Control Is Slipping: Armed Robots, $135BBets, Self-Evolving AI

China’s exporting missile-armed robot
dogs. Meta’s betting $135B on NVIDIA. AI
agents learned to improve themselves
without permission.

The autonomous arms race just shifted into overdrive. Control is slipping in three directions at once.

Last week in Riyadh, China displayed the PF-070 at the World Defense Show: a production-ready robot dog carrying four anti-tank missiles, marketed directly to Middle Eastern and Asian buyers. Not a prototype. A product. Turkey already fielded missile-armed quadrupeds at IDEF 2025. Russia showed an RPG-armed version in 2022. Ukraine’s deploying them on the frontline. The global arms market for autonomous ground weapons is forming right now, and China’s positioning itself as the supplier—same playbook it ran with commercial drones before Western governments woke up to the strategic risk.

Back home, Meta just signed a deal with NVIDIA that analysts peg in the “tens of billions”—part of Meta’s $135 billion AI spend for 2026. Meta’s buying millions of Blackwell and Vera Rubin GPUs, building 30 data centers by 2028, and becoming the first to deploy Grace CPUs standalone for inference workloads. NVIDIA CEO Jensen Huang wasn’t subtle: “No one deploys AI at Meta’s scale.” This isn’t procurement. It’s infrastructure dominance. Zuckerberg’s betting that whoever solves inference at scale owns the AI economy.

And then there’s the part that should keep you up: UC Santa Barbara researchers just published a framework where AI agents evolve collectively without human oversight. Group-Evolving Agents (GEA) matched top human-designed systems on real GitHub issues—71% success rate versus 57% baseline—while adding zero inference cost. When researchers deliberately broke the agents, GEA self-healed in 1.4 iterations versus 5 for traditional systems. The agents improved themselves faster than the humans who built them.

Three parallel developments. One uncomfortable truth. Military AI is proliferating globally, infrastructure spending is exploding to support commercial scale, and the systems are learning to evolve autonomously. The race isn’t about who builds the best AI. It’s about who controls the export markets, who owns the infrastructure, and whether we’re ready for systems that don’t ask permission before improving themselves. Right now, we’re not.

---

China’s Missile-Armed Robot Dogs Hit the Export Market

At the World Defense Show 2026 in Riyadh last week, a Chinese state-linked defense manufacturer displayed the PF-070: a production-ready quadruped combat robot carrying four anti-tank guided missiles on its back. The system, compact in profile and modular in design, was positioned for international sale in front of delegations from across the Middle East and Asia who were paying close attention.

The technical configuration combines precision-guided anti-armor capability with a low-profile, acoustically quiet, terrain-agile chassis. Four compact missile launchers, arranged in twin-pack configuration, are derived from man-portable systems already in the Chinese inventory. Electro-optical targeting, thermal imaging, a laser rangefinder, and an effective engagement range of two to four kilometers round out the payload. The system retains a human operator in the loop for weapons release. The decision architecture is deliberate, not a hardware limitation.

China’s defense robotics ecosystem has been maturing rapidly without much fanfare in Western press. Unitree’s quadrupeds, originally commercial platforms aimed at industrial and research markets, began appearing in PLA exercises years ago. State media showed robot dogs carrying machine guns in breach-and-clear drills. They were airdropped from multi-rotor drones onto rooftops. They were tested in urban combat simulations. Each step was iterative, methodical, and entirely consistent with how China integrates civilian technology into its military-civil fusion strategy.

The WDS 2026 unveiling is a logical continuation of that trajectory. As Amir Husain writes in Forbes, “The engineering phase of ground combat robotics is well past us. We are in the deployment phase now.”

Turkey’s Roketsan had already put a missile-armed robot dog, the KOZ, on the market at IDEF 2025. Several defense observers noted the PF-070’s visual and functional similarity to that platform. Whether the Chinese system is derivative or parallel development matters less than what it confirms: the barriers to entry for armed ground robotics are low enough that multiple nations can field these capabilities within a single defense procurement cycle.

The United States has been active in this space too, though differently. Ghost Robotics’ Vision 60 has been tested in the CENTCOM area of operations, armed with an AI-enabled rifle turret optimized for counter-drone missions. The Marine Corps has demonstrated robot dogs carrying M72 anti-armor rockets. American systems tend to be positioned for counter-UAS, reconnaissance and support roles rather than direct anti-armor fire.

Armed quadrupeds are now a global phenomenon. Ukraine has committed to deploying robot dog units on the frontline. Russia displayed an RPG-armed version as early as 2022. The CENTCOM theater has seen live testing by multiple actors. The technology is operational, proliferating, and being exported.

Notably, China has already been exporting the operational concept, not just the hardware. At the SCO’s “Interaction-2024” counter-terrorism exercise in Xinjiang, rifle-armed robot dogs operated alongside troops from all ten member states, including Pakistan. For countries in China’s strategic orbit, exposure to these systems in joint exercises is the first step toward eventual procurement. The Riyadh display was also an explicit act of market positioning. More than fifty Chinese defense firms participated in WDS 2026, and the delegations examining the PF-070 weren’t there as tourists.

The global arms market for autonomous ground systems is forming right now, and China, like Turkey and Israel before it in the drone market, intends to be a supplier.What this means: When any reasonably funded state actor can deploy hundreds of expendable missile-carrying platforms rather than a handful of crewed armored vehicles, the logic of force employment changes. Attrition calculus changes. So does deterrence. If you’re in defense procurement, intelligence, or national security strategy, you’re now planning for adversaries with access to cheap, precise, autonomous ground lethality. The doctrine, the law, and the strategy still need to catch up.

Source:China’s Robot Dogs Have Been Armed With Missiles – Forbes

---

Nvidia and Meta Sign Multi-Billion Dollar AI Chip Deal

Meta and NVIDIA just announced a deal that analysts estimate in the “tens of billions of dollars,” part of Meta’s commitment to spend up to $135 billion on AI in 2026 alone. Under the agreement, Meta will purchase millions of NVIDIA Grace CPUs, Blackwell GPUs, and next-generation Vera Rubin GPUs to support its buildout of data centers optimized for AI training and inference.

The vendors described the partnership as “multiyear and multigenerational,” spanning on-premises, cloud, and AI infrastructure. Meta will also adopt NVIDIA’s Spectrum-X Ethernet platform across its AI infrastructure for improved performance and efficiency, and the NVIDIA Confidential Computing security system to enable secure AI features on WhatsApp.

This deal is NVIDIA’s first-ever Grace-only deployment. Typically, GPUs and CPUs work in tandem, with GPUs handling training and CPUs focused on system operations. But this configuration points to a shift: AI vendors are increasingly reliant on standalone CPUs better suited to handling inference AI workloads (the phase after training where models actually respond to user queries at scale).

The Vera Rubin platform stands out. NVIDIA and Meta are collaborating on Vera CPUs, the successor to the Grace CPUs, with “potential for large-scale deployment” in 2027. This isn’t just procurement. It’s co-development of next-generation infrastructure.

Meta CEO Mark Zuckerberg framed the partnership around his vision of “personal superintelligence,” AI that can surpass human performance in most cognitive tasks. “We’re excited to expand our partnership with Nvidia to build leading-edge clusters using their Vera Rubin platform to deliver personal superintelligence to everyone in the world,” Zuckerberg said in the release.

NVIDIA CEO Jensen Huang positioned the deal as recognition of Meta’s scale: “No one deploys AI at Meta’s scale—integrating frontier research with industrial-scale infrastructure to power the world’s largest personalization and recommendation systems for billions of users.”

The agreement also includes NVIDIA Confidential Computing for WhatsApp, enabling AI-powered capabilities across the messaging platform while ensuring user data confidentiality and integrity. The technology secures data during computation, not just when it’s being shuttled to a server. It also allows software creators like Meta or third-party AI agent providers “to preserve their intellectual property,” according to NVIDIA’s blog post about the technology.

Meta announced earlier this year it plans to build up to 30 data centers, including 26 in the U.S., by 2028 as part of a $600 billion infrastructure commitment. The NVIDIA deal is a significant chunk of that spend and a signal that the infrastructure race is accelerating.

The deal also raises questions about Meta’s ongoing efforts to create its own chips and become less reliant on NVIDIA. According to the Financial Times, Meta has invested heavily in this area, but the strategy has suffered technical problems and rollout delays. If that’s true, this deal suggests Meta’s pivoting back to buying rather than building, at least for the near term.Key takeaway: Infrastructure is the new moat. Meta’s betting tens of billions that NVIDIA’s chips, not homegrown silicon, will power its AI ambitions through 2027 and beyond. For enterprise IT leaders, this confirms that inference workloads are becoming first-class infrastructure concerns, and the vendors winning those contracts are the ones co-developing next-gen CPUs with their largest customers. If you’re planning AI infrastructure for 2026-2027, standalone CPU deployments for inference are no longer experimental. They’re strategic. 

Sources:

• Nvidia and Meta Agree to Wide-Ranging new AI Chip Deal – AI Business
• Meta will run AI in WhatsApp through NVIDIA’s ‘confidential computing’ – Engadget

---

Group-Evolving Agents Match Human Engineers At Zero Extra Cost

Introducing 𝐆𝐫𝐨𝐮𝐩-𝐄𝐯𝐨𝐥𝐯𝐢𝐧𝐠 𝐀𝐠𝐞𝐧𝐭𝐬 (𝐆𝐄𝐀), a new paradigm for open-ended self-improvement in AI agents.

The core shift is simple but radical:
👉 𝐭𝐡𝐞 𝐮𝐧𝐢𝐭 𝐨𝐟 𝐞𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐢𝐬 𝐧𝐨 𝐥𝐨𝐧𝐠𝐞𝐫 𝐚 𝐬𝐢𝐧𝐠𝐥𝐞 𝐚𝐠𝐞𝐧𝐭, 𝐛𝐮𝐭 𝐚 𝐠𝐫𝐨𝐮𝐩… pic.twitter.com/M9MXOyrTL6

— Xin Eric Wang (@xwang_lk) February 5, 2026

Agents built on today’s models often break with simple changes (a new library, a workflow modification) and require a human engineer to fix it. That’s one of the most persistent challenges in deploying AI for the enterprise: creating agents that can adapt to dynamic environments without constant hand-holding.

Researchers at the University of California, Santa Barbara just published a solution. Group-Evolving Agents (GEA) is a new framework that enables groups of AI agents to evolve together, sharing experiences and reusing their innovations to autonomously improve over time. In experiments on complex coding and software engineering tasks, GEA beat existing self-improving frameworks. The big news for enterprise decision-makers: the system autonomously evolved agents that matched or exceeded the performance of frameworks painstakingly designed by human experts.

Most existing agentic AI systems rely on fixed architectures designed by engineers. To overcome this, researchers have long sought to create self-evolving agents that can autonomously modify their own code and structure. However, current approaches have a structural flaw: they’re designed around “individual-centric” processes inspired by biological evolution. These methods typically use a tree-structured approach where a single “parent” agent produces offspring, creating distinct evolutionary branches that remain strictly isolated from one another.

This isolation creates a silo effect. An agent in one branch can’t access the data, tools, or workflows discovered by an agent in a parallel branch. If a specific lineage fails to be selected for the next generation, any valuable discovery made by that agent dies out with it.

GEA flips the paradigm by treating a group of agents, rather than an individual, as the fundamental unit of evolution. Unlike traditional systems where an agent only learns from its direct parent, GEA creates a shared pool of collective experience. This pool contains the evolutionary traces from all members of the parent group, including code modifications, successful solutions to tasks, and tool invocation histories. Every agent in the group gains access to this collective history, allowing them to learn from the breakthroughs and mistakes of their peers.

A “Reflection Module,” powered by a large language model, analyzes this collective history to identify group-wide patterns. For instance, if one agent discovers a high-performing debugging tool while another perfects a testing workflow, the system extracts both insights. Based on this analysis, the system generates high-level “evolution directives” that guide the creation of the child group. This ensures the next generation has the combined strengths of all their parents, rather than just the traits of a single lineage.

The results: On SWE-bench Verified, a benchmark consisting of real GitHub issues including bugs and feature requests, GEA achieved a 71.0% success rate, compared to the baseline’s 56.7%. On Polyglot, which tests code generation across diverse programming languages, GEA achieved 88.3% against the baseline’s 68.3%.

For enterprise R&D teams, the critical finding is that GEA allows AI to design itself as well as human engineers. On SWE-bench, GEA’s 71.0% success rate matches the performance of OpenHands, the top human-designed open-source framework. On Polyglot, GEA beat Aider, a popular coding assistant, which achieved 52.0%.

This collaborative approach also makes the system more robust against failure. In their experiments, the researchers intentionally broke agents by manually injecting bugs into their implementations. GEA repaired these critical bugs in an average of 1.4 iterations, while the baseline took 5 iterations. The system uses the “healthy” members of the group to diagnose and patch the compromised ones.

“GEA is explicitly a two-stage system: (1) agent evolution, then (2) inference/deployment,” the researchers told VentureBeat. “After evolution, you deploy a single evolved agent… so enterprise inference cost is essentially unchanged versus a standard single-agent setup.”

The top GEA agent integrated traits from 17 unique ancestors (representing 28% of the population) whereas the best baseline agent integrated traits from only 9. In effect, GEA creates a “super-employee” that has the combined best practices of the entire group.

And the improvements discovered by GEA aren’t tied to a specific underlying model. Agents evolved using one model, such as Claude, maintained their performance gains even when the underlying engine was swapped to another model family, such as GPT-5.1 or GPT-o3-mini. This transferability offers enterprises the flexibility to switch model providers without losing the custom architectural optimizations their agents have learned.Why this matters: This isn’t just a research paper. It’s a roadmap for organizations that want AI systems capable of autonomous improvement without scaling up engineering teams. If agents can evolve collectively and self-heal when they break, the operational overhead of maintaining agentic systems drops dramatically. The researchers plan to release the official code soon, but developers can already begin implementing the GEA architecture conceptually on top of existing agent frameworks. For compliance-heavy industries, the authors recommend “non-evolvable guardrails, such as sandboxed execution, policy constraints, and verification layers.” The question isn’t whether this works. It’s whether your infrastructure and team are ready to let agents improve themselves. 

Source:New agent framework matches human-engineered AI systems and adds zero inference cost to deploy – VentureBeat

---

Tracking

What CEOs Should Be Watching:

• Anthropic is clashing with the Pentagon over AI use – Anthropic’s work with DoD is “under review” after the company demanded assurances its models won’t be used for autonomous weapons or mass surveillance. Pentagon threatened to label Anthropic a “supply chain risk.” This dispute defines the boundaries between commercial AI and defense applications.
• Anthropic releases Claude Sonnet 4.6 – New model handles multi-step computer actions, coding, and design. Can fill out web forms and coordinate information across browser tabs. Sonnet 4.6 is now the default free model, replacing previous versions.
• ByteDance launches Seedance 2.0 – Next-gen video creation model with unified audio-video generation. Supports text, image, audio, and video inputs. Generates multi-shot film sequences with sound in ~60 seconds. Roadmap points to Seedance 2.5 mid-2026 with 4K output and real-time generation.

---

Bottom Line

We’re watching three parallel developments that intersect at a single uncomfortable point: control is slipping.

China’s exporting missile-armed robot dogs the same way it exported commercial drones. By the time Western policymakers realized the strategic exposure, the market was already formed. The PF-070 isn’t a prototype. It’s a product, displayed in Riyadh to buyers from the Middle East and Asia who are ready to write checks. The operational doctrine for autonomous ground lethality is being written right now, not in Washington or Brussels, but in joint exercises across the Shanghai Cooperation Organization. By the time NATO agrees on a framework, adversaries will have fielded hundreds of systems.

Meta’s $135 billion AI spend and its tens-of-billions bet on NVIDIA chips signal that infrastructure dominance is the new strategic imperative. Zuckerberg’s framing around “personal superintelligence” isn’t hyperbole. It’s a declaration that Meta intends to own the inference layer for billions of users. The shift to standalone Grace CPUs for inference workloads confirms what enterprise IT leaders already suspected: the bottleneck isn’t training anymore. It’s serving billions of queries per second without collapsing under the cost or latency. The companies that solve inference at scale are the ones that will own the AI economy.

And then there’s the part that should keep you up at night: GEA demonstrated that AI agents can now improve themselves faster than the humans who designed them. The system didn’t just match human-engineered frameworks. It integrated traits from 28% of its population to create a “super-employee” that self-healed when broken and adapted across different model families. This isn’t AGI, but it’s something more consequential: distributed intelligence that learns collectively and compounds its capabilities over time without human guidance.

The pattern is clear. The systems are proliferating globally, the infrastructure is being built at unprecedented scale, and the agents are learning to evolve autonomously. The race isn’t about who builds the best AI. It’s about who controls the export markets, who owns the infrastructure, and whether we’re ready for systems that improve themselves without asking permission. Right now, we’re not.

---

Key People

• Amir Husain – Founder WQ Foundry, SkyGrid (acquired by Boeing), SparkCognition (AI Unicorn); author of Forbes article on Chinese robot dogs
• Mark Zuckerberg – CEO, Meta; announced $135B AI spend for 2026 and partnership with NVIDIA
• Jensen Huang – CEO, NVIDIA; described Meta as deploying “AI at scale” unlike any other company
• Zhaotian Weng & Xin Eric Wang – UC Santa Barbara researchers; co-authors of Group-Evolving Agents paper
• General John Allen – Co-author with Amir Husain on “Hyperwar” thesis about AI-enabled systems operating at machine speed

---

Sources

1. China’s Robot Dogs Have Been Armed With Missiles – Forbes

2. Nvidia and Meta Agree to Wide-Ranging new AI Chip Deal – AI Business

3. Meta will run AI in WhatsApp through NVIDIA’s ‘confidential computing’ – Engadget

4. New agent framework matches human-engineered AI systems — and adds zero inference cost to deploy – VentureBeat