Google Signs Pentagon AI Deal as Copilot Kills Flat Rates

A publication called The Wire by Acutus deployed AI-generated reporters to solicit quotes from real advocacy groups and nonprofits, publishing nearly 100 articles through a fully automated pipeline. The Midas Project's investigation found the site drafted stories, reviewed them with an automated system, then sent bots posing as named journalists to conduct written Q&A interviews. 42 articles went live that the site's own reviewer flagged as not ready.

The terminal-based development environment open-sourced its client codebase under AGPL v3. OpenAI is the founding sponsor, with agentic workflows powered by GPT models through Warp's Oz orchestration platform. The thesis: community contributors supervising coding agents will ship faster than an internal team alone. Warp is multi-model and supports Claude Code, Codex, and Gemini CLI alongside its built-in agent.

Google agreed to let the Department of Defense use its AI models for "any lawful government purpose" in classified settings. The contract gives Google no right to veto government decisions. Over 600 employees signed a letter urging CEO Sundar Pichai to reject the deal, citing concerns about autonomous weapons and surveillance. Anthropic refused the same terms and was blacklisted by the Trump administration.

DeepSeek announced a 75% promotional discount on V4-Pro through May 5 and slashed input cache costs across its entire API to one-tenth of previous levels. Even at full price, V4-Pro already undercuts GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro on per-token costs. The move intensifies a pricing war that DeepSeek started with R1 in January 2025.

Nine new connectors let Claude work alongside professional creative software. Adobe integration covers 50+ Creative Cloud apps including Photoshop and Premiere. Blender gets a natural-language interface to its Python API for scene analysis and scripting. Autodesk Fusion lets designers create and modify 3D models through conversation. Ableton and Splice bring music production workflows into Claude, while Resolume gives live visual artists natural-language control over performances.

HubSpot moved from code completion copilots to custom agents to a unified internal AI platform over three years. The results: 100% of engineers using AI tools, 73% increase in lines of code written, 90% reduction in pull request feedback time, and 60% engineering velocity improvement. The key decision was building a shared AI foundation rather than adopting point solutions, so every new capability compounds across the product.

Structured-Prompt-Driven Development keeps prompts in version control alongside code, uses a REASONS Canvas to structure them around business intent, and enforces a rule: when reality diverges from the prompt, fix the prompt first. Thoughtworks' internal IT team found that developers need three key skills to make this work: alignment with business goals, abstraction-first thinking, and iterative review of generated output.

Wiz Research discovered CVE-2026-3854, an injection flaw in GitHub's internal git protocol that let any authenticated user execute arbitrary commands on backend servers with a single git push. On GitHub.com, the vulnerability exposed millions of public and private repositories on affected storage nodes. GitHub patched the cloud service within six hours, but 88% of Enterprise Server instances remain unpatched. Notably, this is one of the first critical vulnerabilities in closed-source binaries discovered using AI.

Stripe's Radar evaluates over 1,000 signals per transaction, reaching the correct verdict 99.9% of the time on billions of payments. The architecture was overhauled multiple times, including removing an XGBoost component that was actively improving accuracy because it blocked everything else the team wanted to build. The replacement, a "Network-in-Neuron" approach called Shield NeXt, cut training time while giving the system flexibility to incorporate new signal types without full retraining.

Eran Sandler tested wrapping single agent turns in Anthropic's Batch API and found 90-120 second delays per interaction. The economics only flip at fleet scale, where pooling requests across agents masks the latency. The counterintuitive finding: expensive models like Opus benefit most from batching because the absolute savings are larger, while Haiku batches paradoxically take longer to process since the scheduler has fewer idle windows.

ChinaTalk dismantles Jensen Huang's claim that China can match US AI capability by assembling more chips. Training and inference at the frontier demand quality compute: high numerical precision, memory bandwidth, and fast interconnects between GPUs. A large cluster of weaker chips cannot converge a frontier model. The analysis estimates China has roughly 2.5 to 2.8 million H100-equivalents, but the aggregate number obscures the fact that most of that hardware cannot run frontier workloads.

AI-generated code may lack copyright protection without meaningful human authorship, a position the US Copyright Office has reinforced multiple times. Employment contracts at most companies already assign AI-assisted work product to the employer, often without the developer realising. And models trained on GPL-licensed code can quietly introduce copyleft obligations into proprietary codebases. The practical advice: run licence scans, document human contributions, and read your employment agreement before assuming you own anything.

GitHub Copilot will shift all plans to usage-based pricing on June 1, ending flat-rate subscriptions that Microsoft could no longer afford to subsidise. Ed Zitron argues this confirms what the numbers always showed: agentic workloads cost far more than the subscription covered, and Microsoft absorbed the difference to inflate adoption figures. The piece frames Copilot's pricing shift as a preview of what's coming across the industry as AI companies confront the gap between growth metrics and unit economics.

Workday approaches $10 billion in annual revenue with near-perfect retention, not because the product is loved but because leaving is nearly impossible. a16z argues HCM is the last large enterprise category without a serious AI-native competitor. The platform shift from cloud to AI is more transformational than the one that created Workday in 2005, and agent-native architectures will reset the category again.

Pick almost any supply chain incident from the past 18 months and trace it back: you end up reading a GitHub Actions YAML file. The piece catalogues how Ultralytics shipped a crypto miner, tj-actions leaked secrets from 23,000 repos, and Trivy was compromised twice in three weeks, all through Actions features behaving exactly as documented. The core problem is that defaults designed for private enterprise repos were never rethought for anonymous forks and drive-by pull requests.

Tanay Jaipuria lays out when AI application companies should invest in post-training custom models and when they should wait. Three signals justify the investment: improving unit economics at scale, proprietary data competitors cannot access, and specialised pipeline components that frontier labs ignore. The core risk is that new base models arrive weeks apart and can instantly obsolete fine-tuning work. Build data infrastructure now, but train only after you have real scale.

YepAPI released 110 open-source agent skills as markdown context files that work across Cursor, Claude Code, Gemini CLI, Copilot, and 40+ other coding tools. Each skill gives the AI specific knowledge about a framework, API, or workflow pattern rather than letting it guess. Categories span security, UI/UX, infrastructure, data analytics, and framework-specific rules. One install command adds all 110.

An open-source frontend for the ACE-Step 1.5 music generation model that runs entirely on your machine with no subscription, usage limits, or cloud dependency. The project hit 1,600 stars in its first week on GitHub and is currently trending daily. Built in JavaScript with a professional interface for prompt-based music creation, it positions itself as a direct replacement for paid services like Suno.

A new reference implementation lets developers add voice interaction to React apps using the OpenAI Realtime API. defineVoiceTool() converts app actions into voice-callable functions, while a controller pattern manages sessions, transport, and tool execution. The package keeps the app as the source of truth: tools remain app-owned, the UI handles all visible state changes, and voice acts purely as an input layer. Apache 2.0 licensed, 725 stars already.