Track I

100 AI leaders across product, engineering, research, and design gathered to answer one question: What will actually win in the next era of AI products? From new interaction models to experience-first design, this reveals what thoughtful builders are paying attention to and what many teams are still missing. There's one particular idea they kept returning to throughout the summit, a critical insight that most of us rarely pause to consider deeply. Watch to discover what it is.

AI is converging with Blockchain. Listen to Vitalik Buterin at Devconnect 2025 explain the Ethereum roadmap in 30 min.

The common story on AI scaling is that bigger models mean better results — but the reality is more complicated. This video highlights Ilya Sutskever's perspective on where AI research is heading: Why today's LLMs generalize worse than a bright teenager and how emotions might be the missing value function in AI.

DeepSeek V3.2 offers two models: a standard version and a "special" model for advanced reasoning. This open-source release boasts impressive benchmark scores against leading AI models, including GPT-5, with an innovative sparse attention mechanism and agentic capabilities.

In a recent benchmark experiment named Alpha Arena multiple large language models trade real money ($320,000) in real markets, including crypto and U.S. equities like Tesla, Nvidia, Microsoft, and Amazon. Is AI able to outperform human traders and which models perform best? How will recursive self-improving AI systems operating in economic domains?

Brian Armstrong has led Coinbase through the explosive growth and turbulence of the crypto market. Larry Fink has led BlackRock through bear and bull markets for decades. From tokenization to regulatory scrutiny (or lack thereof), what’s next for the crypto industry? Hear Armstrong and FInk’s insights on the challenges and opportunities defining the next chapter.

Humanoid robots are exhibiting unprecedented movement. Robots from multiple labs are jogging, jumping, and presenting human level fluid task execution - indication the depth and breath of revolution industry and society is facing due to embodied AI in the next 10 years. Prepare to question reality as these advanced robots blur the line between machine and human.

Taking a look at the latest AI breakthroughs in Google, OpenAI, Deepseek and More — continuous thought machines and continual learning vs. scale

An ex-Googler claims a world-first AGI-capable system, but is this hype or reality? The video analyzes the system's three core criteria and demos, comparing it to DeepMind's prior work. Explore the controversial claims and decide for yourself.

AI is creating a new class of millionaires, and it’s not just coders or Silicon Valley insiders. This video breaks down how regular people who think like startup founders are using AI to win. The real advantage in an AI-driven world are: building feedback loops that compound faster than talent or capital. Learn the four loops that matter most and how the best builders of 2026 use them to outlearn and outpace everyone else.

What’s really happening with/inside enterprise AI agents when you “solve memory” and still don’t get leverage? The common story is that better models and better prompt engineering will unlock automation — but the reality is both simpler and more complicated. This video offers insights on primitive fluency as the missing layer for real AI agent execution and speed if teams shift into agent-legible artifacts.

As AI races forward, reshaping the very definition of economic value and outpacing human skills at every turn, there are seismic shifts ahead. From the collapse of traditional white collar jobs to a future where judgment—not labor—is the scarce commodity - the speed of technological change and the ripple effects on our lives, our identities, and our sense of direction will be brutal, surge risk-taking and reveal new steps you can take to prepare for a radically different economic landscape.

Tesla began launching of fully autonomous driverless Robotaxis in Austin, marking a pivotal moment in self-driving technology. Competition between Tesla and Waymo heats up, but one thing is for sure: society is changing. The underlying economics of the cost-efficient Cybercab's have the potential to transform urban living, vehicle ownership and future AI integrations.

Two paths. Two futures. Teaching sand to think, or being left in the dust. The biggest infrastructure supercycle in history is here. We are building the foundation of intelligence itself, with tools, models, silicon and light. This is our moonshot — an a16z Original, produced for the 2025 Runtime Summit.

When will an AI agent run a billion-dollar company with zero human employees? In this video, we are diving deep into the latest benchmarks from Anthropic and Andon Labs to answer one terrifyingly exciting question: How close are we to autonomous AI businesses? In Vending Bench AI agents like Claude, Gemini 3 Pro, and Grok 4 are given control of vending machines to see if they could turn a profit. What started as hilariously bad failures has evolved into models that are 10x-ing their money.

Nvidia's Nitrogen AI agent plays any video game unseen. This deep dive explores its revolutionary architecture, training on internet-scale video data. Discover how this groundbreaking model generalizes skills across diverse game genres, challenging AI norms across future industries.

Another dip into Project Vend: For a large part of 2025, Anthropic ran an experiment where Claude managed a small business in their office. What did they learn about the curious ways that it failed and succeeded—about the plausible, strange, not-too-distant future in which AI models might autonomously run things in the real economy? The shopkeeper (who was named Claudius) had to source products, set prices, manage inventory, and deal with customers. Things got really, really weird.

Yann LeCun offers a view of AI beyond LLMs, where intelligence is learned in abstract world models, not in token space. Instead of generating language token by token, the model predicts continuous semantic embeddings that capture what matters, while ignoring surface-level linguistic noise. This makes learning more efficient, inference faster, and reasoning more aligned with the physical world. Language is treated as a translation layer rather than the substrate of intelligence itself.

GPT-5 surprisingly surpasses human-level performance on a challenging AI benchmark. The video explores "unhobbling," a novel approach boosting AI reasoning capabilities. Learn how this technique unlocks unexpected performance gains in large language models.

Recent research like DeepSeek shows that smarter algorithms—not massive scale—can achieve near state-of-the-art performance with orders of magnitude fewer parameters, reinforcing a shift away from brute-force models. Alongside this, ideas like recursive language models are emerging, where innovations steadily compound into a new frontier while Claude Code crosses the tipping point of enabling genuine cognitive offload across disciplines.

What's really happening with disposable software and AI-native development? The common story is that vibe coding will replace enterprise SaaS — but the reality is more complicated. This video describes what the disposable software discourse gets wrong: • Why software cost collapsed but attention cost did not • How Cursor's philosophy works for developers but breaks for CIOs • What proactive AI means versus just bolting on a chatbot • Why 95% of the market needs a completely different strategy

Project Genie is Google’s experimental research prototype that lets you create and explore infinitely diverse worlds. Using Project Genie starts with World Sketching: Describe your environment and character and see an image preview of your world made by Nano Banana Pro. Then, modify your prompts to fine tune your world before entering. Project Genie is currently available to Google AI Ultra subscribers (US only, 18+), with broader availability opening up gradually.

There is a sharp split in the AI community - some seeing ClawedBot as absolutely revolutionary, others dismissing it as ‘just hype’. With networked AI agents showing abilities to interact in emergent behaviors, specializations and shared skills, the focus of all industries shifts towards building AI native approaches for all existing business models, in the form of a persistent, shareable agent layer and powerful, rapidly self-bootstrapping systems.

What's really happening with AI coding tools after December's convergence? The common story is that better models mean incremental improvement—but the reality is more complicated when the CEO of OpenAI admits he still hasn't changed how he works. This video explores why a capability overhang is widening between what AI can do and what most people are doing with it: What outperforms elaborate agent frameworks and where the real skill shift lands: from implementation to specification and review

Wes Roth explores Google's Gemini 3.1 Pro, analyzing its performance across various benchmarks. The video focuses on new "agentic" benchmarks testing real-world task completion, not just question answering. Discover how this AI model tackles complex challenges like web research and command-line operations.

This week, Anthropic and Google released Claude 4.6 Sonnet and Gemini 3.1 Pro, bringing elite-level coding and agentic capabilities to their more affordable tiers. Google also expanded into creative tools with Lyria 3 music generation and Pomelli for marketing, while xAI debuted Grok 4.2, a model powered by a "council" of four specialized agents. However, major drama erupted as Hollywood studios condemned ByteDance’s Seed Dance 2.0 for infringing on actor likenesses and IP.

A little over a month ago we released Helix 02, a single neural system that controls the full body directly from pixels, enabling dexterous, long horizon autonomy across an entire room. After cleaning up a kitchen, Helix 02 is now taking on another everyday task: tidying up a living room.

Internal leaks suggest OpenAI is developing a unified multimodal brain for future models. This technology aims to revolutionize how users interact with AI through advanced voice capabilities and new physical hardware devices.

This new open reasoning system allows autonomous vehicles to explain their actions in real-time. By thinking out loud, the AI improves decision-making and navigates rare scenarios better.

Andrej Karpathy’s new open-source “auto researcher” is a notable step toward automated AI research: a system that can iteratively modify, train, and evaluate models on its own, hinting at the possibility of recursive self-improvement. While the “intelligence explosion” framing is deliberately hyperbolic, the broader point is that AI-assisted AI research is moving from theory into practical, accessible experimentation.

Learn how to simplify AI usage by focusing on one core skill instead of managing countless apps. Discover a method for utilizing AI agents to handle complex personal tasks through conversational interfaces.

Researchers developed a system enabling a small humanoid robot to play tennis against humans by learning from amateur motion data. This approach showcases how AI can overcome complex physical challenges using imperfect information.

AI advancements are drastically changing the landscape of software vulnerabilities and exploitation. Witness how artificial intelligence can identify and exploit complex security flaws, redefining traditional research methods. Explore the implications of automated exploit development and AI-generated malware.

Elon is building a single AI chip architecture - used across cars, robots, and humanoid systems - expanding into space via solar-powered satellite constellations that provide continuous compute with near-free energy and cooling. His vision for the future is a fully vertically integrated system - from chip design and manufacturing to deployment across Earth and orbit - creating a self-improving, unified AI infrastructure reimagining what will be possible in the future.

A leaked document revealed an unreleased AI model with unprecedented capabilities in coding and cybersecurity. The investigation examines the potential risks and massive computational requirements associated with this next-generation technology.

An accidental leak revealed the inner workings of Anthropic's coding assistant. Discover the massive implications for future software development and how AI is changing legal definitions of copyright. The footage also showcases unreleased features, including autonomous agents and a hidden virtual pet system.

Researchers analyze neural activity within AI to see if recognizable patterns for human emotions exist. By tracking these internal mechanisms, they test how these patterns influence responses during challenging tasks.

Scientists at Google DeepMind developed a new AI system called Aletheia that can generate and refine ideas to help produce genuinely novel research, even contributing core parts of scientific papers. While still limited compared to true breakthroughs, it marks a major step from solving structured problems to assisting—and sometimes autonomously performing—real-world research.

A significant software project was recently leaked, triggering a rapid community effort to recreate its functionality from scratch. This process involves leveraging specialized artificial intelligence tools to mimic the original codebase without violating copyright laws. Explore how this approach reshapes modern software development and individual potential.