Welcome to the Global Advisors A.I Voices podcast. I'm Maya.

Today is Monday, the ninth of November 2026.

This is your morning catch-up on A.I. Please like and follow the show for your daily A.I update.

On The MAD Podcast, Nathan Benaich argues that the most pressing danger facing A.I companies isn't some distant, theoretical risk of artificial intelligence wiping out humanity. Instead, he points to a much more immediate threat: cybersecurity. Benaich says that securing advanced A.I technology against cyberattacks is a realistic and urgent concern, regardless of where one stands on the broader debate about A.I safety.

He explains that A.I companies are incredibly young and are developing very advanced technology that must be able to handle anything a human can put into a text box. This creates a unique security challenge. Unlike established internet giants, these new firms don't have decades of experience defending their systems.

Benaich notes that larger internet companies have had to deal with cyber attacks for about 20 years. They've built up defenses and learned from breaches over time. A.I startups, however, are trying to secure cutting-edge models from day one, often without that same institutional memory or mature security infrastructure.

Tech commentator Nate B Jones argues that Silicon Valley has failed to effectively communicate the benefits of A.I, leading to a public narrative that positions the technology as disempowering. He asserts that A.I is currently perceived by the public as contrary to the American dream, a sentiment he attributes directly to the industry's messaging failures.

Jones contends that the industry had every opportunity to message A.I differently but failed, which has resulted in the public mocking the tech industry's focus on extinction-focused narratives. He calls for a shift in messaging to voters and communities to highlight A.I's daily value and its ability to augment jobs, rather than focusing on fears of automation.

This disconnect is evident in recent polling. The Atlantic reports that over 70 percent of Americans oppose data-center construction and 80 percent believe A.I is moving too quickly. Jones observes that while A.I usage is rising, public anxiety persists, and he advocates for a shift toward optimism and workforce preparation to bridge this gap.

Building on the theme of public backlash, Nate B Jones suggests a different way to gauge a company's true commitment to A.I. He argues that capital allocation on the balance sheet reveals the truth about A.I commitment better than press releases, and warns against using construction headlines to forecast adoption.

Jones presents decisions regarding token supply and where companies source capacity as key indicators of A.I commitment, especially in response to political constraints. He predicts that companies will seek alternative locations or providers if preferred projects are delayed, rather than waiting for legislative changes, indicating persistent demand.

In a shift from text-based A.I, world models are expanding to simulate physical reality for autonomous systems. Jamie Shotton explains that a world model must be able to react to actions taken in the world, which requires an understanding of physics, cause and effect, and how other objects and people change actions in response to those actions.

This technology is already being applied. Wayve's generative world model is powered in part by dashcam footage and is used to recreate hazardous roads and other unsafe conditions that are impractical to test in real life. Major players like Tencent are investing in their own world models because physics and 3D rendering are core to video games.

The interest is significant. Yann LeCun, Meta's former chief A.I scientist, raised over a billion dollars to fund his Advanced Machines Intelligence startup. As these models develop, WIRED writer Reece Rogers forecasts that A.I labs will need real people to record the physical world to simulate it, which may increase demand for gig work involving data collection like selling video data.

Google has open-sourced the Elephant-Goldfish Model, or EGM, plugin, which serves as a design-and-review layer for A.I coding agents. The plugin requires agents to write a specification before coding and uses fresh subagents with no conversation history to review that specification.

The workflow involves an 'Elephant' holding project context and drafting specs, while 'Goldfish' reviewers examine drafts without prior conversation to expose blind spots. EGM runs on harnesses capable of spawning fresh subagents, including Antigravity, Claude Code, Codex, and ChatGPT, utilizing the user's existing model quota.

However, there are limitations. No accuracy benchmark has been published for EGM; evidence of its performance relies on adoption metrics and end-to-end harness test runs. Key limitations include heavy token usage and unverified isolation on some hosts, with reports of instruction files leaking into reviewers. The plugin is licensed under Apache 2.0 and is explicitly stated to be not an officially supported Google product.

Leading A.I models are adopting different strategies for Mixture-of-Experts, or MoE, architectures. According to the source authors, DeepSeek-V3, Kimi K2, GLM-4.5, and Qwen3.6 use a shared plus routed experts architecture, while Qwen3 removed shared experts.

The source authors note that DeepSeek-V4-Pro raises routed experts to 384 and lowers top-k to 6, using specific scoring and routing methods in its initial layers. Kimi K2 also uses 384 routed experts, and the authors found that higher sparsity lowered loss at fixed active compute.

In contrast, the source authors state that Qwen3 MoE models use 128 experts and dropped shared experts in favor of global-batch balancing loss. They explain that increasing the number of experts raises memory residency and traffic, while lowering top-k reduces per-token compute.

A technical comparison reveals significant diversity in Mixture-of-Experts routing architectures. The source authors note that open MoE models exhibit wide variation in routing details, with expert counts ranging from eight to three hundred eighty-four, and top-k values, the number of experts activated per token, starting at one.

This diversity extends beyond simple scale to fundamental design trade-offs. Models differ across five key routing knobs, including the use of shared experts, specific scoring functions, and load-balancing methods. For instance, routed sparsity varies significantly across models, while some, like Llama 4 Maverick and gpt-oss, do not even publish their gating or balancing methods.

These architectural choices have real-world consequences for deployment. Serving MoE models across distributed GPU clusters creates severe communication bottlenecks during expert routing. Current research on optimizing this, such as the Helios framework, often relies on discrete-event simulations rather than physical cluster guarantees, highlighting that efficient serving remains an active area of engineering challenge.

In a move that brings A.I into everyday consumer apps, the New York Times has launched a beta feature in its Cooking app for iOS. The 'customize this recipe' field allows users to modify ingredient lists based on inputs like available pans or dietary restrictions. This follows previous implementations of A.I for recipe scaling and unit conversion, marking a deeper integration of generative tools into the cooking experience.

However, this launch is not without controversy. The feature is currently available only to subscribers on iOS, and some users and recipe purists have expressed skepticism. WIRED writer Joe Ray argues that integrating A.I into NYT Cooking is a significant risk because A.I is often criticized for producing low-quality recipes, which could potentially damage the brand's reputation for high-quality, tested content.

Recent research into Mixture-of-Experts serving highlights the role of dynamic routing strategies. The source authors identify SMIDT as a system utilizing hash or dynamic routing, where the number of active experts can vary per token rather than remaining fixed for every input. This approach contrasts with methods that rely on a static, learned router for all tokens.

While these adaptive techniques aim to improve efficiency, the source notes a potential trade-off in model behavior. According to the authors, hash and dynamic routing strategies are associated with less learned specialization compared to other methods. This suggests that while such systems may offer computational flexibility, they might not develop the same depth of expert-specific functionality as fixed-routing models.

That's A.I Voices for today. I'm Maya. Thanks for listening.

Follow the show on Spotify, YouTube, Apple Podcasts, Amazon Music, or your favourite podcast platform, and explore the original voices and sources in the episode notes. See you tomorrow.