OpenAI Deploys 10,000-Agent Swarm to Solve Navier-Stokes Math Prize
IBM
The breakthrough resolution of the Navier-Stokes problem proves multi-agent swarms can conquer grand-challenge mathematics, but the multimillion-dollar compute bill shows advanced AI capabilities remain governed by capital concentration rather than unguided model autonomy.
AI-assisted summary of OpenAI talks GPT-6 Astra and Millenium Prize, researchers create WeWorm exploit & IBM’s US Open app — original source. Check the original for full context.
Source published:
Key ideas
- OpenAI reportedly used its Astra model in a 10,000-agent swarm architecture to solve the Navier-Stokes existence and smoothness problem, consuming an estimated tens of millions of dollars in compute for an 88-hour run verified via Lean in 17 hours. (7:22)
- Cybersecurity startup Caliph utilized AI tooling to uncover an unconventional memory corruption vulnerability in WeChat's VoIP stack, demonstrating an automated, zero-click worm exploit dubbed 'WeWorm' prior to responsible disclosure and patching by Tencent. (23:45)
- IBM deployed an advanced tennis analytics stack for the US Open serving 14 million fans, combining XGBoost and logistic regression models for win forecasting with computer vision limb tracking that processes 1.2 billion skeletal data points across 254 singles matches. (20:00)
- OpenAI's Astra model achieved 95.9% accuracy on the BenchCAD benchmark, outperforming GPT-5.6 and Fable in generating runnable CAD code and spatial 3D models directly from 2D engineering views. (5:00)
Detailed notes
## Background Published by IBM on their RSS/podcast feed on September 11, 2026, this episode (#124) of Mixture of Experts features host Tim Hwang and co-host David Zax discussing fast-moving AI developments. According to the show description, the panel breaks down new model advances from OpenAI, emerging cybersecurity threats, and the latest digital AI tools deployed for US Open fans.
## TL;DR The panel examines OpenAI's flagship Astra model, dissecting its massive hardware footprint, BenchCAD evaluation, and multi-agent swarm architecture used to solve the Navier-Stokes Millennium Prize problem. Aaron Baughman demonstrates IBM's real-time AI capabilities deployed at the US Open, including pre-match prediction models, live momentum tracking, Match Chat, and computer-vision-based serve biomechanics tracking. The episode shifts to cybersecurity with an analysis of Caliph's AI-discovered 'WeWorm' zero-click vulnerability in WeChat's VoIP stack. Finally, Olivia Buzek and the panel weigh the economics of AI-driven vulnerability discovery, concluding that the primary risk lies not in autonomous software, but in which well-funded actors control the massive compute needed to run these exploits.
## Atmosphere & style The conversation balances conversational tech banter with deep, grounded technical scrutiny, bouncing seamlessly between enthusiastic demo showcases and existential cybersecurity dread. The hosts and panelists maintain a lively, collaborative rapport, poking fun at hyperbolic industry claims while treating underlying compute realities and mathematics with serious rigor. It feels like an insider roundtable of pragmatic practitioners who refuse to get swept up in marketing hype, guiding listeners through genuine scientific breakthroughs, impressive enterprise deployments, and chilling threat models with equal clarity.
## Bulletin - `07:22` OpenAI reportedly used its Astra model in a 10,000-agent swarm architecture to solve the Navier-Stokes existence and smoothness problem, consuming an estimated tens of millions of dollars in compute for an 88-hour run verified via Lean in 17 hours. - `23:45` Cybersecurity startup Caliph utilized AI tooling to uncover an unconventional memory corruption vulnerability in WeChat's VoIP stack, demonstrating an automated, zero-click worm exploit dubbed 'WeWorm' prior to responsible disclosure and patching by Tencent. - `20:00` IBM deployed an advanced tennis analytics stack for the US Open serving 14 million fans, combining XGBoost and logistic regression models for win forecasting with computer vision limb tracking that processes 1.2 billion skeletal data points across 254 singles matches. - `05:00` OpenAI's Astra model achieved 95.9% accuracy on the BenchCAD benchmark, outperforming GPT-5.6 and Fable in generating runnable CAD code and spatial 3D models directly from 2D engineering views.
## Key takeaways - `00:00` Flagship model demos that handle routine tasks like DoorDash orders raise questions about compute efficiency versus the necessity of throwing massive models at basic human interface challenges. - `00:00` True Artificial General Intelligence requires autonomous problem decomposition, tool use, and digital labor execution rather than isolated multimodal parlor tricks. - `00:00` Training massive frontier models like Astra demands staggering physical infrastructure, with rumors citing roughly 100,000 Blackwell NVLink 72 systems. - `05:00` Astra's 95.9% accuracy on BenchCAD reflects an industry shift toward instructing models to execute autonomous engineering and spatial design rather than simple query-answering. - `05:00` Enterprise adoption continues to prioritize latency, operational cost, and task-specific architectures over frontier scientific reasoning models of Astra's scale. - `07:22` The reported Navier-Stokes proof relied on an orchestrated multi-agent swarm of 10,000 AI agents sharing chain-of-thought pathways and leveraging Model Context Protocol (MCP) servers. - `10:00` Human agency was essential in solving Navier-Stokes, as human researchers strategically directed compute toward fluid dynamics after the model reached an intermediate milestone across the Millennium Prize problems. - `10:00` Formal verification of complex mathematical outputs using tools like the Lean theorem prover can drastically compress evaluation time relative to multi-day agent generation cycles. - `15:44` IBM's US Open digital platform integrates XGBoost and logistic regression with dynamic point-by-point momentum decay equations to generate live win probabilities. - `20:00` Computer vision pipelines track 21 skeletal joints at 50 frames per second to evaluate serve efficiency and mechanics across sequential preparation, strike, and follow-through phases. - `20:00` Conversational sports interfaces like Match Chat serve as intermediate engagement bridges, translating live context into graph-based agent queries tailored to diverse fan personas. - `23:45` The Caliph WeWorm exploit highlights that AI tools can automate the weaponization of zero-day vulnerabilities into zero-click propagation vectors within ubiquitous consumer apps. - `26:45` AI lowers the barrier of entry for discovering obscure memory corruption flaws across complex, unconventional attack surfaces like VoIP stacks. - `26:45` Automated zero-day exploit discovery remains constrained by the immense compute costs required to brute-force complex attack pathways, shifting the primary security threat to actors controlling massive financial and hardware capital.
## Detailed walkthrough ### `00:00` OpenAI Astra, Demos, and the True Meaning of AGI
Tim Hwang kicks off episode 124 of *Mixture of Experts* with co-host David Zax (staff writer for IBM Think) and panelists Olivia Buzek (staff AI engineer), Aaron Baughman (IBM Fellow), and Bri Kopecki (AI Customer Success Engineer). The episode lines up three central discussions: IBM's tech deployment at the US Open, emerging cybersecurity alarms surrounding WeWorm, and major announcements coming out of OpenAI concerning their new flagship model, Astra. Opening the Astra segment, David describes the flashy product demos that have become standard industry fare—where presenters step into a specialized room to converse naturally with an AI, showcase real-time conversions of 2D renderings into 3D objects, and order food deliveries on the fly. David questions whether this breakthrough matches the claims of NVIDIA's CEO that society has finally reached Artificial General Intelligence (AGI).
Olivia Buzek immediately raises skepticism regarding the immense compute footprint powering such feats. She points out that using huge, token-hungry models simply to manipulate a desktop interface or order DoorDash is fundamentally inefficient, as lightweight harnesses around smaller models could achieve the same utility. Olivia stresses that the field needs to reconsider whether it is tackling foundational challenges correctly:
> “And so I think what worries me the most is like, are we approaching problems the right way as humans, or are we simply throwing bigger and bigger models at problems without thinking through what the actual implications are.”
Aaron Baughman pushes back on premature declarations of AGI, explaining that the community still lacks a unified definition. To him, an AGI would need to autonomously grasp an unseen problem, acquire missing knowledge, devise an execution plan, use software tools, and perform labor like a high-functioning digital employee. More critically, he notes the staggering physical cost of training Astra, reporting rumors that it required roughly 100,000 Blackwell NVLink 72 systems—an infrastructure demand equivalent to the electrical power consumption of an entire continental U.S. region.
- Panelists: Tim Hwang (host), David Zax (co-host), Olivia Buzek (staff AI engineer), Aaron Baughman (IBM Fellow), Bri Kopecki (AI Customer Success Engineer). - Reported compute footprint for Astra: approximately 100,000 Blackwell NVLink 72 systems.
`05:00` BenchCAD Capabilities vs. Enterprise Utility
Aaron highlights Astra’s actual performance on specialized benchmarks, singling out the BenchCAD benchmark. Astra attained a 95.9% accuracy score, while competing models like GPT-5.6 and Fable languished in the low 80s. Rather than just asking text-based queries, the model takes multiple 2D engineering views, interprets the spatial environment, generates runnable CAD code, and fabricates 3D models compared directly against true geometric specifications. Aaron observes this marks a broader industry pivot from merely asking an AI for answers to instructing an AI to conduct autonomous engineering work.
Bri Kopecki reframes the breakthrough through an enterprise lens, noting that corporate clients rarely need a model of Astra’s colossal size or raw scientific reasoning for day-to-day business operations. While she acknowledges the value of pushing technological boundaries and admires NVIDIA's hardware innovations, real-world deployments prioritize speed, cost, and task-appropriate architectures over sheer scale:
> “I haven't yet to see a use case or as of today or as of this, this or last week, where my customers would have needed a model of this magnitude.”
- Astra BenchCAD benchmark score: 95.9% accuracy. - Comparison benchmarks: GPT-5.6 and Fable both scored in the lower 80% range.
`07:22` The Navier-Stokes Millennium Prize and Swarm Architecture
The conversation shifts to Astra’s reported breakthrough in pure mathematics: solving the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems designated by the Clay Mathematics Institute in 2000 with a $1 million prize. David and Aaron explain the physics behind the problem, which has resisted mathematical proof for nearly 90 years. It evaluates whether smooth initial conditions in fluid dynamics can mathematically break down into infinite velocities—a phenomenon David colorfully nicknames the “exploding water problem”:
> “I call it the exploding water problem because apparently they've proven that water can spontaneously explode, which sounds awesome.”
Aaron explains that the true technical marvel was not single-model deduction, but OpenAI's multi-agent swarm architecture. OpenAI unleashed a coordinated network of roughly 10,000 AI agents. Governed by central coordinators, these sub-agents dynamically acquired external tools through MCP (Model Context Protocol) servers, shared chain-of-thought pathways, and systematically divided the problem space to execute wide-ranging mathematical explorations. David adds an element of drama to the announcement, noting the behind-the-scenes "soap opera": rumors circulated that a team of human researchers at NYU and Anthropic were on the verge of publishing their own Navier-Stokes proof, prompting OpenAI to rapidly spend tens of millions of dollars in compute to solve the problem first and claim the $1 million prize.
```mermaid flowchart TD accTitle: Astra Swarm Agent Architecture accDescr: Orchestration workflow showing Astra directing 10,000 agents with MCP tools and chain-of-thought reasoning to resolve the Navier-Stokes equations.
subgraph Orchestration ["Orchestration Layer"] A["Astra Swarm Coordinator"] end
subgraph AgentSwarm ["Swarm Exploration (10,000 AI Agents)"] B1["Sub-Agent Pool 1"] B2["Sub-Agent Pool 2"] B3["Sub-Agent Pool N"] end
subgraph Execution ["Tool Integration & Reasoning"] C["MCP Server Tools"] D["Chain-of-Thought Workflows"] end
subgraph Output ["Target Discovery"] E["Navier-Stokes Singularity / Proof"] end
A --> B1 A --> B2 A --> B3 B1 --> C B2 --> D B3 --> C C --> E D --> E ```
- Background: Clay Mathematics Institute established seven Millennium Prize Problems in 2000, offering $1 million per proof; Navier-Stokes is the second problem claimed to be solved. - Implementation: OpenAI deployed an orchestration of 10,000 AI agents using MCP server tool retrieval and chain-of-thought exploration. - Cost disparity: An estimated tens of millions of dollars of compute time was expended to win a $1 million cash prize.
`10:00` Human Agency in AI Breakthroughs and Real-World Applications
The discussion opens with banter around the real-world implications of solving fluid dynamics equations like Navier-Stokes, joking about whether it is even safe to take a shower now that fluid behavior can be modeled mathematically. David Zax notes that having a proof generated through AI gives mathematicians something concrete to examine, though Aaron Baughman advises caution before declaring it absolute ground truth: “Let's just wait and see what that ground truth actually is.” However, Aaron notes the litmus test looks very promising, particularly because competing mathematicians are already disputing priority and claiming they had reached the solution first—a reliable signal that the AI's result is close to the real answer.
David highlights that the core of the Navier-Stokes achievement is human orchestration rather than an unguided AI breakthrough. When OpenAI’s Astra was running on all seven Millennium Prize Problems, humans guided the process: after Astra solved an intermediate milestone, researchers made the strategic choice to steer massive compute directly into Navier-Stokes. Addressing Bri Kopecki, David asks how she views this human-in-the-loop collaboration. Bri agrees emphatically, explaining that human intellect, taste, and creativity remain essential for setting priorities and directing compute resources:
> “Yeah, I think that I love how you're framing this because it shows to the world that regardless of how powerful AI has become and is becoming, it still requires human intellect, human creativity and human collaboration to get those answers to these huge problems.”
Bri reassures listeners worried about immediate job displacement that human brains are still fundamentally relevant. She expresses immense optimism for downstream applications, arguing that if 10,000 agents running for 88 hours can tackle a Millennium problem—verified via the Lean theorem prover in just 17 hours—similar techniques will revolutionize health and climate science. She cites podcast discussions predicting that AI and quantum physics will eventually cure all cancers and eliminate fatal diseases, noting that this scale of automated reasoning brings humanity vastly closer to solving major global crises.
- Astra initially churned across all seven Millennium Prize Problems before human researchers pivoted compute specifically toward Navier-Stokes after an intermediate breakthrough. - The initial model solution required 88 hours of agent runtime. - Formal verification of the proof into Lean took only 17 hours to evaluate whether the 10,000 agents' output was mathematically sound.
---
`13:46` IBM and the US Open: Bringing AI to 14 Million Tennis Fans
David pivots the conversation from curing cancer and solving climate change to Aaron Baughman’s primary beat: sports analytics. Aaron introduces IBM’s long-standing partnership with the United States Tennis Association (USTA), which spans over 30 years since 1992, designed to transform the digital fan experience for the US Open. While only a small fraction of the tournament's 14 million global fans can attend the matches in person, IBM uses AI to bring match insights and real-time storytelling directly to viewers on digital platforms.
Aaron emphasizes that this year the team elevated the platform by combining predictive machine learning, real-time momentum tracking, and generative agent experiences. They even tackled grand-challenge biomechanics: tracking athlete limbs and joint movements to evaluate player mechanics, especially on serves.
```mermaid flowchart TD accTitle: IBM US Open AI Match Analytics Workflow accDescr: Pipeline showing pre-match forecasting, real-time live match processing, and agentic fan engagement tools.
subgraph PreMatch["Pre-Match Analytics"] A["Media Volume & Sentiment"] --> C["Pre-Match Likelihood to Win"] B["Head-to-Head & Surface Stats"] --> C C -->|"XGBoost + Logistic Regression"| D["Win Odds Forecast"] end
subgraph LiveMatch["Real-Time Live Engine"] E["Point-by-Point Data Feed"] --> F["Dynamic Momentum Equation"] D --> F F --> G["Live Likelihood to Win"] G -->|"Confluent Messaging"| H["Agentic Marketplace"] H --> I["Key Moments Text Generation"] end
subgraph FanExperience["Interactive Layer"] I --> J["Match Chat & 3D Replay"] K["Graph-Based AI Agents"] --> J end ```
---
`15:44` Live Demos: Win Likelihood, Real-Time Momentum, and Match Chat
Sharing his screen during the live US Open tournament, Aaron walks through the digital tools on USOpen.org, focusing on Jessica Pegula’s recent quarterfinal victory over Emma Navarro. Starting with the pre-match preview, Aaron reveals the mechanics behind IBM’s Likelihood to Win:
> “So what happens is we have the system that ranks all of the players. We have pre-match forecasters that attempt to say what the volume of the media is being written about them, what the sentiment of different factors are. And then we also look at their head-to-head stats, if they have any.”
The system processes player surface records (hard court, clay, grass) through two classical machine learning models—XGBoost trees and logistic regression. In the Pegula-Navarro match, the pre-match model assigned Pegula a 77% likelihood of winning, which proved accurate. Aaron notes IBM’s pre-match prediction accuracy consistently hovers in the 60% range across the tournament.
Aaron then demonstrates the in-match capabilities: - **Live Likelihood to Win**: The model is bootstrapped from the pre-match win likelihood and dynamically shifts on a point-by-point basis using an internal momentum equation with situation-specific boosters and decay rates. - **Key Moments Generation**: As every point is won or lost, data streams across Confluent to an agentic marketplace. The system detects turning points and generates automated text recaps (e.g., summarizing Pegula closing out a 2-1 set victory over Navarro) to act as a second-screen companion for fans. - **Match Chat**: Aaron accepts the terms and conditions on screen to demo Match Chat, a conversational tool allowing fans to interrogate match data. The architecture translates fan queries into prompts grounded in live tournament context, dispatching them across racing graph-based agents to fetch and synthesize stats in real time.
`20:00` Limb Tracking, Match Chat Personas, and IBM's Serve Quality
The conversation opens with Aaron Baughman explaining how IBM's tennis analytics platform handles scale and tailors itself to different kinds of sports fans. David Zax asks whether IBM gathers data on who is actually using these digital experiences, noting that his own family embodies two extremes: one fan who just likes to tune in casually and watch the ball go back and forth, and another who is intensely data-driven. Aaron confirms that they track analytics and build for distinct user personas:
> "What, you know, what I showed you like Match Chat is, is for, you know, sort of the middle of the road. It's not necessarily targeted towards that hardcore fan or the novice, but it's someone sort of in the middle, right?"
Aaron explains that Match Chat serves as a bridge for someone who wants data to transition into a hardcore fan, while also remaining accessible to busy users looking for quick surface-level insights. Because users sign in and favorite specific players, the team can correlate these user profiles back to the types of queries asked within Match Chat to map out user behavior.
Turning to IBM's technical capabilities, Aaron showcases an advanced computer vision feature: limb tracking to assess serve quality, built using IBM's watsonx platform. The system tracks 21 distinct skeletal points across a tennis player's body at 50 frames per second. Over the course of the tournament, this pipeline processes 1.2 billion data points across all 254 singles matches. It evaluates two primary metrics: - **Effectiveness:** Whether the ball lands inside the targeted court polygon. - **Efficiency:** The mechanics of the racket swing and body motion.
```mermaid flowchart LR accTitle: Serve Phase Decomposition Pipeline accDescr: Pipeline decomposing a tennis serve into mechanical stages to evaluate quality and efficiency. subgraph Prep ["Preparation"] A["Start Phase"] --> B["Loading Phase"] B --> C["Cocking Phase"] end subgraph Strike ["Strike Execution"] C --> D["Acceleration"] D --> E["Ball Contact"] end subgraph FollowThrough ["Completion"] E --> F["Deceleration"] F --> G["Finish & Follow-Through"] end ```
The system breaks every serve down into sequential phases: start, loading, cocking (raising the racket into position), acceleration downward toward the ball, contact, deceleration, and the finish/follow-through. Rules evaluate the mechanics across these phases. Fans can ask questions about serve quality directly in Match Chat, and mobile users receive a personalized live home-screen card highlighting serve quality insights. Aaron calls the tool a first of its kind and expresses excitement about expanding limb tracking to other shot types. Tim Hwang praises the feature, noting that every time Aaron joins *Mixture of Experts*, the underlying platform has visibly advanced.
`23:45` The WeWorm Zero-Click Exploit and WeChat Security
Shifting from sports and scientific breakthroughs to cybersecurity, Tim introduces what he calls "the dark side" of recent AI developments: a story about researchers using AI to discover vulnerabilities and engineer exploits, specifically a menace dubbed "WeWorm." David Zax shares his genuine anxiety about the discovery, joking that he used to worry about the monster in the shower as a kid, but is now terrified of his phone.
A cybersecurity startup named Caliph used AI tooling to locate a zero-day vulnerability in WeChat—the ubiquitous Chinese messaging, payment, and lifestyle app used by over a billion people. Caliph's AI did not simply discover a bug; they demonstrated that it could be weaponized into an automated, zero-click worm. A malicious VoIP call could compromise a device without requiring the victim to answer or interact. Once compromised, the malware could propagate exponentially to the user's contacts. Caliph adhered to responsible disclosure practices by alerting WeChat's parent company, Tencent, back in July. Tencent patched the flaw before the news was made public months later.
David admits that as a non-technical observer, the incident sounds terrifying:
> "I am personally I'm not a technical person but I'm terrified because this would seem like an exploit on the level of what we are more familiar from really sophisticated state actors."
David asks where everyone's "P-Doom" (probability of doom) should be sitting given how quickly AI appears to be accelerating both bug discovery and automated exploit generation. Bri Kopecki validates the concern, recalling her time living in Shanghai between 2015 and 2016 when she used WeChat to pay her rent, highlighting how deeply embedded the platform is in daily life. However, Bri cautions against outright panic:
> "I think the key takeaway here is that AI... Did it make the hacking more powerful? I mean, a hacking case is always significant event, it made it more accessible."
She emphasizes that high-profile vulnerabilities surface monthly, and while AI lowers the barrier to entry for uncovering them, the tech industry must rely on security researchers to develop defensive countermeasures.
`26:45` Unconventional Attack Surfaces and the Economics of AI Exploit Discovery
David turns to Olivia Buzek to unpack the technical details behind Caliph's disclosures. Caliph reported that the vulnerability stemmed from a memory corruption bug inside WeChat's VoIP stack, warning that this represented an "unconventional attack surface" common across modern messaging platforms. David asks why AI seems to be suddenly surfacing vulnerabilities in obscure, overlooked corners of legacy software stacks.
Olivia offers both a philosophical and technical explanation. Software engineering, like all human craftsmanship, is fundamentally imperfect:
> "There have always been rough edges in everything that humans have ever created. It's actually one of the core beauties of being human is that the things that we create are not generally perfect objects."
Drawing a comparison to *wabi-sabi*—the Japanese art form of repairing broken pottery to embrace flaws—she points out that building entirely vulnerability-free software is impossible. Historically, finding obscure memory corruption flaws required painstaking human reverse-engineering. Today, while language and reasoning models do not possess independent agency or act unprompted, their ability to execute creative reasoning against broad, open-ended prompts (such as "go off and look for vulnerabilities") has expanded dramatically.
However, Olivia stresses that this automated capability is heavily constrained by economics and compute resources: - Caliph's disclosure did not reveal the token or dollar cost required to discover and construct the WeWorm exploit. - By comparison, solving the Navier-Stokes existence and smoothness problem with OpenAI's Astra model reportedly consumed approximately $15 million in compute to win a $1 million Millennium Prize reward. - Finding complex zero-click exploits via AI brute-force similarly requires massive, multi-million-dollar compute budgets.
Olivia argues that fear should not be directed at the models themselves, but at who commands the capital:
> "So personally, my fear is a little bit proportional to who holds those resources and what their priorities are."
Tim concludes that while lowering technical barriers is concerning, massive capital was already an asymmetrical weapon in cybersecurity: previously, an adversary with $15 million could simply hire elite black-hat hackers to find the same zero-day flaws. Tim wraps up the episode by thanking Bri Kopecki, Aaron Baughman, and Olivia Buzek for providing a blend of hope, grounded technical nuance, and existential terror, before signing off until next week's episode.
## Notable quotes > `00:00` And so I think what worries me the most is like, are we approaching problems the right way as humans, or are we simply throwing bigger and bigger models at problems without thinking through what the actual implications are. — Olivia Buzek
> `26:45` There have always been rough edges in everything that humans have ever created. It's actually one of the core beauties of being human is that the things that we create are not generally perfect objects. — Olivia Buzek
## Mentioned OpenAI, GPT-6 Astra, Millennium Prize Problems, Clay Mathematics Institute, Navier-Stokes equations, WeWorm, IBM, US Open, United States Tennis Association, Tim Hwang, David Zax, Olivia Buzek, Aaron Baughman, Bri Kopecki, NVIDIA, Blackwell NVLink 72, BenchCAD, GPT-5.6, Fable, Model Context Protocol, Anthropic, New York University, Lean, Jessica Pegula, Emma Navarro, XGBoost, Confluent, Match Chat, watsonx, Caliph, WeChat, Tencent, DoorDash
## Video description Visit Mixture of Experts podcast page to get more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts It's been a week that proves AI is moving faster than anyone can fully keep up with, unless you’ve got a good backhand. On episode 124 of Mixture of Experts, host Tim Hwang and co-host David Zax are joined by Bri Kopecki, Aaron Baughman, and Olivia Buzek to talk about new advances from OpenAI, new cybersecurity threats, and new tools for US Open fans from IBM. First, OpenAI...
[Source](https://cdnapisec.kaltura.com/p/1773841/sp/177384100/playManifest/entryId/1_tvxm15e9/format/download/protocol/https/flavorParamIds/0/name/media.mp3)