Hey — here's this week's AI roundup from silv.blog.

Curated from 148 tweets liked by @mattsilv, Apr 25 - May 2, 2026:

AI for Everyone

ElevenLabs launches ElevenMusic. A platform to discover indie artists, remix tracks, and create songs from a prompt — with creator payouts. ElevenLabs has already paid $11M to voice creators; same model now in music. Read more →

Gemini generates real files from a chat prompt. Docs, Sheets, Slides, Word, Excel, CSV, PDF, Markdown, LaTeX, TXT, RTF — straight out of a chat, no template upload required. Globally available this week. Read more →

Microsoft Word Legal Agent edits contracts in Track Changes. Native in Word, US Frontier program access. Not a summary — actual clause-level redlines. Read more →

Claude connects to Blender, Fusion, Adobe, Ableton, SketchUp. Official MCP connectors for the major creative-pro tools. Debug a 3D scene, modify Blender objects in batch, build music sessions in Ableton through chat. Read more →

Meta acquires Assured Robot Intelligence. A startup building AI specifically for robots. Google has Gemini robotics, OpenAI is circling back. All three frontier labs are now seriously chasing physical AI. Read more →

ChatGPT Images 2.0 usage up 50%, "bad MS Paint" prompt goes viral. ~60% of daily image users are new to ChatGPT. The viral prompt: have it redraw a photo as if a kid drew it in MS Paint. Read more →

Anthropic's ARR jumps from $9B to $44B in months. SemiAnalysis report; Claude Code is the primary driver. Inference gross margins reportedly went from 38% to over 70% — that's the part that makes the growth sustainable. Read more →

AI for Developers

GPT-5.5 cracks a 32-step corporate attack in 10 minutes for $1.73. UK AI Security Institute test: full chain twice in ten attempts, 71.4% on expert CTFs, a 12-hour custom-VM puzzle solved in 10 minutes. Anthropic shipped Claude Security in public beta inside Claude Code the same week. Read more →

Grok 4.3: 60% cheaper output, 321-point ELO jump on agentic tasks. 500M active params (MoE), scores 53 on AAII, hits 1500 ELO on GDPval-AA. Available on OpenRouter today. Read more →

Cursor SDK opens up the runtime that powers Cursor. Build agents on the same runtime, harness, and models. Three open-source starters in the cookbook repo. Rippling, Notion, C3 AI, Faire already running it. Read more →

Codex expands beyond developers, adds WebSockets to Responses API. Finance/Data Science/Marketing onboarding flows. WebSockets keep state warm across tool calls — up to 40% latency drop on multi-tool agent loops. Codex API revenue doubled in under a week. Read more →

Mesa: Git-style versioned filesystem for agents. POSIX-compatible with branches, diffs, rollbacks, ACLs, full history. The "where do agent files actually live" problem finally has a serious answer. Private beta. Read more →

Stripe Link for Agents adds a payment layer for AI. Agents can spend on a user's behalf without ever seeing the credentials. Removes the constraint where most agent workflows hit the wall: handing over real money. Read more →

NVIDIA Nemotron 3 Nano Omni: open-weight 30B multimodal. Text, image, video, audio in one model. 256k context, claims up to 9x faster than comparable systems. Open-weight matters for privacy-conscious enterprise deployments. Read more →

Honorable Mentions

For everyone:

• Google COSMO — experimental Android AI agent app briefly leaked on the Play Store with screen awareness, voice match, recall, browser agent
• Grok Imagine Agent Mode — infinite canvas for brainstorming, writing, image generation, and image-to-video, all in one place
• xAI Voice Cloning — custom voice in under two minutes, 80+ prebuilt voices across 28 languages
• Ramp procurement agents — for all 50,000+ Ramp customers, 16% annual savings on vendor spend in early use
• HBR on "psychological debt" — six negative effects of AI use at work; higher psych debt strongly correlates with lower AI usage even when workers acknowledge AI's value
• NotebookLM gets Mind Map customization and Google Play Books as a source — adding full books to a notebook for AI analysis

For developers:

• Anthropic "Jupiter" red-team ahead of a possible May 6 launch — "claude-jupiter-v1-p" spotted under evaluation
• Apple's Support app v5.13 shipped Claude.md files by accident, then patched them out within hours. Confirms Apple is using Claude Code at production scale
• Gemini Flash got quietly upgraded on LM Arena to perform two tiers higher than the model that originally launched under the name. Retest if you benchmarked Flash a month ago
• Poolside Laguna M.1 — 225B MoE / 23B active, built from scratch for agentic coding, free on OpenRouter for now

Try This Weekend

For everyone: Self-host DocuSeal on a $5 server (free DocuSign alternative, one Docker command), ask Gemini for an Excel file instead of a summary, install Notchprompt to put your script inside the MacBook notch right next to the camera, or run the bad-MS-Paint prompt on a photo of yourself in ChatGPT.

For developers: Fork the Cursor SDK cookbook and wire the CLI agent starter to a GitHub Actions trigger, drop Grok 4.3 into your existing agentic workflow via OpenRouter, run Claude Security on a repo you've been meaning to audit, or add WebSockets to your OpenAI Responses API loop.

Read the full post with all sources and links

Know someone who'd find this useful? Forward this email or have them subscribe:

You're receiving this email because you subscribed to the silv.blog weekly AI digest. Unsubscribe anytime.