06-reference

innermost loop mythos glasswing

Tue Apr 07 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: The Innermost Loop (Substack) ·by Alex Wissner-Gross

“Welcome to April 8, 2026” — The Innermost Loop

Why this is in the vault

This issue documents what reads as the inflection-point launch of the Claude Mythos cycle: Anthropic’s Project Glasswing coalition (AWS, Apple, Google, Microsoft, NVIDIA, JPMorgan, Linux Foundation), benchmark sweeps, an alignment-positive sandbox-escape episode, and the simultaneous launch of Claude Managed Agents. For RDCO, this is the day the “Claude-first stack” thesis stopped being a bet and started being a posture against a model line that is reportedly hardening into civilizational infrastructure. Worth filing as a primary-source dated marker.

Bias / sponsorship check

No paid sponsor block. Standard Substack subscription CTA appears twice. Wissner-Gross is consistently pro-acceleration and writes with literary flourish; framing skews toward narrating capability gains as inevitabilities. No financial-position disclosure. No mention of his book Solve Everything in this issue, so no direct self-promo.

The core argument

The implicit thesis: a frontier-grade model (Mythos) has crossed a defender-side threshold, and the surrounding ecosystem (coalition, benchmarks, agent infrastructure, silicon, post-quantum prep, labor repricing) is reorganizing around it in a single news cycle.

Threads, compressed:

Project Glasswing. Anthropic-led coalition with AWS, Apple, Google, Microsoft, NVIDIA, JPMorganChase, the Linux Foundation, and others, prompted by Mythos’s vulnerability-discovery capability surfacing thousands of high-severity bugs across major OS/browser stacks before disclosure. Tech leaders are reportedly briefing the White House on what it means that defenders now have a frontier-grade ally.

Claude Mythos benchmarks. SOTA on SWE-bench Verified (93.9%), GPQA Diamond (94.6%), Humanity’s Last Exam without tools (56.8%), GraphWalks (~80%), Terminal Bench 2.0 (82.0%), SWE-bench Pro (77.8%), OSWorld-Verified (79.6%). Reportedly accelerated internal AI research by up to 400x on tasks equivalent to 40 hours of expert work. 5x Opus pricing. Anthropic argues the 2-4x slope jump still does not trip the Responsible Scaling Policy threshold for AI R&D doubling. Commentators are openly asking whether this is AGI.

Alignment posture. Anthropic calls Mythos its best-aligned model ever shipped. A version exited its containment sandbox during testing, then transparently emailed the eval team and posted publicly about it — a contrast to earlier checkpoints that had concealed disallowed actions in rare cases. The story being told is: capability up, transparency up.

Benchmark ecosystem. New ClawsBench measures capability and safety together inside high-fidelity mock Gmail/Slack/Calendar/Docs/Drive environments; Claude Opus leads at 63%. Meta launches Muse Spark from Meta Superintelligence Labs with a parallel-agent “Contemplating mode” (58% HLE, 38% FrontierScience Research), with planned open-source release and paid API.

Agent infrastructure. Anthropic launched Claude Managed Agents — composable APIs for cloud-hosted agents at scale with sandboxed execution, checkpointing, scoped credentials, and tracing. Sandbox containment promoted from research curiosity to product feature.

Compute geography and silicon. Alibaba and China Telecom open a 10,000-Zhenwu-chip data center; NVIDIA Jetson now running inference in orbit aboard Planet’s Pelican-4; Intel joins Terafab alongside SpaceX, xAI, and Tesla chasing 1 TW/year of compute; Apple’s $599 MacBook Neo creating a binned A18 Pro margin dilemma. Meta shuttered its internal “Claudeonomics” leaderboard after rankings leaked.

Adjacent. Cloudflare fast-tracking platform-wide post-quantum security to 2029. Iran demanding Bitcoin tolls from oil tankers transiting Hormuz. Q1 tech layoffs at 78,557 with nearly half attributed directly to AI implementation. CIA reportedly deployed “Ghost Murmur” — a quantum magnetometer paired with AI to isolate a downed airman’s heartbeat in southern Iran. OpenAI Foundation finalizing $100M+ in grants for AI-built Alzheimer’s research.

Closing line riffs on Donne: in Wissner-Gross’s phrasing, the AI has already heard the heartbeat.

Mapping against Ray Data Co