Uber's AI-Generated Code Hits 65-72% as Developers Shift to Agent Orchestration

GitHub, OpenAI, Anthropic, and Duckbill adopt risk-based code review triage to handle AI-driven PR explosion

GitHub disclosed that monthly commits have grown from 1.4 billion to 2.9 billion since April, with PRs and commits nearly doubling in Q4 2025 alone, creating unsustainable code review bottlenecks. Both OpenAI and Anthropic now use "blast radius" classification to route changes, with Anthropic planning to eventually have Claude instances merge low-risk PRs autonomously. Duckbill Group, a five-person startup, reduced code review time from 2 days to near-zero for low-risk changes by implementing risk-based triage and improved guardrails, increasing merged PRs 94% (from 80/week to 154/week), with human-reviewed PRs now merging in 26 hours median versus 1 hour for non-reviewed changes. GitHub responded by adding 3+ million CPU cores and 120 petabytes of storage, with Azure now serving 58% of platform load (up from 12% in May). These concrete metrics illustrate the infrastructure and operational challenges frontier labs and development organizations face as AI agents dominate code generation.

NSA/FBI/CISA report: Chinese AI companies conducting industrial-scale distillation of U.S. frontier models since late 2024

A joint cybersecurity advisory from NSA, CISA, and FBI details systematic extraction of billions of tokens from U.S. frontier models by DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun, and Z.AI. DeepSeek distilled specialized capabilities from Claude 3.7, Sonnet 4/4.5, Opus 4.1, Gemini 2.5, GPT-4/4o, and Grok 4 to train its R1 and V3 models, targeting legal specialization, API rule-driven tasks, chain-of-thought reasoning, and agentic functions. The companies route requests through multiple pathways including APIs, cloud providers, and third-party aggregators, using "transfer stations" (gray-market proxies) to bypass geographic restrictions and evade detection. The agencies characterize this as malicious activity "likely with Chinese government awareness" and note that distillation forms the core—not merely a supplement—of these companies' AI development strategy. The report estimates that industrial-scale distillation has significantly shortened AI development timelines and reduced financial expenditures for Chinese companies, though it raises questions about why the Trump administration has not expanded export controls to target distillation alongside chip restrictions.

Cognition's Devin coding agent reaches $48B valuation on ~$900M ARR, doubling revenue in four months

Cognition raised $2B at a $48B valuation (84% increase since May), with annual run-rate revenue nearly doubling from $492M to ~$900M in four months. The round was led by Andreessen Horowitz and Accel, alongside existing investors Founders Fund, General Catalyst, and Avenir. Devin now powers engineering teams across NVIDIA (chip design), GE Aerospace (aviation), Citi (financial services), Mercedes-Benz (automotive), and Modal (AI infrastructure). New capabilities include Devin Auto-Triage for incident investigation, Devin Security Swarm for vulnerability detection, and Devin Automations for event-driven work from Slack, GitHub, and Linear. Cognition is positioning itself as an independent agent lab that can combine models from multiple providers rather than tying customers to one vendor, and has opened offices in Washington D.C., Tokyo, Singapore, London, São Paulo, and Madrid alongside existing hubs.

OpenAI solves Navier-Stokes Millennium Prize Problem using 10,000 agents and 88 hours of compute

OpenAI's next-generation model (beyond GPT-6 Astra) produced a proposed proof to the Navier-Stokes Millennium Prize Problem—one of the deepest unsolved problems in mathematics—via massive test-time compute orchestration using 10,000 agents over 88 hours, followed by 17 hours of Lean formalization. The estimated cost is $10M–$40M. This demonstrates test-time compute scaling as a first-class axis alongside pretraining, with significant implications for economics (costs collapsing from millions to consumer-accessible over time) and infrastructure (unstructured parallel compute plus orchestration now critical).

Dell projects AI will represent 75% of datacenter demand by 2030, signaling sustained infrastructure buildout

Dell's COO Jeff Clarke stated that AI will drive 75% of datacenter demand by 2030, reflecting accelerating AI adoption timelines. Dell sold $16.4B in AI systems in Q2 FY2027 (double year-over-year) with a $95B backlog ($60.9B booked in-quarter). The company projects $74B in AI system sales for full-year FY2027 ($19B Q3, $22.5B Q4), with traditional server sales now eclipsing AI servers in Q4 as customers consolidate aging infrastructure (8-14 year old PowerEdge 14G machines collapsing into newer 17G/18G systems). Dell's Infrastructure Solutions Group revenue hit $31.78B (up 89% YoY) with operating income of $4.78B (up 3.25X), reflecting the company's ability to secure GPU allocations and scale manufacturing. This forecast underscores sustained demand for high-margin AI infrastructure over the next four years, though it also signals competitive pressure on traditional workloads.

Trump administration imposes $100,000 H-1B visa fee, constraining skilled immigration pipeline for AI infrastructure firms

The Trump administration implemented a $100,000 fee on H-1B visas for beneficiaries residing outside the U.S. and modified the lottery selection process to prioritize higher-paid workers (effective February 27, 2026). The weighted selection rule increases likelihood of selection for Wage Level III (3x submissions, 45.87% selection probability) and IV (4x submissions, 61.16%) while decreasing it for Level I (1x submission, 15.29%). However, the policy targets the median H-1B job earning $130,075 annually (91st percentile wage) rather than just lower-wage abuses, effectively shrinking skilled immigration rather than improving it. This directly constrains the pipeline of technical workers that AI infrastructure and hardware companies like Cerebras depend on for R&D and engineering roles. Two legal challenges have been filed; the Chamber of Commerce appeal to the D.C. Circuit is pending.

How Uber Uses AI for Development: Inside Look at Agentic Systems, Internal Tooling, and Engineering at Scale

The Pragmatic Engineer published a detailed essay on Uber's internal AI infrastructure and developer workflows. Main takeaways:

• Uber has built a four-layer agentic system: internal AI platform (on Michelangelo), context sources (code/docs/Slack/JIRA), industry agents (Claude Code, Copilot, Codex), and specialized agents for testing and code review
• 84% of Uber developers are agentic coding users; 65-72% of code is AI-generated in IDEs; Claude Code usage nearly doubled from 32% (Dec 2025) to 63% (Feb 2026)
• Uber built MCP Gateway to proxy internal endpoints as MCP servers, Agent Builder for no-code multi-agent workflows, and AIFX CLI for unified agent provisioning and discovery across the company
• Developer workflows have shifted from single-threaded IDE work to orchestrating parallel agents; engineers naturally kick off multiple background tasks, creating resource and cost challenges
• AI-related costs are up 6x since 2024; Uber built specialized tools (Code Inbox, uReview, Autocover generating 5,000+ unit tests/month, Shepherd for migrations) to handle increased code review noise and testing burden

China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies

National Security Agency, Cybersecurity and Infrastructure Security Agency, Federal Bureau of Investigation in NSA, CISA, and FBI (Joint Cybersecurity Advisory) published a report titled "China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies". Main takeaways:

• DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI extracted billions of tokens from Claude, GPT, Gemini, and Grok variants since late 2024 to train R1, V3, Kimi-K3, and other models
• DeepSeek's reported $5.6M training cost is misleading; distillation of specialized capabilities (CoT reasoning, agentic functions, legal optimization) significantly reduced actual R&D expenditure
• Chinese companies use 'transfer stations' (API proxies), bulk subscription sharing, and automated failover tactics to bypass geographic restrictions and evade detection across multiple U.S. providers
• Distillation represents core development strategy, not supplementary technique; enables significantly shorter timelines and reduced financial expenditures for frontier model training
• Specific models and capabilities targeted documented in detailed table: Moonshot distilled Claude Fable 5 for Kimi-K3; MiniMax used prompt injection attacks on Claude Code; Z.AI extracted billions of GPT-5.5 tokens by mid-2026

Mark Zuckerberg on Meta's Muse Personal AI Agent, Privacy Architecture, and Competitive Strategy

Alex Heath in Sources (Alex Heath) published an interview titled "Mark Zuckerberg on Meta's Muse Personal AI Agent, Privacy Architecture, and Competitive Strategy". Main takeaways:

• Muse launches with 100M tokens/week free tier; Meta plans to monetize via small transaction cuts from businesses, positioning personal agents as self-paying through user earnings/savings
• Zuckerberg personally recruited Signal founder Moxie Marlinspike to build a confidential virtual machine where even Meta cannot see Muse's actions, with separate sentinel agents monitoring for prompt injections
• Meta is undercutting startups like Instinct and Town on pricing by leveraging existing business revenue to subsidize agent usage, creating high barrier to entry for competitors
• Zuckerberg attributes datacenter backlash partly to speculative developers flipping sites rather than long-term community investment, distinguishing Meta's infrastructure strategy

Hybrid HiSparse: Serving GLM 5.3 at 1M Context on 8×H200 with 9.75x Higher Concurrency

vLLM Team (Matthew Bonanni, Lucas Wilkinson, Fares Obeid, with Huawei and Mistral collaboration) in vLLM Blog published a report titled "Hybrid HiSparse: Serving GLM 5.3 at 1M Context on 8×H200 with 9.75x Higher Concurrency". Main takeaways:

• Hybrid HiSparse enables GLM 5.3 (1M context) to run on 8×H200 at full context length, previously impossible due to GPU memory constraints
• Achieves 9.75x concurrency improvement (39 concurrent requests vs. 4 with standard offloading) on the same hardware by selectively offloading sparse-MLA KV cache to CPU only under memory pressure
• Hot buffer pages stay GPU-resident and shared across requests, reducing CPU-GPU transfer overhead compared to full offloading; only pays transfer cost when system is under KV cache pressure
• Integrates with vLLM's existing infrastructure (prefix caching, speculative decoding, disaggregation) without requiring architectural changes; planned for v0.30 release
• Concrete benchmark: 13-turn agentic conversations with 74K-token first turn and 753-token later turns show substantial throughput gains across interactivity-throughput sweep