Zuckerberg: Teams of 50 Now Need 10. Anthropic Eyes $900B.

Microsoft reported $37 billion in annualised AI revenue for Q3 2026, more than doubling the prior year. Azure grew 40%, and the company raised its calendar-year capex forecast to $190 billion. An amended OpenAI partnership gives Microsoft royalty-free access to frontier models, cementing the relationship as both companies restructure their deal. The CFO attributed $25 billion of the capex increase to rising memory chip prices.

Samsung's first-quarter net profit reached $30 billion, with 94% from semiconductors. SK Hynix and Micron posted similarly record results. Memory prices jumped nearly 100% quarter over quarter as AI inference demand consumed both HBM and conventional DRAM supply. The three firms are collectively projected to earn $350 billion in 2026, with Samsung vaulting to the world's second most profitable company behind Nvidia.

Anthropic is in early talks for a funding round that would value it above $900 billion, per Bloomberg. The company previously turned down multiple offers at $800 billion. Google recently committed $10 billion at the older $350 billion valuation, with up to $30 billion more tied to performance targets. Anthropic is also weighing an IPO starting in October.

Mistral released Medium 3.5, a 128B dense model with a 256K context window, as open weights under a modified MIT licence. The model powers new remote coding agents in Vibe that run in the cloud and notify you when they finish. A new Work mode in Le Chat extends this with a multi-step agent for research and cross-tool tasks. It runs self-hosted on as few as four GPUs.

Zuckerberg told Meta employees that AI infrastructure costs are forcing headcount cuts. "We have two cost centres: compute and people. If we invest more in one, we have less for the other." He described teams that once needed 50 to 100 people now needing 10, and floated plans to build 50 new apps with the freed capacity. Meta's chief people officer did not rule out further layoffs.

Norwegian robotics firm 1X started production of its NEO humanoid at a 58,000 square foot facility in Hayward, California. The factory can produce 10,000 robots annually, scaling to 100,000 by 2027. First-year capacity sold out within five days of the October launch. Each robot runs on NVIDIA's Jetson Thor for real-time AI inference and is designed for household tasks like mobility support.

Anthropic researchers introduced introspection adapters, a technique for training a single LoRA adapter that causes fine-tuned models to state what behaviours they acquired. The adapter generalises across models fine-tuned in very different ways, achieving state-of-the-art results on an existing auditing benchmark. It also detects encrypted fine-tuning API attacks where adversaries smuggle hidden behaviours through the training pipeline.

The Shepherd Model Gateway disaggregates tokenisation, detokenisation, and request routing from GPU inference entirely. The problem: Python's GIL creates a single-threaded ceiling on CPU-bound work that sits in the serving path. At large-scale disaggregated serving, GPUs worth hundreds of thousands of dollars sit idle waiting for input. SMG replaces the entire CPU layer with pure Rust, and the PyTorch team published benchmarks showing significant throughput gains.

CVE-2026-31431 targets a 2017 optimisation in the kernel crypto API that lets an unprivileged local user corrupt page-cache pages and escalate to root. The same binary works unmodified on Ubuntu, Fedora, RHEL, and SUSE. Shared dev boxes, CI runners, and multi-tenant containers are highest risk. Patch your kernel or disable the algif_aead module. The vulnerability was surfaced by Xint Code in about an hour of automated scanning.

Ryan Law at Ahrefs chained 23 custom Claude Code skill files to generate publish-ready blog drafts. The system mirrors human editorial workflow: research, outline, draft, each step saved as an artefact for debugging. They have published around 15 articles and updated 30 more with the pipeline. The key insight is encoding editorial expertise into skill files rather than relying on generic prompting.

Kent C. Dodds asks what remains valuable one step before AI takes over all implementation. Not typing code, not choosing libraries, not designing the architecture. The last valuable thing is judgement: deciding what is worth building, what constraints must hold, and what trade-offs are acceptable. He frames this "product engineering" as the most durable skill and the hardest for AI to replicate.

Cursor was the fastest-growing software business in history: $2 billion ARR in 13 months, 70% of the Fortune 1000. Its founders looked at the path to $100 billion independent and decided they would not underwrite it. The signal: Anthropic does not tolerate middlemen. Every model lab whose customers are reachable through someone else's distribution ends up watching that intermediary own the user relationship and route to the cheapest model.

Sarah is the person who knows why a service was split, why an abstraction exists, and why the obvious fix was rejected three quarters ago. When agents work through a codebase without that context, they inherit the output without the reasoning. Documentation is no longer a courtesy to future humans. It is the primary way to ensure durable context survives the shift to agent-driven development.

The Zig project has one of the strictest anti-LLM policies in open source. The rationale: reviewing PRs is an investment in individual contributors, not just their code. Bun, which runs on a Zig fork, recently achieved a 4x compile performance improvement but cannot upstream it because the patch involved LLM assistance. Simon Willison calls this the best articulation yet for a blanket ban on AI-assisted contributions.

A deeply reported NYT investigation finds near-universal agreement inside AI labs that advanced AI will displace millions of jobs. Anthropic's Dario Amodei predicts 50% of entry-level white-collar roles gone by 2030. OpenAI published a policy paper proposing a 32-hour workweek and public wealth fund. Anthropic's own engineers run multiple Claude Code agents overnight while privately worrying about the world they are building.

The Cursor SDK exposes the same runtime and models that power the Cursor desktop app, now accessible with a few lines of TypeScript. Agents can run locally or on Cursor's cloud against a dedicated VM with sandboxing, repo cloning, and a fully configured dev environment. The SDK targets CI/CD integration, workflow automation, and embedding coding agents into products. Public beta, available via npm.

Stripe's Link CLI lets agents grab one-time-use payment credentials from a Link wallet without storing real card details. The tool handles authentication, payment method listing, and spend requests with user approval via push notification. It also runs as an MCP server for integration with coding assistants. This is infrastructure for the emerging pattern of agents that need to make purchases on a user's behalf.

Browserbase released a JavaScript SDK that gives Claude agents web browsing capabilities through a standardised skills interface. The project reached 793 stars on GitHub and is trending daily. It fills a practical gap in the agent toolchain: letting Claude-based agents fetch and interact with live web pages rather than relying on static context or pre-fetched data.