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

Curated from 148 tweets liked by @mattsilv, May 3 - May 11, 2026:

AI for Everyone

GPT-5.5 Instant rolls out as the default for every ChatGPT user. Available in the API as gpt-5.5-chat-latest. Factuality gains in medicine, law, and finance, plus a new "memory sources" panel that shows what context ChatGPT pulled from your chats, files, and Gmail. Full-duplex voice mode coming. Read more →

Coinbase cuts 14% and says the quiet part out loud. Brian Armstrong's email blames it directly on AI productivity, not market conditions. Coinbase is moving to AI-native pods, one-person teams, 5 org layers max, and every leader as an individual contributor. Read more →

Google launches an AI Health Coach, a $99 screenless Fitbit, and a 14,000-patient study. Fitbit app rebranding as Google Health, Gemini coach connects to U.S. medical records. The study found structured AI symptom interviews beat passive symptom entry by 27% in diagnostic accuracy. Read more →

Kevin Rose quietly relaunches Digg as an AI news aggregator. Live alpha at di.gg with 9M graph connections, 15+ AI judges, real-time X ingestion, influence-flow tracking. Matt Van Horn already shipped a Digg CLI with Claude Code + Hermes skills. Read more →

Apple's camera AirPods hit late-stage testing. Gurman confirmed both earbuds carry cameras capturing low-resolution visual data for Siri. AirPods Pro 3-style design, September target depends on the Gemini-powered Siri rebuild hitting Apple's quality bar. Read more →

Anthropic and Wall Street form a $1.5B joint venture. Blackstone, Goldman, Hellman & Friedman are partners; targets private-equity-owned portfolio companies. The JV is supposed to actually rebuild workflows, not resell API access. Read more →

Anthropic co-founder Jack Clark puts 60% odds on recursive self-improvement by end of 2028. Based on reading hundreds of public AI development sources. One tweet, but an unusual source: he's on the inside and not a hype account. Treat as a planning input. Read more →

AI for Developers

Claude Managed Agents get Dreaming, Outcomes, and multi-agent orchestration. Dreaming reviews past agent sessions to curate memories — Harvey reported ~6x completion improvement (research preview, waitlisted). Outcomes runs a rubric-driven self-grading loop until it clears your bar (public beta, +10pp task success, Wisedocs reports 50% faster reviews). Multi-agent orchestration runs specialists in parallel on shared filesystem. Read more →

Claude Code rate limits doubled, SpaceX compute deal announced. Pro/Max/Team plans got 2x the 5-hour limit, peak-hours throttle removed, Opus API limits raised. Claude Code is up 15x since January 1. Read more →

OpenRouter ships free response caching with one header. Add X-OpenRouter-Cache: true and identical calls return in 80-300ms at zero token cost. Best for agent retries and test suites. Works alongside prompt caching. Read more →

Voice AI convergence day (May 7): OpenRouter Audio APIs, OpenAI GPT-Realtime-2, ElevenLabs cuts prices 55%. OpenRouter now routes TTS/STT across OpenAI, Google, Mistral, Whisper, Chirp 3, and Groq. GPT-Realtime-2 brings GPT-5-class reasoning to voice (instruction retention 36.7% → 70.8% APR). Read more →

Firefox fixed 271 security bugs in April using Claude Mythos. 423 total bugs shipped (vs. the normal 20-30/month baseline) including a 20-year-old XSLT bug. Mozilla built an agentic harness on existing fuzzing infrastructure; Claude created reproducible test cases to validate hypothesized vulnerabilities. Read more →

Codex ships a Chrome extension for background tab work. Per-site permissions, cross-tab debugging, parallel DevTools, no browser takeover. The right answer for browser agent work. Read more →

Pareto Code routes your coding calls to the cheapest capable model. OpenRouter's free experimental router: set min_coding_score, get the cheapest model that clears your bar. DeepSeek V4 Pro, GPT-5.4 Mini, and Gemini 3.1 Pro are at the top right now. Read more →

Hermes Agent v0.13 ships with Autobrowse that self-optimizes 80% cheaper. Demo: browser automation task dropped from 102s to 35s, $1.46 to $0.28 in two iterations, by switching from step-by-step clicking to direct JS eval. Still very low-level; Claude Cowork is the friendlier starting point. Read more →

Honorable Mentions

For everyone:

• Claude for Excel, Word, and PowerPoint hit general availability, with Claude for Outlook in public beta. Context carries across Microsoft apps. (source)
• NotebookLM Mind Maps got prompt-driven steering, renaming, and sharing. Scope a map to a specific topic instead of letting it auto-generate. (source)
• iOS 27 will let you pick Claude or Gemini instead of ChatGPT for Apple Intelligence. Opens the current OpenAI-only escalation path. (source)
• Google Finance Beta added AI-powered key moments that explain stock price swings and jump you to the relevant part of the earnings call. (source)
• Gemini agent for macOS in development per 9to5Google. Uses Screen Access to organize files, convert to Sheets, batch-rename, draft email summaries from meeting transcripts. (source)
• A Gemini "Omni" video model leak surfaced today. First output claimed by @chetaslua; @kimmonismus speculates it ships at Google I/O on May 19, possibly as Veo 3.1's successor. (source)

For developers:

• Gemini 3.1 Flash Lite went GA on May 7 at $0.25/M input, $1.50/M output. Gemini 3.2 Flash also briefly visible before Google I/O on May 19. (source)
• Grok 4.3 launched on the xAI API at $1.25/M input, $2.50/M output, 1M context. Tops Artificial Analysis on instruction following and Vals AI on case law + corporate finance. (source)
• Anthropic's leaked "Orbit" feature for Claude Cowork would connect Gmail/Slack/GitHub/Calendar/Drive/Figma and generate proactive briefings. Dev gate "tibro enabled" (orbit backwards). (source)
• OpenRouter's GPT-5.5 cost analysis: short-prompt workloads cost ~92% more on GPT-5.5 vs GPT-5.4; 49-69% more for medium/long. (source)
• Claude Platform now supports keyless auth via AWS, GCP, or Azure cloud identity. Anthropic's most-requested security feature. (source)
• Codex "Ultrafast mode" briefly appeared in the GitHub repo before being deleted: "The fastest available responses for latency-sensitive work." Unintended push. (source)
• OpenAI's October 2025 employee secondary stock sale totaled $6.6B, per WSJ. 600+ current and former employees, $11M average per person. (source)

Try This Weekend

For everyone: Open ChatGPT and look at memory sources the next time you ask something personal, try Google Finance Beta on a stock you actually follow to see the AI key moments annotations, steer a NotebookLM Mind Map with a prompt like "focus on the risks", or bookmark Digg's AI alpha and see if it beats your current habit of scrolling X.

For developers: Add X-OpenRouter-Cache: true to your existing OpenRouter test suite, drop Pareto Code into one call by setting min_coding_score, install the Codex Chrome extension for background-tab work, or set up one cron job that runs Claude Code on a recurring chore (Boris Cherny's PR babysitter pattern).

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.