Podcast··43m

Episode 13 — Outnumbered by Robots

Fable gets banned worldwide by US export controls, Grok goes to the battlefield, GLM 5.2 matches Fable at 15x less cost, and Figure AI's robots now outnumber their human workers.

Listen onYouTube

Episode notes

The US government banned a software model for the first time, Grok shipped a military AI the next day, Chinese lab GLM 5.2 matched Fable's intelligence at a fraction of the cost with open weights, and Figure AI's robot fleet overtook its human workforce.

Chapters

  • 05:00 — Fable banned worldwide by US export controls, Grok militarized overnight
  • 13:00 — Sakana's Fugu multi-agent router, Apple's AI stagnation, and the usage reality check
  • 19:00 — GLM 5.2: Fable-level intelligence at 15x less cost, fully open weights
  • 28:00 — Figure AI robots outnumber humans, Mondo's filming robot, and the headphone chip exploit

Episode 13 — Outnumbered by Robots

Description: Fable gets banned worldwide by US export controls, Grok goes to the battlefield, GLM 5.2 matches Fable at 15x less cost, and Figure AI's robots now outnumber their human workers.

The first software export ban in US history landed this week, and it landed on an AI model. One day later, Grok announced a military-grade model already directing targeting decisions on the battlefield. Meanwhile, a Chinese lab released an open-weight model that matches Fable's intelligence at 15 times lower cost, and a robotics company crossed a symbolic threshold where robots outnumber humans on its own factory floor.

Fable Banned Worldwide: The First Software Export Control

The Bureau of Industry and Security, the US agency that normally handles hardware export controls, banned Fable for the first time as software rather than hardware. Because Anthropic cannot reliably verify who is a US citizen versus a non-citizen at the traffic level, the restriction applies to everyone worldwide, including apparently Anthropic's own non-US staff members.

  • Internal contradiction: Anthropic almost certainly uses Mephos internally for development and possibly training. Roughly half the team codes with Fable 5, the other half with Opus 4.8. Banning non-US employees from their own model is a self-imposed wound the company cannot realistically enforce.
  • The jailbreak trigger: The restriction was not driven by commercial concerns but by a security finding. Internal US government research determined that Fable could be jailbroken too easily, similar to the Pliny exploit. The guardrailed model itself was not the problem; the ease of bypassing those guardrails was deemed an unacceptable security risk.
  • The Europe question: The ban underscores how dependent Europe remains on US intelligence infrastructure. The GPS parallel is instructive: three separate satellite networks now exist (American, Chinese, European) precisely because relying on a single military-controlled network was untenable. AI is heading the same direction, but Europe is hella far behind, and Mistral alone cannot fill the gap.

Grok Goes to War: Military AI the Day After the Ban

One day after Fable was restricted, xAI announced a Grok model adapted for military use. It is already confirmed to be directing targeting decisions in live active military operations.

  • Speed of deployment: Going from "Mega Hitler" persona six months ago to battlefield targeting decisions is a velocity that raises its own questions about guardrail maturity.
  • The ethics asymmetry: Anthropic faces intense scrutiny over ethics and safety, while Grok ships directly to the battlefield with comparatively little public debate. The split in who gets access to intelligence and who does not is now visible not just across borders but across use cases.
  • Token economics: A model that reasons through long thinking loops and rereads its own output to arrive at better answers is fundamentally expensive in tokens, regardless of per-token pricing.

Sakana's Fugu: A Router Disguised as a Model

A small Japanese lab, Sakana AI, released a model called Fugu that it advertises as Mephos-class intelligence. The architecture is not a single model but a multi-agent routing system.

  • How it works: Fugu orchestrates across multiple closed and open models, routing queries to whichever model is best suited. Sakana built the router and orchestration layer on top of its own models, creating intelligence through diversity of training data and perspective rather than raw scale.
  • The swarm approach: This mirrors what OpenRouter does at the platform level, but packaged as a single model endpoint. Multiple models with different training data and different views onto the problem collectively produce better outputs than any single model alone.
  • Cost question: Token usage across a multi-model router is necessarily high. The intelligence-per-dollar math depends heavily on how efficiently the router can minimize unnecessary calls.

Apple's AI: Still Missing in Action

Apple announced its latest AI features at WWDC, shipping in September. The headline Siri revamp demonstrated by asking "who's playing tonight at the World Cup?" and pulling up the game.

  • Wrong audience: A World Cup score lookup as the flagship demo signals Apple is targeting the average iPhone user, not developers or power users. The team's consensus: the Siri revamp looks useless for anyone who already uses Gemini or ChatGPT daily.
  • Gemini under the hood: Apple's new Siri is effectively Gemini-powered. The visual intelligence and interface layers are where Apple differentiates, but the intelligence itself comes from Google.
  • The bubble reality: A widely circulated chart shows each dot representing 3.2 million people who never use AI. Free chatbot users, paying users, and power users form sharply distinct tiers. The AI market is still far from saturation, which means Apple's mass-market strategy is not wrong, it is just not aimed at us.

GLM 5.2: Fable Intelligence at 15x Less, Fully Open

Chinese lab Zhipu released GLM 5.2, and if you trust the benchmarks, it sits at Fable-level intelligence at roughly 15 times lower cost. It is a one-trillion-parameter model, and everything is published: weights, architecture, and already a Q2 quant.

  • Open weights: The full model and a Q2 quantization are available. The quant brings VRAM requirements down from 1.56 terabytes to a "reasonable" 260 gigabytes, still massive but tractable on high-end hardware.
  • Hardware reality: Running the full model requires 1.56 TB of VRAM. The Tiny Box from Tiny Corp, which can run GLM 5.2, costs $20 million. A smaller version with six RTX 6000s (96 GB each) runs $50,000 to $80,000. This is not a hobbyist setup.
  • Reasoning over scale: GLM 5.2 achieves its intelligence through extended reasoning loops, rereading and iterating within the model in a pattern reminiscent of Karpathy's auto-research loop. It thinks harder rather than being trained larger, which is a fundamentally different cost structure.
  • Daily driver: For practical use, GLM 5.2 is already a daily-work tool at Fable-grade quality, and the open weights mean the local-to-cloud ladder works within a single model family.

Figure AI: Robots Now Outnumber Humans

Brett Adcock, founder and CEO of Figure AI, shared that robots have outnumbered humans at the company. The numbers: 660 employees and roughly 740 robots manufactured, building approximately one robot per hour.

  • Not all active: The 740 figure includes robots in production, in testing, and ready for customer delivery, not 740 walking around doing dishes. But the manufacturing rate of one robot per hour is a real production milestone.
  • BMW testing: Figure robots have been testing at BMW for roughly a year and a half, well past the pilot stage. Demand is picking up, and the production ramp reflects real orders, not just warehoused inventory.
  • The singularity question: Whether robots are now building robots at Figure remains unclear, but the crossover point where a robotics company's output exceeds its headcount is a meaningful inflection.

Mondo Robotics: A $499 Filming Robot That Jumps

Mondo Robotics shipped a ground-based filming robot that hits 25 km/h, jumps, flips, self-rights after falling, and handles all terrains including grass and mud. Price: $499. Processing happens on-device.

  • The use case: It follows you via a wristband tracker, giving a ground-level tracking shot that drones cannot replicate. Drones win on altitude flexibility; this wins on proximity and terrain interaction.
  • Design lineage: It looks like a DJI product, and likely shares components. The form factor and dual-camera setup mirror DJI's industrial design language.
  • The verdict: Not a tool for everyone, but for skating, cycling, and action sports footage, it fills a gap between aerial drones and static tripods. As a modern RC car replacement for kids, it is a straight upgrade.

Black Market Subscriptions and the Bluetooth Headphone Exploit

Two security stories rounded out the episode. First, marketplaces are selling Claude Code subscriptions for $6, likely sourced from stolen accounts rather than regional pricing arbitrage. Second, two researchers from Heidelberg discovered an exploit in a Bluetooth chip from a Taiwanese manufacturer used in over 100 headphone models.

  • The chip exploit: The chip had an open maintenance port that should have been closed. Within 10-meter Bluetooth range, an attacker could extract the pairing key, connect to the headphones, and send commands to the paired phone, including accessing the phone's memory.
  • Scope: Over 100 headphone models were affected, including Beats Studio Pro and Nothing earbuds. Apple's own AirPods use a proprietary chip and were not affected.
  • Patch timeline: Apple took roughly a month and a half to fix the issue, significantly longer than competitors. The vulnerability was not a zero-day in the traditional sense, just a manufacturer forgetting to close a maintenance port.

Conclusion

The signal this week is that AI is now weaponized, regulated, commoditized, and physical all at once. A US agency treating a language model like a munition export is a precedent that will shape policy for years. Open-weight models matching frontier intelligence at a fraction of the cost will pressure every closed lab's pricing. And when a robotics company builds one robot per hour and its fleet outnumbers its staff, the question stops being whether robots will be everywhere and starts being when the rest of us notice.