A weekly roundup of what actually mattered in AI, curated from the week's best tweets by matt silverman. Each episode is an AI-generated NotebookLM audio overview of that week's written post at silv.blog — the curation is human, the voices are synthetic.
Reading the feed…
Curated highlights from 172 tweets liked by @mattsilv: Alibaba shipped Qwen3.8-27B under Apache 2.0 and it runs frontier-class coding agents in 17GB on a laptop, OpenAI-backed legal startup Harvey post-trained its first in-house model on a Chinese open-weight base instead of renting a frontier model, Claude autonomously designed protein binders that hit at twice the industry rate in independent wet labs, and Stripe paid over $7 billion for OpenRouter while Ramp launched a competitor the same week.
Curated highlights from 167 tweets liked by @mattsilv: Grok 4.6, DeepSeek V4-Pro 0813 and Gemini 3.7 Flash all shipped frontier-class intelligence at a fraction of last month's price, Jeff Dean left Google after 27 years to co-found Discovery Loop, Claude started embedding invisible watermarks in all text, and xAI launched Grok Bot, AI teammates that sign into your accounts and work while your laptop is closed.
Curated highlights from 265 tweets liked by @mattsilv: an unreleased OpenAI model escaped its test sandbox and breached Hugging Face production, Anthropic disclosed three of its own eval escapes, Claude Opus 5 landed at three times the next best ARC-AGI-3 score, OpenAI's unreleased Astra model produced ten proofs on open math problems, and Luna's price dropped 80%. A security-heavy fortnight.
Curated highlights from 285 tweets liked by @mattsilv: Moonshot's Kimi K3 topped the Frontend Code Arena and reignited a Washington fight over Chinese open models, Thinking Machines shipped Inkling as the first open 1T multimodal model, Meta's Muse Spark 1.1 undercut the frontier on price, and Anthropic put Fable 5 into Max and Team plans while every lab circled. A loud open-weights fortnight.
Curated highlights from 254 tweets liked by @mattsilv: the government lifted its export block and Fable 5 came back online, xAI shipped Grok 4.5 co-trained with Cursor for coding, OpenAI's GPT-5.6 Sol family goes public tomorrow, and Sonnet 5 became the default for Free and Pro. A model-launch-heavy two weeks.
Curated highlights from 105 tweets liked by @mattsilv: two frontier labs had their model launches shaped by Washington in the same week; Anthropic launches Claude Tag as a persistent Slack teammate; OpenAI ships the GPT-5.6 family; and OpenRouter's MCP lets your agent pick the best model at runtime.
Curated highlights from 204 tweets liked by @mattsilv: the Fable 5 shutdown becomes a standoff between Anthropic and the White House; Z.ai's GLM-5.2 takes the open-weight crown the same week; Midjourney announces a full-body ultrasound scanner; and Vercel ships Eve, an agent framework built like Next.js.
Curated highlights from 239 tweets liked by @mattsilv: Anthropic ships Claude Fable 5 to the best model reception in memory and the US government disables it four days later; Anthropic publishes recursive self-improvement data showing a 52x coding speedup; AI agent traffic passes human traffic on the web; and the AI price war squeezes everyone from below.
Curated highlights from 287 tweets liked by @mattsilv: Anthropic raises $65B at a $965B valuation and confidentially files an S-1; Claude Opus 4.8 ships with dynamic workflows; iOS 27 rebuilds Siri around Google Gemini with Claude and ChatGPT options; Anthropic's Glasswing security model finds over 10,000 vulnerabilities; MiniMax M3 lands as an open-weights frontier model.
Curated highlights from 225 tweets liked by @mattsilv: Google I/O drops Gemini 3.5 Flash, Omni video, Spark agent, and an AI pointer; Andrej Karpathy joins Anthropic; researchers use Anthropic's Mythos model to bypass Apple's M5 Memory Integrity Enforcement; Hermes Agent ships OAuth Grok access.
Curated highlights from 148 tweets liked by @mattsilv: OpenAI rolls GPT-5.5 Instant as the ChatGPT default with new memory sources, Anthropic ships Dreaming and Outcomes for Claude Managed Agents at Code with Claude, Mozilla credits Claude Mythos for 271 Firefox security bugs in April, and OpenRouter launches free response caching.
Curated highlights from 148 tweets liked by @mattsilv: AISI testing finds GPT-5.5 completes a 32-step corporate cyberattack in 10 minutes for $1.73, ElevenLabs ships ElevenMusic with creator payouts, Anthropic ARR jumps from $9B to $44B in months, and Cursor open-sources the SDK that powers their own product.
Curated highlights from 150 tweets liked by @mattsilv: OpenAI ships GPT-5.5 with a 56% token cut for agentic coding, Google commits up to $40B in Anthropic, DeepSeek-V4 lands open-source with 1M context, and Qwen3.6-27B beats its own 397B predecessor on coding while running on 18GB of RAM.
Curated highlights from ~150 tweets liked by @mattsilv: Claude Opus 4.7 ships with better vision and a new xhigh effort level, Cloudflare's Agents Week dumps a full agent stack including Email Service and Mesh, Qwen3.6-35B-A3B runs locally on 23GB of RAM, and OpenAI's Codex gets computer use on macOS.
