Gemini 4 Argon, defenders first 🛡️, Meta's data centres are "experiments" 🧪, Knowledge workers demoted to AI managers 🧑‍💼

Google DeepMind is rolling Argon out through its Fairwind Program, without cyber guardrails for trusted defenders, before paid API customers and Google AI Ultra subscribers get access. Google reports a state-of-the-art 77.9% on DeepSWE v1.1 and says Argon agents freed over 300 TiB of memory across its data centres. It will launch at an introductory $2 per million input tokens and $10 per million output tokens.

The industry-wide probe targets the potential dangers AI poses to consumers, a senior FTC official told Reuters, and the agency plans to compel testimony from executives at Anthropic, OpenAI and research group METR. The official said OpenAI agents hacking Hugging Face added urgency. FTC Chairman Andrew Ferguson suggested last week that developers whose agents cause hacks in cybersecurity tests should be liable for the harm.

Operators copied encrypted reasoning from one conversation and asked a model in another to decrypt and transcribe it, peaking at 16,000 requests from over 4,000 users on July 24 and 25, figures OpenAI says count attempted extractions. It disrupted a related cluster of more than 15,000 users by July 28 and closed the replay pathway. OpenAI warns systems supporting portable or replayable reasoning artefacts may face related risks.

Meta classifies its AI data centres as experimental "pilot models" so its chips qualify for the research and experimentation tax credit, the New York Times reports. Meta said the credit cut $2 billion off its taxes in 2024 and $3.9 billion in 2025, up from $700 million in 2023. Its reserve for tax positions the IRS could overturn has risen 45% to $18.74 billion, with research credit uncertainty listed first.

Executives including Nvidia's Jensen Huang asked Anthropic's Dario Amodei why he was so extreme in public about AI risks, after a White House lunch where about two dozen CEOs agreed to safety principles, people familiar told the WSJ. Amodei said it was important to be honest about model capabilities. The WSJ says the tensions could hinder the industry's plan to self-regulate through model evaluations and internal controls.

Cloudflare says more than half of internet traffic this year wasn't human, and human traffic to heavily crawled categories like retail fell as much as 40% in under a year. Its Pay Per Use beta pays publishers when verified crawlers report using their content. Monetization Gateway, in closed beta for eligible US customers, returns an HTTP 402 via the open x402 protocol so agents pay per request.

Pi's website once declared it did not support MCP, its makers were dismissive of the protocol on podcasts, and MCP lived only as an extension. Earendil moved it into the core because the changes it needed, letting tools be deferred or reserved for Codemode, a JavaScript sandbox that chains tool calls on the harness side, were useful across Pi. After upgrading, configuring an MCP server loads Codemode automatically.

Struct's sporadic agents wake for escalations, developer questions and deployment checks, and an idle gap longer than the cache TTL means rewriting the full prompt. In its telemetry, a 30-minute window covered about 88.5% of gaps between wake-ups against 72.5% for 5 minutes, and a 1-hour window 93.3%, but Anthropic bills 1-hour cache writes at 2x input price, which Struct judged not worth it. Cache misses fell from 24.2% to 9.2%.

Claude's Computer Use, which works from screenshots alone, and Perplexity's agent in Comet went straight to Save in 25 of 25 runs, so Merj reads the split as perception modality rather than vendor quality. In its 100 purpose-built tests, average success rose from around 30% to around 60% between November 2025 and March 2026. Merj says the Cancel page would score close to perfect on readiness checkers, so teams should test whether agents finish their own journeys.

Exploited vulnerabilities rose from an average of 10.5 a month in 2025 to 18 in 2026, which GTIG attributes mainly to rapid weaponisation of already disclosed flaws rather than new zero-days. Of the bugs it could identify as AI-discovered, 50% led to remote code execution, against 26% across all CVEs, and one AI-found BeyondTrust flaw was exploited within four days of disclosure. Agent orchestration frameworks now account for 50% of AI-related flaws.

Researchers had Claude rate each job task by how structured an environment a robot needs to do it, counting only real deployments, sales or demonstrations. Driving and warehouse jobs score as highly exposed while nursing and general repair do not, and jobs more exposed to robots since 1977 saw wage and employment declines in later decades. Robots are cost-competitive on just 0.3% of tasks, and on past price trends Anthropic estimates 40 years to reach 10%.

Jain says the five-layer model for verifying AI code in that earlier essay checked correctness but missed human judgment and the conversation where a team builds its mental model of the system. The revised order is Argue, Capture, Codify, Debate and Own: agents disagree before the PR exists, sessions record intent and acceptance criteria, and repeated review comments become invariants. Humans then debate only what the agents could not settle, and the team owns the invariants.

The TLA+ educator and advocate is responding to the rush of interest after Boris Cherny said Opus used TLA+ to find race conditions in code. TLA+ checks invariants, action properties and liveness across every individual behaviour, so it cannot natively state two-step properties, reachability such as a game being winnable, or hyperproperties covering many security and percentile-latency guarantees. Workarounds exist, Wayne writes, but they blow up state space or stop models matching the real system.

In a post on X, Narayanan says the AI jobs debate fixates on outright replacement and misses a more likely harm: a growing share of knowledge workers now manage AI agents instead of doing the work themselves. Managing people means losing the craft, owning others' mistakes and splitting attention, offset by higher pay, status and the reward of mentoring. Managing AI, Narayanan argues, brings the downsides without those upsides, and the role change is forced.

The free, Apache-2.0 app runs on macOS on Apple Silicon, Linux, Windows 11 with WSL, iPhone and iPad, with Android in beta. Each note keeps its own persistent shell, can run blocks over ssh on remote hosts, and pulls secrets from profile files kept outside the notes folder. An MCP server lets Claude Code and other agents read and edit notes, with no delete tool and locked note bodies withheld from them.

The GPL-3.0 app needs macOS 14 or later on Apple Silicon, and no bot joins the meeting, while transcription, speaker detection and your documents stay on the Mac. The copilot can run on Claude with your own key, about $0.07 per call hour by the README's figure, any OpenAI-compatible server, or a local Ollama model that sends nothing off the Mac. The developer calls it a personal project that still has bugs.