Oracle blames AI for 21k cuts 💼, ship the slop, just not always 🧑‍💻, AI's authority is grammar 🤖

Oracle's annual regulatory filing says it cut 21,000 roles over the past year, shrinking headcount from 162,000 to 141,000, and states that deploying AI across its operations has resulted in workforce reductions. That is roughly 13% of staff and about $1.8 billion in restructuring costs. It makes Oracle one of the first blue-chips to put the AI-for-headcount trade in writing, rather than hiding it behind the usual talk of efficiency.

OpenAI expanded its Daybreak security effort with the full GPT-5.5-Cyber model, a Codex Security plugin that scans a codebase and returns severity reports with fixes, and Patch the Planet, a programme built with Trail of Bits and HackerOne. It has already merged dozens of patches across 19 projects including cURL, Python, and Go. The shift is from finding bugs to landing fixes, each one human-reviewed first.

Meta left data from a divisive employee-monitoring programme, keystrokes, mouse clicks, and on-screen content from US staff laptops, accessible to anyone inside the company, per an internal security notice. The data was being collected to train AI models. After Wired began reporting, Meta said it is pausing the collection indefinitely. For anyone wiring workplace telemetry into model training, the exposure surface here is the data lake, not the model.

Cursor has acquired Continue, the open-source coding-assistant alternative, and is winding it down rather than absorbing it. Its site now tells users they have until July 15 to export their data before deletion, and recurring billing is already off. The deal landed as SpaceX confirmed its $60 billion purchase of Cursor, so the rocket company now owns Continue too. If you run Continue, export now and pick a replacement.

Cloudflare announced PACT, or Private Access Control Tokens, a protocol it is developing with Mozilla, Google, Microsoft, and Shopify to verify whether web traffic is a human or an authorised agent without CAPTCHAs, logins, or invasive tracking. Anonymous tokens would prove legitimacy at the edge instead of fingerprinting users. For anyone shipping agents or bots, it signals a coming browser-backed standard for proving sanctioned automated access.

Google moved its Interactions API to general availability and named it the primary way to build on Gemini. The single endpoint folds in server-side state, background execution, tool combination, and multimodal generation, removing orchestration plumbing teams currently hand-roll on top of chat completions. If you build Gemini agents, it is worth evaluating a migration, because server-side state and background runs change how you structure long-running, multi-step work.

Consulting firm Bain is building AI-generated replicas of acquisition targets' software to show buyers how cheaply the technology could be reproduced. What began in 2023 as a dedicated engineering team is now hundreds of rough prototypes built by regular consultants. It is already shaping deals: one private equity investor said a Bain-vibecoded analytics-platform clone contributed to walking away from a bid. Defensibility now sits in data and distribution, not code.

Researchers from NVIDIA, CMU, and Berkeley built ENPIRE, a feedback loop that lets coding agents improve robot policies in the physical world, not just in simulation. The agent resets the scene, executes a policy, verifies the outcome, and refines the next iteration, reaching a 99% success rate on dexterous manipulation. It points at a repeatable harness for robotics research that has long needed heavy human supervision.

Cloudflare's Images service, built in Rust on the hyper HTTP library, began returning 200 responses with the image data silently cut short, but only intermittently and only for larger files. The team spent six weeks tracing it to a race between the writer's buffer flush and connection shutdown on slow readers. The fix was four lines. Suspect the transport layer, not your app, when truncated responses come back as successes.

On 10,000-line AI-generated diffs, Simian Words says, nitpicking individual lines is wasted effort, because the model already clears that bar and catches high-severity issues. Your real value as a reviewer is out-of-distribution knowledge the LLM cannot have: a service being deprecated, a codebase convention, a decision from last week's architecture meeting. Run the model for the mechanical pass and spend attention where only you hold the context.

Tomasz Tunguz breaks down why so many AI companies reselling inference are running a zero-margin payment rail, not a software business. Cost-plus markup caps willingness to pay at the raw API price and compresses toward zero as inference commoditises. Holding 30-plus points of gross margin means charging for the work done, or compounding an optimisation moat that drives delivered cost down over time.

Works in Progress traces the AI build-out's real bottleneck to moving electricity, not generating it. OpenAI and SoftBank's Stargate campus in Abilene alone will draw 1.2 gigawatts at peak, as much as 313,000 homes, with total AI compute projected near 100 gigawatts. The binding constraint is grid interconnection and transmission to where the data centres sit, essential context for modelling the cost and siting of large-scale compute.

Here is the unsettling part: AI systems acquire real organisational authority through the grammar of their outputs, not through accuracy. A line like a budget reduction is required erases the person who decided and reframes a debatable choice as necessity, so people treat it as if a legitimate process already ran. Generated inside the workflow at scale, probability quietly becomes policy. Audit how AI-written reports phrase recommendations.

Doug Slater, no fan of unreviewed AI code, argues the right amount of slop is not zero. His rule is that rigour should be proportional to risk, with narrow cases where shipping LLM-generated code without reading it is defensible. He walks through his own criteria for skipping review and pushes back on both the never-trust-vibe-code camp and the idea that an LLM reviewing its own output is enough.

Inspecting Claude Code's on-disk session logs, the author found the thinking blocks stored as a 600-character signature with no readable text. Per Anthropic's docs, the model encrypts its reasoning into that signature, Anthropic holds the key, and the API returns a summary of the reasoning, not the reasoning itself. The warning: do not promise anyone an audit trail from those logs, because what you can read is a lossy summary.

AWS introduced Lambda MicroVMs, a serverless primitive that gives every user session or job its own VM-isolated environment to run untrusted or AI-generated code, with snapshot-based fast starts and suspend-resume to cut idle cost. It is built on Firecracker, the same virtualisation behind Lambda functions. AWS names AI code-execution sandboxes, interactive dev environments, and agent evaluation as the target cases, the isolation problem agent-code teams keep hitting.

Recall tackles Claude Code's cold-start problem by keeping a local log of your sessions and condensing it into a resume-ready summary, entirely on your machine. The summarisation runs a classical Python algorithm, not an LLM, so persistent memory costs nothing beyond the subscription you already pay for. Each session loads a compact context file of one to two thousand tokens instead of re-explaining the project. Nothing leaves your machine.

Oak is an open-source version control system designed for agents rather than humans: branch-per-session as the unit of work, branch descriptions instead of per-commit messages, and content-addressed lazy mounts that get an agent editing any repo in seconds. Because it hydrates on demand, it claims to be faster than git for agent workloads. Bring your own agent, Claude Code, Codex, or Cursor, and Oak is the substrate underneath.