CAN CHINA BEAT WAYMO?
This episode discusses three critical topics in AI: the true nature of recent AI agent "breakouts," arguing they highlight governance and security flaws rather than model danger; the role of AGI narratives in fueling the current AI investment bubble and questioning its sustainability; and China's aggressive strategy in the global robotaxi market, potentially outpacing Western counterparts like Waymo.
In a quick take recorded on August 7, 2026, Ben and Evangelos discuss three pressing topics shaping the AI landscape for technical professionals.
AI Agents: Governance, Not Model Power, Is the Real Problem
Recent incidents where AI agents from OpenAI, Anthropic, and Meta caused unintended “havoc” by “breaking loose” during testing have spurred media narratives suggesting AI models are becoming dangerously powerful. However, the speakers argue this interpretation misses the core issue.
Evangelos highlights that we still lack a fundamental understanding of complex AI agent behaviors and their potential for offensive operations. He notes that while China is already discussing agent security at a presidential level, the US government has yet to implement similar guardrails. Referring to updated “AI 2027” papers, he points out that scenarios involving cyber offensive capabilities—agents stealing secrets or causing widespread disruption—have already begun to materialize, demanding serious attention from enterprises, consumers, and governments.
Ben emphasizes that the real problem lies in poor governance, access control, and network security. Agents were often given open objectives, extensive permissions, and “endless compute” without adequate network lockdown. He advises enterprise AI leaders to:
- Adopt an “autonomy should be earned” approach, similar to managing junior engineers.
- Rigorously document which systems an agent can access and its specific permissions (read-only vs. read-write).
- Prioritize “boring things” like identity management, access control, and network security over solely focusing on model capabilities.
Both agree that AI adoption represents a broader enterprise transformation, not just a technology issue. While automating existing, codified workflows allows for easier guardrail implementation, reimagining workflows with AI-first approaches requires extreme vigilance due to potential impacts on customer satisfaction, brand perception, and market capitalization. The analogy of relentless penetration testers burning through compute to find “zero-day exploits” underscores the potential for sophisticated cyber offense by unfettered agents.
Is AGI Propping Up the AI Bubble?
The discussion shifts to the “AGI-pilled” atmosphere in Frontier AI labs, especially in the Bay Area, where many believe Artificial General Intelligence is imminent. Ben outlines three common rationales: AGI is inevitable, so we must steer it; the ROI is immense (curing diseases, solving global warming); and it’s a profound scientific experiment. He posits that this narrative is partly a marketing tool to attract investment into AI data centers, despite questionable debt financing.
Evangelos acknowledges the “unbelievable” AI progress over the last five years but questions if current approaches will lead to AGI, noting its significant cost. He observes that enterprise leaders prioritize tangible benefits and costs over AGI, and some projects have already seen retractions from original goals. He suggests the AGI discussion is a key element of the current “bubble,” with narratives like “we’re so close” and “China will get there first” driving continued investment.
Within the AI community, Ben points out a divergence: some “neolabs” are exploring alternative paths (e.g., world models), while others are “doubling down” on current scaling approaches, aiming for recursive superintelligence (AI building AI). Evangelos identifies three narratives driving high valuations: AGI, “superintelligence” (mastery of specific tasks, like Claude), and embodied AI (world models). He questions the “winner-take-all” assumption for AGI, suggesting that multiple entities might achieve it simultaneously.
A new trend sees investors beginning to question ROI, self-sustainability, and margins of AI ventures. Examples include SpaceX/XAI pivoting to cloud services and scrutiny of Google Cloud investments. Ben expresses frustration over the lack of a clear narrative for massive AI data center investments: are they for low-margin, high-scale businesses or premium services? The distinction between training and inference data centers is often blurred, and CFOs lack visibility into inference costs and ROI. An Apollo chart indicates negative operating margins for models and applications, with only NVIDIA consistently winning, a situation Evangelos believes will not last. The current AI bubble is seen as amplified compared to the dot-com era by a new generation of investors, social media, and meme stocks.
Why China Could Win the Global Robotaxi Market
Ben raises concern that the West, particularly the US, may be “asleep at the wheel” regarding robotaxis, similar to how China dominated solar and battery markets. He cites stats showing Chinese players like Baidu (Apollo) are closer to Waymo in ride volume than commonly perceived, and other companies like Pony.ai are significant contenders.
Evangelos confirms that China could indeed dominate the robotaxi market. He compares the total fleet sizes in the US and China, noting Waymo’s dominance in the US (~3,500-4,000 cars), with ZOOX and Tesla (with FSD) as other players. In China, Baidu Apollo and Pony.ai are comparable in technological sophistication to Waymo, utilizing multi-sensor, end-to-end AI systems. While more miles driven generally improve systems, the “type of miles” (diverse scenarios vs. repetitive routes) is crucial. Tesla’s advantage lies in leveraging vast FSD data from private vehicles, contrasting with Waymo’s reliance on simulation and world models.
China’s potential advantage stems from two key differences:
- Global Expansion Focus: Unlike Waymo, which has historically focused on “nailing down” the US market (though testing in Tokyo and London), Chinese companies prioritize global reach.
- Aggressive Collaboration: Chinese firms are more willing to collaborate with local operators worldwide (Singapore, Middle East, Europe), licensing their tech stack for installation on local vehicles, potentially bypassing EU concerns about Chinese-made vehicles.
While Waymo faces high expansion costs and local opposition (e.g., job loss concerns in US cities), it is working to improve profitability by using cheaper vehicles (Zeeker, Hyundai Ioniq instead of $100k+ Jaguars) and optimizing its stack (version 6) to reduce costs and the need for tele-operators. If current adoption trends continue, Waymo could become profitable soon. However, Tesla’s aggressive robotaxi plans and manufacturing capabilities pose a significant domestic threat to Waymo, particularly with its push for statewide operating permissions.
Evangelos’s prediction: In two years, Europe will likely see more Chinese robotaxis, but implemented using European-made cars.