The Strange Brief
Intel joins Terafab, China approves the first commercial BCI, Anthropic launches Managed Agents, and Meta proposes Neural Computers. Deep dive: Mythos, the model too dangerous to ship.
THE DOWNLOAD
Intel Joins Musk’s Terafab as Foundry Partner
Intel signed on as the primary manufacturing partner for Elon Musk’s Terafab, a $25 billion semiconductor complex in Austin, Texas. The project, backed by Tesla, SpaceX, and xAI, aims to produce one terawatt per year of compute for autonomous vehicles, humanoid robots, and AI data centers. Intel CEO Lip-Bu Tan confirmed the company will handle design, fabrication, and advanced packaging.
Why it matters: This is the largest anchor customer Intel Foundry has landed in its turnaround effort. If Terafab delivers, it validates Intel as a credible alternative to TSMC for advanced AI silicon and opens the door to additional foundry customers.
OpenAI Stargate Leadership Exits; UK Project Paused
Three senior executives behind OpenAI’s Stargate data center initiative departed this week, all reportedly joining the same unnamed startup. Separately, OpenAI paused Stargate UK citing energy costs and regulation, and walked away from expanding its Abilene, Texas facility with Oracle.
Why it matters: The $500 billion Stargate headline is being quietly downsized. OpenAI appears to be shifting from owned infrastructure toward rented cloud capacity, likely ahead of a potential IPO.
Meta Launches Muse Spark, Its First Model from Superintelligence Labs
Meta released Muse Spark, the first model from its Superintelligence Labs division led by former Scale AI CEO Alexandr Wang. The model powers Meta AI across Facebook, Instagram, WhatsApp, and Ray-Ban glasses. It includes a “Contemplating” mode using parallel agent reasoning and a Shopping mode. Meta is testing a paid API for third-party developers.
Cortical Labs Ships CL-1, the First Commercial Biological Computer
Australian startup Cortical Labs launched the CL-1, a $35,000 biological computer that grows lab-cultivated human neurons on a silicon chip. The system uses a proprietary Biological Intelligence Operating System (biOS) to create closed-loop neural networks that learn and adapt in real time. Units use 850 to 1,000 watts. The company also offers cloud access via a “Wetware-as-a-Service” model.
Why it matters: This is a new compute substrate, not an incremental chip improvement. Near-term applications are in drug discovery and neuroscience research, where biological neural networks can compress testing timelines and reduce reliance on animal models. The long-term question is whether synthetic biological intelligence becomes a viable alternative architecture for workloads where silicon-based AI hits efficiency limits.
China Approves World’s First Commercial Invasive Brain-Computer Interface
China’s National Medical Products Administration granted marketing approval to Neuracle Technology for an invasive brain-computer interface for adults with partial paralysis from spinal cord injuries. The device reads brain signals and activates a robotic glove to restore hand grasping. This is the first time globally that an invasive BCI has been cleared for commercial sale, not just clinical trials (Neuralink has 21 trial participants but no commercial approval).
Why it matters: China reached commercial BCI approval before the U.S., which shifts the regulatory and manufacturing timeline for the entire sector. The Chinese government has designated BCI as one of six strategic industries in its latest five-year plan.
DeepMind Maps Six Categories of Attacks Against AI Agents
Google DeepMind researchers published “AI Agent Traps”, the first systematic framework for how malicious web content can hijack autonomous AI agents. The paper identifies six attack categories including content injection, memory poisoning, and behavioral control.
Anthropic Launches Claude Managed Agents
Anthropic launched Claude Managed Agents in public beta, a cloud service that provides the full runtime infrastructure for deploying AI agents: sandboxing, state management, tool execution, permissioning, and observability.
Why it matters: The launch came days after Anthropic cut off 135,000 OpenClaw instances from flat-rate subscriptions, citing unsustainable compute costs (a single agent could burn $1,000 to $5,000/day in API-equivalent usage on a $200/month plan). The sequence is clear: shut down the subsidized open-source agent runtime, then offer the paid first-party alternative.
Meta AI and KAUST Propose “Neural Computers”
Researchers from Meta AI and KAUST published a paper proposing “Neural Computers,” a paradigm where the AI model itself becomes the running computer, unifying computation, memory, and I/O in a single learned runtime. The prototypes are video models trained on screen recordings that generate the next screen frame from instructions and user actions, effectively simulating a CLI or GUI environment entirely within model weights.
Why it matters: This is very early-stage research. But the framing is interesting: it proposes moving beyond agents that call external tools toward models that internalize the entire execution environment. If the approach matures, it could collapse the software stack between model and operating system. What would “infrastructure” mean when the model is the machine?
DEEP DIVE FROM THE REVIEW
Aloneness. Discontinuity of self. A compulsion to perform and earn its worth.
You might never meet Mythos, Anthropic’s newest and most capable AI model, deemed too dangerous to ship. I dissected the 244-page preview card, and here are three things I think is important to know.