Seven signals to understand AI this week: model leaks, code engines, personnel management

Author: Tara Tan / StrangeVC

Compiled by: ShenChao TechFlow

ShenChao Briefing: This week’s roundup is extremely dense—seven independent signals covering some of the most critical trends in the AI industry.

Most worth watching: Anthropic accidentally leaked a new model’s internal codename, “Capybara,” via a CMS configuration mistake; its tier sits above Opus.

Full text below:

Over the past few months, we’re almost certainly past some agentic threshold. What used to take four to six weeks to build five years ago now takes less than five minutes. Six months ago, the same task was still one to two hours plus a lot of debugging.

This is a fairly significant phase transition that we may not have fully absorbed yet. The collapse of the distance between ideas and working products will rewrite the entire industry. It’s a step-change in the tools humans use to build, create, and solve problems.

Related to this, OpenClaw has clearly become more stable since OpenAI’s acquisition. It has a clear path to becoming one of the most important open-source projects in the AI space.

Now, into this week’s content.

Anthropic’s Claude Mythos leak reveals the next model tier

Due to a CMS configuration mistake, Anthropic accidentally exposed details of an unreleased model named Claude Mythos. The leaked draft describes a new “Capybara” tier, positioned above Opus, with major breakthroughs in programming, reasoning, and cybersecurity capabilities. Anthropic confirmed it’s testing the model with early access customers, calling it a “step-change” and “the most powerful model built to date.” (Fortune, The Decoder)

Why it matters: Beyond the model itself, there are two things worth paying attention to even more. First, the leaked draft warns that the model’s cybersecurity capabilities are “far beyond any other AI model,” which drove the movement in cybersecurity stocks within a single trading day. Second, introducing a fourth model tier (Capybara sits above Opus) suggests Anthropic is building pricing room for enterprise customers—not just building performance room for benchmark tests.

Claude Code is becoming Anthropic’s core growth engine

Claude Code currently accounts for about 4% of all public GitHub commits, and is expected to reach 20%+ by year-end. Anthropic’s overall revenue is estimated to have an annualized run rate of $14 billion, while Claude Code alone has an annualized run rate of roughly $2.5 billion. The users of this tool have expanded from developers to non-technical users, who are learning terminal commands to build projects with it. (SemiAnalysis, Uncover Alpha, VentureBeat)

Why it matters: Claude Code compresses customer acquisition costs to near zero through organic developer adoption. By extending to non-developer roles via Cowork, the addressable market expands far beyond the global 28 million professional developers.

Cheng Lou’s Pretext: Text layout without relying on CSS

Cheng Lou is one of the most influential UI engineers of the past decade (React, ReasonML, Midjourney). He released Pretext, a pure TypeScript text measurement algorithm that completely bypasses CSS, DOM measurement, and browser reflow. The demo effects include: virtualized rendering of tens of thousands of text boxes at 120 frames per second, tightly packed chat bubbles with zero pixel waste, responsive multi-column magazine-style layouts, and variable-width ASCII art. (X post)

Why it matters: Text layout and measurement have long been the hidden bottlenecks blocking the next generation of UI. CSS was designed for static documents—not for the fluid, AI-generated, real-time interface design that has now become mainstream. If Pretext delivers on the demo, it will remove one of the last foundational constraints on AI-native interfaces in terms of appearance and experience.

Arm ships its own chips for the first time in 35 years

Arm has released an AGI CPU: a 136-core data center processor built on TSMC’s 3nm process and jointly developed with Meta. This is the first time in the company’s history it has sold finished chips rather than licensed IP. OpenAI, Cerebras, and Cloudflare are the first partners; mass shipments are expected to begin by the end of the year. (Arm Newsroom, EE Times)

Why it matters: Today’s AI data centers are dominated by GPUs. GPUs handle training and model execution, while CPUs mainly process data streams and scheduling. But agentic workloads are different. When thousands of AI agents run at the same time—each coordinating tasks, calling APIs, managing memory, and routing data across systems—those orchestration jobs land on the CPU. Arm claims this will drive a 4x increase in CPU demand per gigawatt of data center capacity. (HPCwire, Futurum Group)

NVIDIA and Emerald AI turn data centers into grid assets

NVIDIA and Emerald AI announced an alliance with AES, Constellation, Invenergy, NextEra, and Vistra to build a “flexible AI factory” that participates in grid balancing services by adjusting compute load. The first facility, Aurora, is located in Manassas, Virginia, and is scheduled to open in the first half of 2026. (NVIDIA Newsroom, Axios)

Why it matters: The biggest bottleneck for AI infrastructure expansion isn’t chips—it’s the grid interconnection timeline, which in most regions requires 3 to 5 years. Data centers that can demonstrate grid flexibility can interconnect faster and face less regulatory friction. This redefines the energy thesis for AI infrastructure investors: the winning argument isn’t “more power,” but “smarter power.”

China restricts Manus AI executives from traveling abroad

After Meta’s $2 billion acquisition of this Singapore-registered AI startup, Chinese authorities restricted Manus CEO Xiao Hong and chief scientist Ji Yichao from traveling abroad. This month, the National Development and Reform Commission summoned the two executives to Beijing and imposed travel restrictions during the regulatory review. (Reuters, Washington Post)

Why it matters: This isn’t a trade restriction—it’s a personnel restriction. China may be signaling that AI talent with mainland backgrounds is a controlled asset, regardless of where the company is registered.

400B-parameter model runs locally on iPhone 17 Pro

An open-source project called Flash-MoE demonstrates a 400B-parameter mixture-of-experts model running entirely on-device, using the A19 Pro chip in iPhone 17 Pro; the model weights are streamed from SSD to GPU. The model (Qwen 3.5-397B, 2-bit quantization, 17.0B active parameters) runs at 0.6 tokens per second, with 5.5GB of RAM remaining. (WCCFTech, TweakTown, Hacker News)

Why it matters: This is a proof of concept, not a product. A 400B-parameter model can run on a phone with 12GB of memory because, at any given moment, only a small portion of the model is active (mixture-of-experts); the rest is streamed from the phone’s built-in SSD on demand rather than resident in memory. But applying the same trick to much smaller models—say 7B or 14B parameters—on next-gen faster mobile chips with storage, and you get truly usable AI with full on-device local inference and dialogue speed, without relying on the cloud.

AI Agents autonomously completed an entire particle physics experiment

MIT researchers published a framework called JFC (Just Furnish Context) demonstrating that an LLM agent built with Claude Code can autonomously execute a full high-energy physics analysis pipeline: event selection, background estimation, uncertainty quantification, statistical inference, and paper writing. The system runs on open data from the ALEPH, DELPHI, and CMS detectors. (arXiv 2603.20179)

Why it matters: This is one of the clearest demonstrations that agentic AI can automate end-to-end scientific workflows in domains where methodological rigor is extremely high. The direct investment implication points to re-analyzing legacy datasets in physics, genomics, and materials science—decades of archived data that still haven’t been fully mined.