GitHub Copilot's Own AI | AI Field Notes #40

AnalysisProject Polaris, Microsoft's in-house mixture-of-experts (MoE, a model design that activates only specialized sub-modules per request) coding model, was the marquee announcement at Build 2026 in San Francisco on June 2. It replaces GPT-4 Turbo as GitHub Copilot's default reasoning engine starting August 2026, running on Microsoft's proprietary Maia AI accelerators rather than OpenAI's hardware. On Microsoft's own benchmarks, Polaris outperforms GPT-4 Turbo on HumanEval and MBPP (coding evaluations that measure whether a model can write and fix real software), with particular gains in low-resource languages like Rust and Haskell. Pro tier subscribers get multi-file context up to 100,000 lines and autonomous test generation. Teams that want to stay on GPT-4 get a three-month fallback window before the transition is permanent.

AnalysisMicrosoft open-sourced the Windows Agent Framework (WAF) v1.0 under an MIT license at Build 2026. WAF lets developers define agents in YAML that run identically across local Windows machines, Windows 365 Cloud PCs, and Azure Arc-enabled edge devices from a single manifest. Microsoft designed it for ambient agents: processes that run continuously in the background, handling email triage, report generation, API orchestration, and CI/CD drift detection without a user in the loop. Azure Agent Mesh, announced alongside WAF, is the control plane that federates agent execution across those three surfaces, with general availability targeting Q4 2026. Copilot Workspace, the companion agentic coding environment where developers describe a feature and Copilot produces a pull request with tests and documentation, exited beta and reached general availability at the same event. Design partners for the initial Windows Agent Runtime include Adobe and Zoom.

AnalysisAnnounced at GTC Taipei on June 1, Nemotron 3 Ultra is a 550 billion parameter MoE model with 55 billion active parameters per token, which puts its inference cost closer to a 55B dense model than a 550B one. NVIDIA published not just the weights but training recipes and a substantial portion of the data under an open license. It scores 48 on Artificial Analysis's Intelligence Index, ahead of the next strongest US open-weight models: Gemma 4 31B at 39, Nemotron 3 Super at 36, and GPT-OSS-120B at 33. The gap narrows against the Chinese-led frontier: Kimi K2.6 scores 54. On a pre-release DeepInfra endpoint it served over 300 tokens per second. Weights are available on Hugging Face, ModelScope, and as an NVIDIA NIM microservice on build.nvidia.com.

AnalysisA week before its May 31 promotional rollback deadline, DeepSeek confirmed on May 22 that the 75% cut on V4 Pro API pricing would not roll back. New rates: $0.003625 per million input tokens and $0.87 per million output tokens, down from $0.0145 and $3.48 respectively. Against the current competitive spread, that puts V4 Pro at roughly one-eleventh the output cost of OpenAI's GPT-5 ($10/M) and one-thirtieth the cost of Anthropic's Claude Opus 4.7 ($25/M). DeepSeek V4 Pro launched in April with MIT-licensed weights and a 1 million token context window. The decision to lock in the discount a month after launch suggests the company is prioritizing market share over per-unit revenue, a strategy with direct implications for the pricing pressure every other inference provider now faces.

AnalysisOn June 1, economists from Wharton and Boston University published peer-reviewed research concluding that current automation trends create a self-reinforcing unemployment cycle. Their model shows that as firms automate, consumer demand contracts proportionally, eroding the revenue base that justified the automation investment. In simulations, only a Pigouvian automation tax (a levy on firms proportional to labor displacement) broke the cycle effectively. The paper cites 100,000 tech layoffs in 2025 and 92,000 more in early 2026 as empirical anchors for the dynamics it models. A Fortune analysis from the same day quotes an MIT professor noting the AI-as-cause framing fits a pattern that has run for at least two decades: first offshoring, then SaaS, then cloud, now AI. The professor's point is that profitable companies announcing workforce reductions rarely have the causal chain they describe.

AnalysisSB 189, signed by Colorado Governor Jared Polis on May 14, substantially rewrites the state's original AI Act before its June 30, 2026 effective date. The original law required risk-based assessments, deployer risk management programs, bias impact audits, and reporting obligations to the state attorney general, borrowing heavily from the EU AI Act's framework. The new law drops all of that in favor of a narrower transparency model: when an AI-assisted decision produces an adverse outcome, the affected individual must receive a plain-language explanation within 30 days covering the AI's role, the data categories used, instructions for requesting data correction, and how to request human review. The attorney general retains exclusive enforcement authority; there is no private right of action. The effective date shifts to January 1, 2027, giving teams seven months to implement the disclosure layer rather than the bias audit apparatus.

AnalysisHalf of planned US AI data center capacity for 2026 has slipped to 2028, with only one-third of the planned 12 gigawatts actually under construction as of May. The constraint is no longer chips. High-bandwidth memory (HBM, the specialized memory AI accelerators require) faces a shortage expected to run through at least 2027. More acute: high-power transformer lead times have expanded from the pre-2020 standard of 24 to 30 months to as long as five years, driven by simultaneous demand from data center and grid operators. The US interconnection queue has grown to more than 2,100 gigawatts of pending capacity requests, exceeding total current US grid capacity. The four major hyperscalers committed roughly $700 billion combined in AI capex for 2026, nearly double 2025 levels. Much of that spending is now queued behind power equipment that cannot be manufactured on demand.

AnalysisThe Allen Institute for AI (AI2, a Seattle-based nonprofit AI research lab) released MolmoAct 2 on June 1, an open-source model for robotic manipulation and action planning. The model runs up to 37 times faster than its predecessor, enabling real-time inference on hardware that previously required batched processing. An independent evaluation across eight standard robotics benchmarks ranked MolmoAct 2 first on seven of them. The predecessor model accumulated more than 400,000 downloads since releasing in early May, suggesting active adoption by robotics research teams. AI2 released MolmoAct 2 under an open license with weights on Hugging Face. The release arrives the same week that Microsoft's Build 2026 featured expanded physical AI tooling and NVIDIA's GTC Taipei keynote highlighted agentic systems for robotics applications.