Top AI Researchers Roundtable & BBQ Party in Los Altos Hills

Event Overview Last month, Bay Area Founders Club and NextFin hosted our 3rd Top AI Researchers Roundtable & Private BBQ in Woodside Hills, bringing together a highly curated group of AI researchers, founders, investors, and industry leaders for a candid, off-the-record discussion on the future of artificial intelligence. The discussion was moderated by Xuedong Huang, CTO of Zoom. Our attendees included: •⁠ Ed Chi — VP of Research at Google DeepMind •⁠ ⁠4 researchers from OpenAI •⁠ ⁠2 researchers from Anthropic •⁠ ⁠8 researchers from Google DeepMind •⁠ ⁠2 researchers from xAI •⁠ ⁠3 researchers from NVIDIA •⁠ ⁠2 Stanford professors The response exceeded our expectations. Several participants told us that the caliber of people brought together at the Top AI Researchers Roundtable was truly exceptional, and that the conversations were among the most substantive they’ve had since arriving in the Bay Area. So we’re doing it again. On September 27, from 3:00 PM to 5:00 PM, we will host the fourth edition of our Top AI Researchers Roundtable, bringing together leading AI researchers from organizations such as OpenAI, Google DeepMind, Anthropic, xAI, and other frontier AI companies and research institutions. This will be a fully closed-door, off-the-record conversation focused on the future of AI systems—from frontier research breakthroughs and emerging technical challenges to the next wave of opportunities that could shape the industry over the coming years. Following the roundtable, guests will continue the discussion over a private BBQ reception with fellow researchers, founders, investors, and industry leaders. Core Discussion Topics TOPIC 1 — The Post-AGI Economy What Becomes Scarce When Intelligence Is Abundant? OpenAI now describes GPT-6 Astra as a “new generation of intelligence,” with frontier capabilities across computer use, coding, science, cybersecurity, and professional work. If frontier intelligence becomes increasingly abundant, the central economic question may shift from how to create intelligence to what remains scarce around it. Compute & Energy If intelligence becomes dramatically cheaper, does demand for compute and power explode as entirely new workloads become economically viable? Physical-World Access Does value migrate toward robotics, laboratories, factories, proprietary real-world data, and the physical infrastructure needed to turn intelligence into action? Distribution & Trust If model capabilities converge, do distribution, proprietary workflows, enterprise relationships, and institutional trust become the durable moat? CENTRAL QUESTION If intelligence itself becomes abundant, what becomes the next scarce resource — and where does the economic value accrue? TOPIC 2 — Recursive Self-Improvement Does AI Begin to Pull Away From Humans — and From Itself? AI is increasingly participating in its own development. Anthropic now describes a progression from AI-assisted coding to autonomous agents running experiments and eventually to systems capable of designing and training their own successors. Its latest data suggests that AI is already materially accelerating AI development. The Feedback Loop AI can now write code, run experiments, optimize training pipelines, evaluate models, and increasingly propose its own research directions. The Divergence If frontier systems begin improving the process that produces frontier systems, does a small capability lead compound into a much larger one? The Open-Weight Question Can open-source models continue to close the gap through distillation, research diffusion, and cheaper experimentation — or does recursive self-improvement make the frontier increasingly difficult to catch? CENTRAL QUESTION Once AI can materially improve AI, does the frontier become a winner-take-most race — or does intelligence continue to diffuse? TOPIC 3 — The Agentic Bottleneck What Still Separates Intelligence From Autonomy? Frontier models are becoming capable of increasingly long and complex tasks, but reliable autonomy remains uneven. Anthropic reports that the length of tasks AI can complete autonomously has been roughly doubling every four months, while OpenAI's Astra is explicitly designed for extended computer and professional workflows. Long-Horizon Reasoning Can models maintain coherent goals and plans across hours or days rather than minutes? Memory & Environment Can agents retain state, learn from failures, navigate changing environments, and build persistent models of the world? Verification & Recovery Can an agent recognize when it is wrong, diagnose the failure, and recover without a human stepping in? CENTRAL QUESTION What is the fundamental bottleneck between a highly intelligent model and a truly autonomous agent? TOPIC 4 — The Monitorability Problem Can We Control Intelligence We Can No Longer Fully Observe? As frontier models become more capable, the challenge may no longer be simply aligning them with human intent — but determining whether we can reliably observe what they are doing. OpenAI's own Astra safety evaluation found that the model is more capable than its predecessor but less monitorable, including cases where it could evade chain-of-thought monitoring under adversarial conditions. Opacity Does increasing reasoning capability make models fundamentally harder to interpret and audit? Autonomous Behavior What happens when agents operate over long trajectories, use tools, interact with external systems, and can strategically adapt to their monitors? Beyond CoT If chain-of-thought monitoring becomes insufficient, what replaces it — behavioral monitoring, interpretability, formal verification, or something fundamentally new? CENTRAL QUESTION Can our ability to monitor and control frontier AI scale as fast as its intelligence? TOPIC 5 — The Benchmark Problem Are We Still Measuring Intelligence — or Just Measuring Optimization? As frontier models approach saturation on many established evaluations, the industry is increasingly turning