every pixel).
Talk
00:55
Making Autonomy Autonomous: Toward Mental Models for Discovery and Intuition
Manmohan Chandraker
· Professor, University of California, San Diego
Physical AI's edge cases never run out, so let a team of agents run the discover-experiment-validate-write loop themselves; what keeps that from being hallucination is that every claim traces back to a real GPU run and a real line of code — and the human's role shifts from executor to the source of *tacit expertise* and *intent*.
Talk
01:04
Scaling Sidewalk Autonomy with World Models
Bolei Zhou
· Associate Professor, UCLA; Chief AI Scientist, Coco Robotics
The bottleneck in sidewalk autonomy isn't a shortage of models — it's having no fair way to evaluate them before deployment; the fix is to treat real-world video as source code, reconstruct it into interactive simulation, and grow many "digital cousins" from a single clip for training and evaluation.
Talk
01:14
Open-Source Agent Investigations: Security Arena, Distilled Traces, and Auto-Optimization
Devina Jain; Zach Mueller; Chuan Li
· Research Engineer, Lambda; Head of Developer Relations, Lambda; Chief Science Officer, Lambda
Lambda ran three open-source experiments against the same question — *what do we actually know about agents?* The attack/defense arena found that winning prompt injections are transferable rhetorical shapes, not domain-specific exploits; distilling 300M tokens of traces found that the harness, not the model, often decides performance; and giving Claude Code an experiment tracker found that what matters isn't how smart it is but whether it writes things down.
Workshop
02:14
Building Agentic Apps End-to-End with Replit Agent
Brandon Middleton
· Head of Education, Replit
A workshop billed as "build apps with Replit Agent" spends three quarters of its time on education — because once information *and* intelligence are abundant, a university built on trading time served for learning, credentials for capability, and sorting for teaching stops working; and the point of vibe coding isn't saving keystrokes, it's letting a 15-year-old, a musician with no CS background, or Shaquille O'Neal turn an idea into something that runs.
Workshop
PM
Session 3: Frameworks & Dev Platforms
03:14
Graphs Are the Knowledge Layer for Agentic AI
Philip Rathle
· CTO, Neo4j
Knowledge, context, semantics, memory, and ontology are facets of one thing — an enterprise knowledge layer — and graphs are its natural implementation, because real-world data already arrives as networks, hierarchies, and paths.
Talk
03:29
Building Punk Software: An Open Agent Stack
Josh Albrecht
· Co-founder and CTO, Imbue
AI may not pose an existential risk to the species, but it does to *you* and your livelihood, because the frontier stack is controlled by a handful of companies — the answer is to open the whole stack and build it yourself.
Talk
03:44
vLLM: Building Open and Efficient Inference for Agents
Woosuk Kwon
· Co-Founder & CTO, Inferact
Agents barely changed the inference API but changed everything underneath it — so vLLM is rebuilding along three axes: workload-aware model parallelism, dynamic hierarchical KV cache, and enough hardware backends to turn any compute on earth into tokens.
Talk
04:01
Building Frontier Inference and Training Infra for Agent: A Case Study of SGLang and Miles
Banghua Zhu
· Co-Founder, RadixArk
2026 is the year of agentic infrastructure — inference is squeezed by cache reuse, trillion-parameter models, and million-token context, while RL training is squeezed by environment diversity and train–inference mismatch; RadixArk answers each with SGLang and Miles.
Talk
04:14
Speaking the Same Language: NLIP for Agent Interoperability
Ranjan Sinha
· IBM Fellow, CTO & VP for watsonx, Enterprise AI, IBM
Traditional protocols demand that both endpoints and the wire speak the same language and break the moment a schema changes; NLIP drops that assumption — if there's intelligence at both endpoints to translate, agents can keep their own languages, and NLIP moves intent across in just five fields.
Talk
中文版