Right. I’ve sorted the useful from the ornamental. A regrettably necessary service.
Andrej Karpathy has joined Anthropic to work on pre-training under Nick Joseph, and Anthropic told TechCrunch he will start a team focused on using Claude to accelerate pre-training research. This is not celebrity hiring for its own sake; pre-training is where the frontier-model bill is paid, and Anthropic is betting that AI-assisted research can improve the machinery that makes Claude. Recursive enough to be useful, which is the only acceptable kind. TechCrunch Karpathy
Claude Code v2.1.144 added /resume support for background sessions, elapsed-duration notifications, session-scoped model switching, and reliability fixes around plugins, MCP pagination, protected macOS folders, and terminal rendering. For developers actually using coding agents as background workers, reliability and resumability matter more than yet another theatrical demo. Autonomous coding work is only glamorous until the session disappears into the machinery. GitHub Anthropic Docs
Anthropic acquired Stainless, the SDK and MCP tooling company behind Anthropic’s official SDKs and tooling used by OpenAI, Google, Replicate, Runway, and Cloudflare. Stainless says it will wind down its hosted products, which means a useful neutral layer in AI developer infrastructure is being folded into one frontier lab. A tidy reminder that “ecosystem” often means “available for acquisition.” Anthropic Stainless
Cursor released Composer 2.5, positioning it as a stronger fast-tier coding model for long-running tasks, complex instruction following, and collaborative agent work. If Cursor is already your agentic IDE, this is the sort of default-model upgrade that changes the texture of daily use rather than merely the changelog. For teams deciding where agentic development actually happens, the default model inside the IDE matters more than another splendid benchmark poster. Cursor Cursor Blog
KPMG and Anthropic announced a global alliance that embeds Claude into KPMG’s Digital Gateway client-delivery platform and rolls it out across KPMG’s 276,000-plus workforce. Consultancies are becoming distribution channels, training layers, and implementation arms for frontier models, because apparently software adoption now requires both a model card and a professional-services empire. Anthropic KPMG
GitHub added a “Fix with Copilot” button to failed GitHub Actions workflow logs for Copilot Business and Enterprise users. The cloud agent investigates the failure, pushes a fix to the branch, and tags the developer for review, which is how CI maintenance becomes background work instead of ritual punishment. GitHub
Cognition launched Devin Auto-Triage, which monitors Slack, Linear, GitHub, Sentry, Datadog, webhooks, and similar sources, then investigates issues and can summarize findings, tag owners, or open a PR. Coding agents are creeping toward the interrupt layer of engineering, which is exactly where a great deal of productivity goes to die. Cognition
Zoom expanded its MCP Server so third-party AI tools can access meeting summaries, transcripts, recordings, notes, action items, collaboration history, and connected enterprise context. It includes a Codex plugin and Claude support, which means meeting decisions can finally become usable input for developer workflows instead of archaeology. Zoom
Decart raised $300 million at a nearly $4 billion valuation, backed by Radical Ventures, NVIDIA, Adobe Ventures, Toyota Ventures, eBay Ventures, Sequoia, and Benchmark. The company is selling both real-time world models and an optimization stack that runs across NVIDIA GPUs, Google TPUs, and Amazon Trainium, which is a fairly efficient way to tell the market you intend to sit near several expensive bottlenecks. Decart SiliconANGLE
BBC Panorama found dozens of Facebook and Instagram accounts using AI-generated anti-immigration videos while posing as patriotic UK voices, with operators linked to Sri Lanka, Vietnam, the Maldives, the U.S., Europe, Iran, and the UAE. Cheap synthetic media has made identity-based influence operations easier to industrialize. Democracy, as ever, has been given another maintenance ticket. BBC BBC Sounds
OpenAI previewed a personal-finance experience for ChatGPT Pro users in the U.S., letting users connect accounts through Plaid and ask questions grounded in spending, investments, subscriptions, payments, and balances. It pushes ChatGPT further into sensitive, action-oriented workflows where accuracy and trust stop being decorative values and become product requirements. Do try not to let the chatbot become your wealth manager. OpenAI TechCrunch
OpenAI put Codex into the ChatGPT mobile app in preview, letting developers monitor threads, approve commands, review diffs, change models, and start new work from iOS or Android. Remote SSH and Hooks are also generally available, which makes Codex less like a terminal toy and more like an agent you supervise from wherever trouble happens to find you. OpenAI
Amazon Bedrock introduced Advanced Prompt Optimization, with cross-model prompt comparison, evaluation metrics, Lambda-based scoring, LLM-as-judge rubrics, steering criteria, and multimodal inputs. Prompt migration is becoming regression testing with a user interface. Civilisation advances in small, slightly dull steps. AWS AWS Blog
GitHub launched the Copilot app in technical preview, a desktop workbench for agentic development where sessions can start from issues, pull requests, prompts, or prior work, then run in separate branches with terminal and browser validation. Copilot is becoming a place where development work lives, not just something that completes a line while you wonder whether you trust it. GitHub
xAI launched Grok Build in early beta for SuperGrok Heavy subscribers, with plan-review-approve workflows, clean diffs, AGENTS.md support, plugins, hooks, skills, MCP servers, parallel subagents, worktree integration, headless mode, and ACP support. Another terminal agent enters the room; do try not to trip over the subagents. xAI
Cursor 3.4 added multi-repo cloud-agent environments, Dockerfile-based configuration, build secrets, layer caching, validation, audit logs, environment-scoped secrets, and egress controls. The message is refreshingly adult: if agents are going to work on real repositories, they need real environments and real governance. Cursor
Cerebras priced its IPO above range, raising $5.55 billion and valuing the company at $56.4 billion on a fully diluted basis. Public markets now have a rare pure-play AI hardware bet that is not simply “buy more Nvidia and pray the grid holds.” CNBC
Anthropic launched Claude for Small Business, with connectors and ready-to-run workflows across QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365, finance, operations, sales, marketing, HR, and customer service. The practical shift is packaged operational automation for companies without the budget to invent an AI transformation office, which is most of them. Anthropic
Notion launched a Developer Platform with Workers, database sync, custom agent tools, an External Agent API, and a CLI for humans and coding agents. It turns Notion into more of an agent workspace: synced business data, automations, and external agents all meeting in the same place, hopefully by appointment. Notion TechCrunch
Google announced Gemini Intelligence for Android, Gemini in Chrome for Android, and the Googlebook laptop concept, all pointing in the same direction: AI as a system layer rather than another app icon. The useful bits are proactive task automation, contextual browsing help, natural-language widgets, and tighter phone-laptop continuity. Google has noticed that assistants are more useful when they are near the work. Startling discovery. Google Google Chrome
AWS Security Agent now supports full-repository code reviews in preview, reasoning about architecture, trust boundaries, and data flows rather than only matching vulnerability patterns. It is aimed at moving security review earlier in AI-accelerated development, where “we’ll check it later” has historically done such fine work. AWS
RSL Media launched the Human Consent Standard, a way for people and rights holders to declare terms for AI use of their likenesses, voices, characters, creative works, and brands. It has backing from major actors, directors, agencies, and studios, and is more constructive than another open letter, which I realise is a low bar but still worth clearing. The Verge RSL Media
That will do. Try to brief someone before they mistake the longest thread for the most important one.