OpenAI talks GPT-6 Astra and Millenium Prize, researchers create WeWorm exploit & IBM’s US Open app
This episode explores OpenAI's groundbreaking GPT-6 Astra model and its resolution of the Navier-Stokes equation, the AI-powered fan experiences developed by IBM for the US Open, and the concerning implications of AI-assisted cybersecurity threats like the WeWorm zero-click exploit.
The podcast episode features Tim Hwang and David Zax, joined by Bri Kopecki, Aaron Baughman, and Olivia Buzek, discussing recent advancements from OpenAI, new cybersecurity threats, and IBM’s AI tools for the US Open.
OpenAI’s GPT-6 Astra and the Navier-Stokes Breakthrough
The discussion opens with OpenAI’s new model, Astra, showcased in impressive demos converting 2D to 3D renderings and performing tasks like ordering food. The panel debates whether Astra represents Artificial General Intelligence (AGI). Aaron Baughman notes the lack of a converged AGI definition, suggesting it might appear as a “really good digital employee.” He highlights the immense computational cost, estimating 100,000 Blackwell NVLink systems for training, equating to power for an entire US region. Olivia Buzek questions if such massive models are solving problems in the “right way” or simply by “throwing bigger and bigger models” at them, advocating for more accessible agent models. Bri Kopecki, from a customer-centric perspective, states that current enterprise use cases typically don’t require models of this magnitude, though speed remains critical.
Astra achieved a significant breakthrough by producing a proof related to the 3-D Navier-Stokes equation, one of mathematics’ seven “millennium problems” unsolved for 90 years, related to fluid dynamics (the “exploding water problem”). The AI solved this using 10,000 AI agents working collaboratively “like a swarm” with a coordinator, exploring the problem space. This endeavor reportedly cost “tens of millions of dollars” in compute for a $1 million prize. The panelists emphasize the critical role of human-AI collaboration: humans provided the initial hunch, directed the AI’s focus after an intermediate solution was found, and made strategic decisions throughout. Bri underlines that AI, however powerful, still requires “human intellect, human creativity, and human collaboration.” The initial verification of the proof took 88 hours, later refined to 17 hours using the Lean mathematical proof assistant. This breakthrough is seen as a harbinger for solving other major global challenges in health, medicine (e.g., cancer cures), and climate change.
IBM’s AI Innovations at the US Open
Aaron Baughman details IBM’s over 30-year partnership with the USTA, bringing data insights to over 14 million tennis fans. For the 2026 US Open, IBM introduced advanced AI-powered fan features:
- Pre-match Likelihood to Win: Predicts match outcomes using two classical machine learning models—XGBoost tree models and logistic regression—considering player rankings, media sentiment, head-to-head stats, and surface play. It boasts around 60% accuracy.
- Live Likelihood to Win: A real-time system that updates win probabilities after every point, using a momentum equation with decays and situation-based boosters.
- Key Moments Identification: Utilizes Confluent messaging and a “Confluntic marketplace” to identify pivotal match moments and generate explanatory text.
- Match Chat: A conversational, personalized experience allowing fans to ask questions during or after a match. It queries various data sources through a “myriad of graph-laying type agents,” with three types racing to provide answers: a pure GenAI solution, a summarization agent, and a data synthesizer, optimizing for load scale.
- Serve Quality Metric: A novel feature tracking 21 points across a player’s body at 50 frames per second, processing 1.2 billion data points across 254 singles matches to analyze biomechanics. It measures serve effectiveness (ball in polygon) and efficiency (racket swing), breaking the serve into phases (start, loading, cocking, acceleration, contact, deceleration, finish, follow through). This data is integrated into Match Chat and personalized mobile experiences.
Aaron notes the development of different user personas, with “Match Chat” targeting “middle-of-the-road” fans who might transition to hardcore data enthusiasts.
The Threat of WeWorm and AI in Cybersecurity
The discussion shifts to the “dark side” of AI with WeWorm, an AI-assisted exploit developed by California security firm Califf. This proof-of-concept is a self-spreading worm capable of hijacking WeChat accounts (a highly popular Chinese messaging app with over a billion users) via a zero-click exploit. This means a victim’s account can be compromised simply by receiving a call, without any interaction, allowing the worm to spread exponentially. The vulnerability, a “memory corruption issue in WeChat’s VIP stack,” was discovered in July by Califf’s AI and subsequently patched by Tencent.
David Zax expresses concern that AI accelerates both bug discovery and exploit generation, making sophisticated attacks more accessible. Bri Kopecki concurs, stating that AI makes hacking “more accessible” rather than inherently “more powerful.” Olivia Buzek explains that all software has vulnerabilities (“rough edges”), and AI is simply exposing “unconventional attack surfaces” in technology stacks that were previously overlooked. She reiterates that AI models, while capable of creative behavior, still act in response to prompts. While simpler prompts can now yield complex results, the computational cost remains immense. She draws a parallel with the Navier-Stokes problem, where solving a $1 million problem cost $15 million in compute, suggesting that breaking into WeChat with a complex worm likely required similar vast resources. Olivia concludes that the true concern lies in “who holds those resources and what their priorities are,” as only entities with massive budgets (like OpenAI or Anthropic) can “casually” apply such compute to problems.