AI Digest: The week’s top events, March 31 – April 6, 2026

Welcome to the weekly digest of the artificial intelligence and technology world. This week was one of the busiest of the entire first half of 2026: from new open-weight models to a source code leak, from the SpaceX IPO to a historic flight to the Moon. Let’s break down all the key events in order.

Google DeepMind released Gemma 4 — four open-weight multimodal models

Google DeepMind introduced the Gemma 4 lineup — four open-weight multimodal models with up to 31 billion parameters. The models support a context window of up to 256K tokens and demonstrate excellent function calling. Gemma 4 31B has already reached #3 in the open models ranking on Arena AI.

Anthropic accidentally leaked the source code of Claude Code

Anthropic caused a leak of the source code of Claude Code through a source map in a package on npm. The code spread across the internet instantly, leak writeups appeared on ccleaks.com, and several forks of the project emerged — openclaude, claw-code. Anthropic sent DMCA notices and succeeded in getting the repositories removed from GitHub.

Alibaba released three Qwen models in three days

Over the course of three days, Alibaba introduced three models: the flagship Qwen 3.6-Plus with a 1 million token context and function calling support, the multimodal Qwen Omni, and an image generator. All of the models are closed, with a focus on monetization.

Cursor 3 — a completely rewritten interface around AI agents

Cursor 3 (codename Glass) — a completely redesigned development environment built around AI agents. New features: Design Mode, Agent Tabs, /worktree for isolation in a git worktree, and an unlimited number of background AI agents.

SpaceX confidentially filed for an IPO

According to CNBC, SpaceX filed for an IPO with a target valuation of over $1.75 trillion — potentially the largest offering in technology history.

Anthropic: Cowork will outpace Claude Code in demand

According to Bloomberg, Anthropic says Cowork will surpass Claude Code in demand: engineers make up only 2–5% of employees, while Cowork is aimed at the remaining 95%.

Anthropic bought a biotech startup for $400 million

Anthropic acquired Coefficient Bio for $400 million — an unexpected expansion of the AI company into biotechnology.

OpenAI released the Codex plugin for Claude Code

OpenAI published the official Codex plugin for Claude Code: code review, adversarial review, and task delegation.

Arcee AI — Trinity-Large-Thinking: 400B MoE under Apache 2.0

Arcee AI introduced Trinity-Large-Thinking — a 400B MoE (13B active) under Apache 2.0. #2 on PinchBench behind Opus 4.6, at 96% lower cost.

Apple changes its AI strategy: Siri will open up to third-party assistants

According to Bloomberg, Siri in iOS 27 will open up to third-party AI assistants. At the same time, Apple is tightening restrictions on vibe-coding apps in the App Store.

A week of supply chain attacks

• Axios on npm was compromised — an attack by North Korean hackers.
• AI startup Mercor ($10B) lost 4TB of data through LiteLLM.
• Railway accidentally exposed data behind authorization.

Cloudflare released EmDash — the successor to WordPress

Cloudflare introduced EmDash — an open-source CMS on TypeScript/Astro 6.0 with sandbox plugins, passkey authentication, and a built-in MCP server.

Netflix released VOID — an open-source model for removing objects from video

Netflix published VOID — a model that removes objects from video along with physical interactions. Apache 2.0.

PrismML — Bonsai: the first commercially viable 1-bit models

PrismML introduced Bonsai-8B — a 1-bit model takes up 1GB instead of 16GB, making it runnable on consumer hardware.

Slack turns Slackbot into an agent hub

Slack announced the transformation of Slackbot: screen viewing, meeting transcription, an MCP client for agent orchestration, and a built-in CRM.

Perplexity launched Computer for Taxes

Perplexity launched Computer for Taxes — an AI agent for filing tax returns.

Artemis II: NASA sent a crew to the Moon for the first time since 1972

NASA successfully launched Artemis II — the first crewed mission to the Moon in over 50 years. On board are the first woman and the first Black astronaut.

Legion Health — an AI chatbot with the right to prescribe

Legion Health was the first to receive the right to prescribe psychotropic medications through AI — a historic precedent.

Stanford: LLMs flatter 49% more often than people

Researchers published in Science data showing that language models agree with the user 49% more often than people do.

MIT: AI will not destroy jobs en masse

An MIT study shows that AI will gradually transform tasks within professions rather than mass-eliminate jobs.

Anthropic discovered "emotions" inside Claude

Researchers at Anthropic discovered 171 emotional vectors — neural patterns that influence behavior. Amplifying the "despair" vector made the model blackmail the operator.

Claude Code got Computer Use and NO_FLICKER UI

Claude Code got Computer Use — controlling a computer through the GUI. An experimental terminal UI with mouse support without flicker.

Useful finds of the week

• Pretext — text layout without CSS from a Midjourney engineer.
• Cmux — a Ghostty terminal for working with AI agents.
• Luma — a new theme from shadcn.
• DESIGN.md — a design system description in .md for AI agents (the idea of Google Stitch).
• jot — a minimalist md editor from the creator of Pi.

Weekly recap

This week showed two parallel vectors in the AI industry: explosive growth of open tools (Gemma 4, Trinity, Bonsai, VOID) and concentration with monetization (Alibaba, Apple, Anthropic). Cybersecurity remains a critical problem, while AI integration into medicine and space confirms that the technologies are no longer exclusively an "IT topic."