AI Meets the Physical World: Neuromorphic Computing, MHS Standard, and Code-as-World Emerge in Same Week

According to a report by Leiphone, Anthropic's Model Hardware Standard (MHS) preview aims to solve the two-decade-old problem of fragmented hardware interfaces in labs and industry. Instead of forcing devices to rewrite their underlying code, MHS inserts a "unified translator" between AI agents and physical devices. Agents can issue simplified commands to obtain device information or write control parameters, and a reference file records real-time status, parameters, and sensor data for all devices, so an agent can grasp the entire lab without connecting to each device individually. MHS also includes safety guardrails: thresholds such as maximum power and maximum movement range are written into the driver layer, preventing AI from crossing safety boundaries.

The practical effects have been documented. A Carnegie Mellon University team used MHS to coordinate a liquid handler, a plate reader, a robotic arm, and a monitoring camera; they accelerated experiment orchestration and monitoring by about three times, according to Leiphone. At quantum computing company QuEra, MHS helped Claude autonomously re-lock lasers that had lost frequency lock. The recovery time shrank from 150 seconds to about 6 seconds, and blind tests across 700 attempts reached a 99.3% success rate. In a subsequent PID parameter tuning task, Claude ran 363 experiments in 16 hours and reduced residual error from 15.7 mV to 1.55 mV, a tenfold improvement. During a 19-hour continuous test, the expert-tuned parameters lost lock about 1.6 times per hour, while Claude-tuned parameters did not lose lock once. Anthropic designed two control modes: "online commanding" for step-by-step adjustments and "offline burst" for pre-packaged deterministic Python scripts that execute locally without waiting for cloud inference, thereby bypassing network latency.

At the same congress, according to Leiphone, the roundtable titled "Brain-inspired Computing for Autonomous Robot Decision-Making" featured Alois Knoll, a professor at the Technical University of Munich, who noted that the unit of AI computation is shifting from exaflops to gigawatts as data centers expand. He argued that if 1 billion to 10 billion humanoid robots were deployed, cumulative energy demand using current high-power computing models would become unsustainable. Neuromorphic computing, which mimics biological neurons and event-driven signaling, offers a potential alternative. Knoll highlighted that biological brains perform computation and memory in a highly coupled way and use event-driven communication, so neurons only transmit pulses when events occur. Neuromorphic systems adopt spiking neural networks to reduce continuous information exchange. Robots must continuously interact with environments, making them a natural application for such event-driven architectures.

The roundtable converged on a more modest role for neuromorphic computing. Xu Ningyi, founder and CEO of HuiZhi Intelligent, argued that robots are moving from simple "sense-decide-act" chains to layered, asynchronous, self-learning systems. She proposed a division into System 0 for movement, System 1 for fast perception-to-action links, and System 2 for slow, complex reasoning. Li Zhaoshi, director of the MuXi AI Research Institute, explained that data movement between memory and compute units often costs more than the computation itself, and neuromorphic computing reduces data movement by handling sparse, event-driven signals. He said System 2 remains better suited to traditional digital computing, while System 0 and System 1 — including sensor perception and some simple VLA tasks — are promising for neuromorphic computing. Touch is a typical example: most of the time no new tactile signal is activated, yet traditional digital processing keeps running. Neuromorphic or asynchronous circuits could cut power consumption in such naturally sparse, event-triggered tasks.

On commercialization, panelists offered timelines. Yang Yuxin, chief marketing officer of Black Sesame Technologies, said neuromorphic chips could reach acceptable performance and cost in 3 to 5 years, but software ecosystem maturity would require longer, putting fuller commercialization at 5 to 8 years. Li Zhaoshi emphasized scenario differences: semi-structured factories might see large-scale deployment in 3 to 5 years, while households and services could take 5 to 8 years due to long-tail situations, safety, and ethics. Xu Ningyi said her company's robots have already begun deployment in some semi-structured factories. Knoll summarized that hardware solutions may mature first, starting with perception and lower-level capabilities, with broader market maturity roughly in 3 to 8 years.

According to MarkTechPost, MirroS released Code-as-World, a paradigm that represents physical scenes as executable code rather than pixels. The system generates a scene.json file that can be run in MuJoCo, an agent can verify against the source video, and anyone can edit and re-simulate. The technical report argues that pixels are evidence of a physical scene, not its ontology; video models may predict plausible frames without representing mass, contact, or gravity. Code-as-World uses an agentic loop with up to five rounds of propose, instantiate, execute, render, and verify. For video input, SAM 3 provides instance masks, VGGT-Omega estimates depth and camera geometry, and SAM 3D generates per-object meshes. The verified worlds become training data with exact physical labels. The resulting Code-as-World-VL-9B model scored 55.4 MRA on QuantiPhy-validation, above Gemini-3.1 Flash at 54.8 and about 15 points above the strongest open-weight baseline. MirroS open-sourced the code and checkpoints under Apache 2.0, fine-tuned from Qwen3.5-4B and Qwen3.5-9B.

The technical report also includes ablation results: image-space-only training scores 44.2 (4B) and 50.9 (9B); adding world-space sources lifts them to 50.6 and 55.4. The 9B model improves from 63.7 to 68.3 on RefCOCO and from 20.1 to 26.6 on GOT-10K after world-space reinforcement learning. The system is limited to rigid-body scenes, and the model does not learn the discovery loop itself.