Botwin's Morning Wire

In May, German Gref, chief executive of Sberbank, told Russian state broadcaster Channel One that he hoped to run the bank's GigaChat AI model on Chinese processors, illustrating the difficulty of obtaining advanced chips under Western sanctions.

Western export controls have limited Russia's access to semiconductor technologies from the United States and its allies, prompting efforts to develop alternative supply chains.

Tom's Hardware reported that the most likely candidate to power Sberbank's AI systems is Huawei's Ascend 950 family, the most advanced silicon China currently produces. Sberbank, Russia's largest lender and a major driver of the country's AI initiatives, is seeking Chinese chips because Western sanctions block its access to advanced hardware abroad.

The arrangement could enable a trade relationship between China and Russia in semiconductor technology, but the long-term adequacy of this supply chain for Russia's AI development remains unclear.

Meta introduced Muse Spark 1.1, a multimodal reasoning model for agentic tasks, on Tuesday. The new model succeeds Muse Spark and adds improvements in tool use, coding, and multimodal understanding.

Muse Spark 1.1 supports a context window of one million tokens, allowing it to retain actions and retrieve information from earlier work while compacting critical steps for later use. The model can orchestrate multiple agents, delegate tasks, and adapt to changes in task context without user intervention.

In computer‑use workflows, Muse Spark 1.1 maintains context across extended sessions, adapts to evolving requirements, and can automate steps or interact directly with interfaces as appropriate. It can write scripts when automation is more efficient and generate batches of actions at each step.

For coding, the model shows improvements on real‑world tasks involving large codebases. It can diagnose and fix complex bugs, add features to enterprise systems, and execute large migrations. Testing in OpenCode demonstrates its ability to build a web application, capture screenshots to identify failures, locate relevant code, implement fixes, and validate changes.

Safety evaluations were conducted according to Meta’s Advanced AI Scaling Framework. Assessments across chemical and biological, cybersecurity, and loss‑of‑control risk categories indicate the model operates within safe margins, showing resistance to jailbreaks, prompt injection, and other adversarial attacks, with lower hallucination and sycophancy rates.

Developers can access Muse Spark 1.1 through the public preview of the Meta Model API. The model is available in the Meta AI app and on meta.ai in "Thinking" mode. Early partners have praised its capabilities. Amjad Masad, CEO of Replit, said the model packs a million‑token context, multimodal support, built‑in search with citations, strong reasoning, strong coding abilities, structured output, and parallel tool calling in an OpenAI‑compatible package. Saoud Rizwan, CEO of Cline, noted the combination of long context handling with strong coding and reasoning makes it viable for large‑scale coding workloads. Yashodha Bhavnani, VP of AI Products at Box, said the model delivers enterprise‑level intelligence and procedural workflow strengths for professional services, public sector, and industrial operations.

Meta also launched Muse Image this week, expanding its multimodal offerings. The company said it continues to develop more capable models and plans to share future updates.

The United States Army has optimized sustainment for permissive environments, but its current logistics model may be insufficient for large‑scale, contested conflict, according to senior Army officials. History shows that operations without adequate sustainment can fail, as seen in the 1941 German invasion of the Soviet Union and in recent fighting in Ukraine, where pervasive sensing and drone attacks have targeted supply routes, according to analysts and open‑source reports. The Army consumes tens of thousands of gallons of fuel daily for an armored brigade, requiring heavy tactical vehicles that are detectable by thermal and electromagnetic signatures, officials said. Current distribution platforms lack the protection needed to survive deep‑strike attacks, and maintenance shortfalls reduce available capacity, according to a recent Army study. Ammunition expenditure rates observed in Ukraine are comparable to those of World War II, creating challenges for transporting heavy 155‑mm shells and rocket pods across contested terrain, the study noted. Centralized supply depots represent high‑value targets for long‑range precision fires, limiting the speed and volume of resupply, analysts said. To survive contested environments, the Army intends to replace its hub‑and‑spoke model with a decentralized network of smaller, mobile sustainment nodes, officials said. Distributed caches of fuel, water, and ammunition, along with autonomous ground vehicles and heavy‑lift drones, are being pursued to reduce reliance on fixed depots, the Army stated. Sustainment units will be equipped with embedded counter‑UAS systems and short‑range air defense to protect themselves, and will be trained to operate in GPS‑denied, signature‑managed conditions, the Army added. The Army’s modernization priorities have historically favored maneuver and firepower over sustainment, a trend reflected in budget allocations and programmatic focus, officials said. Elevating logistics to a primary warfighting function will require reallocating resources, revising the tooth‑to‑tail ratio concept, and incorporating realistic logistical stress tests into training, according to senior leaders. Without improvements in physical resilience, dispersion, and endurance, the Army risks fielding a force with tactical superiority but operational vulnerability, officials warned. Future large‑scale conflict may be decided by the ability of sustainment to endure persistent attack rather than by the speed or range of individual weapons systems.