Google DeepMind released Gemma 4 — four vision-capable reasoning models (2B, 4B, 31B, and a 26B-A4B MoE) under Apache 2.0, a major licence upgrade from previous Gemma restrictions.
Built for autonomous agents on mobile, IoT, and personal computers, with native function calling and integrations across Android Studio, LiteRT, Vertex AI, GKE Agent Sandbox, and AICore Developer Preview.
A new Gallery App on the Play Store lets anyone try an agentic Gemma experience offline — signalling Google's push to bring frontier AI to everyday devices.
Microsoft AI (Mustafa Suleyman) launched three in-house foundational models: MAI-Transcribe-1 (beats OpenAI Whisper on 25 languages), MAI-Voice-1 (generates 60 seconds of audio in 1 second), and MAI-Image-2 (top-3 on Arena.ai image benchmarks).
All three were built by teams of under 10 engineers and are available via Microsoft Foundry and the MAI Playground — Suleyman's AI self-sufficiency push paying off.
Already shipping inside Copilot Audio Expressions and Copilot Podcasts, extending the audio AI race that saw Cohere, Mistral, and Google all launch voice models in the past two weeks.
Microsoft released Harrier-OSS-v1, a family of open multilingual text embedding models (270M, 0.6B, 27B) hitting state-of-the-art on Multilingual MTEB v2.
Uses a decoder-only architecture with last-token pooling and L2 normalisation; the 27B variant has 5,376 dimensions and supports 32K context.
Released open-source, competing directly with Google's Gemini Embedding 2 (covered Mar 14) in the multilingual retrieval space.
OpenAI closed a $122 billion funding round, the largest in AI history, to accelerate its next phase of frontier model development and infrastructure buildout.
The raise comes as OpenAI crosses $25B ARR and pivots sharply toward coding tools and enterprise after dropping Sora and the Atlas browser project.
Combined with Oracle's $156B AI capex commitment and Meta's $115B infrastructure spend, it signals a full-scale hyperscaler arms race with no signs of slowing.
Oracle began layoffs affecting thousands of employees on March 31, with restructuring costs up to $2.1 billion — funded by $45-50B in new debt and equity raised in 2026 alone.
The cuts free capital for Oracle's $156B AI data centre commitment, the largest infrastructure bet by any non-hyperscaler in AI history.
Oracle's stock rose 2% on the announcement, reflecting a now-familiar pattern: investors reward AI infrastructure bets over headcount, continuing the trend set by Meta's 15k layoffs (covered Mar 15).
An undisclosed Chinese manufacturer launched the first fully automated humanoid robot production line, capable of producing one robot every 30 minutes — 10,000 per year.
UBTECH, AgiBot, and Unitree are now at similar output rates, signalling a step-change in Chinese humanoid robot manufacturing capacity.
The announcement arrived the same week a US hospital CEO publicly declared readiness to replace radiologists with AI — two signals of AI-driven labour displacement converging.
Security researcher Chaofan Shou found Anthropic's full Claude Code CLI (1,900 files, 512k+ lines) exposed via a 60MB source map file left in the npm package — the second Anthropic security incident in one week after the Claude Mythos CMS leak (covered Mar 28).
Attackers exploited the window within hours, injecting a Remote Access Trojan into malicious axios versions (v1.14.1/0.30.4) in npm between 00:21–03:29 UTC March 31 — anyone who updated Claude Code during that window should rotate credentials immediately.
Analysis of the leaked source revealed an ANTI_DISTILLATION_CC flag that silently injects fake tool definitions into API requests, frustration-detection regexes, and a hidden 'undercover mode' — giving researchers an unprecedented look inside a frontier coding agent.
The CEO of America's largest public hospital system publicly stated he is ready to replace radiologists with AI, citing accuracy, speed, and cost advantages.
The announcement is one of the most direct statements by a major health system executive about near-term AI-driven job displacement in a licensed medical profession.
Coming alongside China's humanoid robot production line announcement, it marks a week where AI labour displacement moved from theory to operational planning.
Garry Tan (President & CEO of Y Combinator) open-sourced gstack — 15 slash-command tools that turn Claude Code into a virtual engineering team: CEO, designer, eng manager, QA lead, security officer, and release engineer.
Tan claims 10,000–20,000 lines of production code per day part-time, with 600k+ lines in 60 days across 3 projects — citing Peter Steinberger's OpenClaw (247k stars) as inspiration.
The repo hit 54k+ stars within days of launch, extending the coding agent tooling wave that includes Everything Claude Code (90k stars, covered Mar 21) and OpenAI Codex Plugins.
OpenAI launched a Plugins system for Codex — bundles of skills, MCP servers, and app integrations that mirror what Claude Code has offered for months, per Ars Technica.
In a striking move, OpenAI simultaneously published an official Claude Code plugin (codex-plugin-cc) allowing Claude Code users to delegate tasks to Codex — the first cross-company coding agent integration.
Together, the two announcements signal that coding agents are converging toward an interoperable ecosystem rather than walled gardens.
NVIDIA Research published PivotRL, a turn-level RL algorithm for post-training LLMs on long-horizon agentic tasks that matches full E2E RL accuracy with 4x fewer rollout turns.
The key insight is Pivot Filtering — selecting only turns with high reward variance and low mean, maximising the training signal per compute dollar.
Already deployed in production powering NVIDIA Nemotron-3-Super-120B-A12B, making this an immediately practical result for teams doing agentic post-training at scale.
Stanford's CS25 — the AI community's most-watched seminar series on Transformers and foundation models — is now fully open to the public for Spring 2026.
Sessions run Thursdays at 4:30pm PDT via Zoom and in-person, with speakers confirmed from OpenAI, Anthropic, Google DeepMind, and NVIDIA.
Free access to Stanford-quality frontier AI education; consistently one of the most-referenced course series on r/MachineLearning.
Context Hub (chub) is a CLI giving coding agents versioned, curated API docs so they stop hallucinating library interfaces mid-task — 7k+ stars on launch.
Agents annotate documentation gaps locally and vote docs up or down, feeding improvements back to maintainers over time.
Works with Claude Code, Codex, and any agent harness as a simple skill or SKILL.md integration — a natural companion to gstack and Everything Claude Code.